A research programme (proposal and log on the project page) asking whether a stock humanoid, a mobile manipulator with a lift or a purpose-built gantry can hang a 4x8 sheet of wall sheathing and land all 62 nails where the panel maker and the code want them, which error source decides it, and what a clean sheet costs; it gathers track 20's 26 buildbot-* projects into one question aimed at Automation in Construction.
With a 0.5 mm stud reading, a robot aiming at the design window has no tolerance for sheet pose at all; aiming from the measured sheet edge (to 0.5 mm) allows 0.44 mm of translation and 0.031° of rotation (1σ), and halves the weight of the stud error. Along a 40 ft wall, sheets butted with a 1/8 in gauge drift 3.2 mm a sheet: no third sheet is clean with full panels, and panels trimmed 1/16 in keep a joint window for four sheets; setting each sheet from its own studs keeps every sheet at 98% clean.
Show the rest of the summary (64 more paragraphs, 5,286 words)
Framing the joints with a 3x or two fastened 2x (named for adjoining panel edges in the shear-wall standard, SDPWS 4.3.7.1) widens the window by 1/2 in a side, and three reads then give 100% clean sheets at both warp levels; twist costs 1–2 points.
Milestone 4 asks how coarsely a robot may hang the sheet if it tacks it with two field nails, measures where the edge landed and aims the joint nails from that, or re-hangs it when the window looks too tight. On single-2x joints the sheet must still hang to about 0.03° and 0.4 mm (1σ) for 95% clean sheets; re-hanging a 1 mm / 0.1° hang peaks at 87% (2.1 hangs a sheet), and a held sheet may sway only 0.19 mm under the nail. On two 2x fastened at the joint, tack-then-measure allows 0.54° and 10 mm at the lower published warp, and a held sheet may sway 4.1 mm: the placement target for the humanoid and lift experiments.
Milestone 3, first part asks whether a ledger board under the sheet can do the squaring a robot cannot. In MuJoCo, a 21.8 kg sheet lowered 0.25–2° off plumb onto a level ledger and released ends within 0.0001° of it for any press up to 100 N; above 115–380 N (wall friction 0.7 to 0.2) friction holds the whole error, at 0.87–0.99 of the closed form, and a wrist that keeps over about 7,500 N·m/rad of stiffness at 1° keeps part of its own error.
Seated, what is left is the panel (PS 2-18: 0.083° out of square, a 1/16 in edge bow that rocks it up to 0.149°) and a laser line rated 1/8 in at 30 ft: 100% clean sheets on doubled joints in every setting, 84–86% on single 2x with panels drawn up to the PS 2 limits, and 95% only if panels are within a quarter of them.
Milestone 3, second part puts a stock Unitree G1 or H1 (MuJoCo Menagerie, static whole-body model, torque limits, no foot slip) in that holder's place. Neither can carry the sheet to the wall: the G1 holds at most 2.8–8.7 kg in both hands at 0.6–1.3 m and the H1 12.4–18.8 kg. With the sheet on the ledger, the G1 presses 66–110 N and the H1 128–281 N. Sliding the seated sheet to its stud line tips it unless pushed below W/(2μl) (MuJoCo within 1.6% in 14 cases); a G1 manages the slide at ledger friction 0.3 but not 0.5, and not through palm friction at all.
At Unitree's published joint gains (G1 arms 60 N·m/rad) the sheet's stiffness about the landing corner is 2.9 to 31 times under the limit at which a hold keeps its error, so the ledger seats it; only an H1 with a rigid clamp is too stiff (it keeps 0.46–0.99° of a 1° error). The catch: that softness would let the sheet's weight turn a hold 10–14° if the ledger did not carry it. Hang and sway errors are swept, not measured on a robot; the humanoid model is static with gains taken as its only compliance, and the panel distributions within PS 2's limits are assumed.
Monte Carlo checked against buildbot-studfind and closed forms. Stud, edge, gauge and width errors are swept, not measured; the gap, edge distance and trim come from APA pages seen in search results; stud warp means and limits come from search snippets of USDA and Wood and Fiber reports, the warp distribution shape is assumed (two are shown), and nothing is tested on hardware.
Milestone 5 puts the stages in one steady-state cycle along a row: a lift stands the sheet on the ledger, the robot slides, presses and tacks it, reads the three new studs and both edges, nails with edge aim (timed by buildbot-mobilebase, 54 or 121 s) and re-drives every missed nail. On two 2x at the joint a single read a stud gives 0.9994 first-pass clean sheets at the lower warp and is the fastest choice in all 108 swept time cells: 76–216 s a sheet.
On single 2x no sensing policy passes 0.59 (0.7–2.2 nails missed a sheet), three reads are fastest, and the best sheet is 17.9–65.5 s slower, all of it rework: at 30 s a re-driven nail, framing the extra joint stud may take 30–42 s a sheet and still win. Firing without waiting out a 4 N·s recoil costs up to 7.4 points of clean sheets on single 2x (a conservative model) and nothing on doubled joints, and on doubled joints a lift that hangs and tacks the sheet to 0.25° and 5 mm needs no ledger (0.978). Every stage time except the nailing is swept, not measured.
Milestone 7 prices those seconds: ownership (price, life, interest and upkeep over the wall sheets of 10 to 1,000 houses a year, capped at 2,000 machine hours), supervision and the person who still carries the sheet at a loaded carpenter wage, and the extra joint stud ($3.41–5.56 retail plus a framer's time). A hand crew costs $5.21 a sheet (median; $1.93–14.46 over the taskbook's sweep of its minutes). A G1 on a drive-and-lift base costs $54, $19, $7.03, $3.48 and $2.20 at 10, 30, 100, 300 and 1,000 houses a year and beats the crew in 37% of draws at 100 and 74% at 300; a walking H1 costs about 2.5 times as much and never beats the base; a gantry in a plant ($40k–600k, unsourced) is cheaper than the base in the median draw only from about 500 houses a year.
Every robot picks single 2x joints with three reads: the extra stud costs $5.42 a sheet and saves under a dollar of rework, so doubled joints pay only if a missed nail costs about $6.50 more than its re-drive. Machine prices but the humanoids', lives, upkeep and stage times are swept, not sourced.
Milestone 6 rebuilt the ledger seat and slide in a second physics engine, Drake (SAP contact): a lightly pressed sheet still seats, leaving 0.0014° (1σ) against MuJoCo’s 0.00001°, and every ledger pipeline’s clean-sheet rate moves by 0.05 points at most; the press at which wall friction stops the ledger working is 97–100% of the closed form in Drake and 87–99% in MuJoCo. How far a bowed bottom edge rocks the sheet depends on the engine (0.11–0.15°, both under the bound charged), and a soft held wrist keeps up to 0.016° in Drake.
Milestone 6b shows that engine dependence is an artefact of modelling the bow as three feet: a parabolic 1/16 in bow has a radius of 117 m against a centre of mass 1.22 m above it, so the sheet is a stable rocker that rolls back to level and keeps only what wall friction holds (θs = μFnr̄ / (mg(R−h0))). Both engines seat inside that band in all 48 cells (99th percentile 0.06° over 2,000 seats, against up to 0.149° charged before), the best single-2x pipeline rises from 0.560 to 0.625 clean sheets, and doubled joints move by under half a point. A kink, or 1 mm of grit on the ledger, keeps a rock of 0.05–0.15° that depends on how the sheet lands.
Milestone 8 writes milestones 1–7 up as a paper draft for Automation in Construction, with a script that checks each of its 70 quoted figures against the results file it came from: three of the proposal's five hypotheses held, the gantry's crossing volume was refuted as registered, and the edge-aim hypothesis held only at a 0.5 mm stud read.
Milestone 9 prices the carpenter's alternative to a doubled joint stud: pushing the bowed stud onto the line. As a Stud-grade 2x4 bending flatwise between pinned plates it takes 48 E I b / L3 — a median of 28 N and a 99th percentile of 126 N at the published warp, 376 N at the grade limit — which a stock G1 can just manage (81–88 N sideways at stud mid-height). Straightening lifts single-2x joints from 56% to 71% clean sheets at the lower warp and from 31% to 68% at the higher, but not to 95%: on perfectly straight studs the same pipeline gives 74%, so bow was costing up to 37 points and the sheet's own pose on the ledger costs the rest. Pushing each nail's point in turn, rather than once at mid-height, needs up to 1,258 N, so push first and then nail.
Milestone 10 splits that 74–81% ceiling by Shapley share over six error sources: the nail gun’s 0.9 mm scatter carries 45–47% of the misses, the sheet’s pose (slide, ledger slope, panel geometry) 39–44%, the stud and edge reads the rest, and neither the tool nor the pose misses on its own (0.957 and 0.995–1.000 clean): each narrows the ±3.14 mm window the other must hit. Sliding the sheet along the ledger to centre its gap and shimming it plumb before tacking lift straight studs to 92–94% and pushed studs to 89–92%; 95% then needs a tool held to about 0.8 mm (1σ). Slide and shim accuracies are swept, not sourced.
Milestone 11 measures six installed studs from the Adel et al. (2024) laser scan of a robot-built 2x4 wall (CC BY 4.0): their bows scaled to 8 ft (median 0.106 in) sit on the published mean, but 0.3–0.9 mm RMS of waviness remains after a parabola, and on the measured shapes the best single-2x pipeline gives 0.43–0.47 clean sheets against 0.58 for parabolas at the same bows; doubled joints give 1.000.
Part of the waviness is common to every scanned member and is treated as scan stitching; one wall is a small sample.
Milestone 12 runs milestones 9 and 10’s remedies on those measured studs: reading each stud at every nail height lifts single 2x from 0.425 to 0.530 (the waviness cost falls from 15 to 5 points), a mid push measured from the end chord adds 13 points, and per-nail reads, push, slide and shim together give 0.771 against 0.860 on the same studs’ parabolas; 95% then needs a tool held to 0.61 mm. Doubled joints give 1.000. Slide, shim and push errors are swept, not sourced.
Milestone 13 prices those remedies with milestone 7’s cost model (G1 on a lift base, 4,000 draws, every missed nail re-driven), charging their seconds: a line-laser per-nail read, the push and the slide and shim (swept, not sourced). First-pass quality is cheap: three reads alone cost a median $7.10 per clean sheet at 100 houses a year (0.425 clean), the full stack $7.42 (0.771). The doubled joint stud reaches 1.000 but costs $12.17, more than the cheapest single-2x pipeline in every draw at 100, 300 and 1,000 houses a year; it pays only if a missed nail costs a median $5.51 more than its re-drive (milestone 7: $6.47 on parabolic studs). A line laser beats three point reads above about 0.23 m/s.
Milestone 14 asks what a missed nail that is never found does to a shear wall, with a rigid-plastic fully anchored model (Källsner and Girhammar 2009) that tracks SDPWS’s spacing table within 6%. One joint-edge miss costs 2.0–2.2% of the sheet; with three reads on measured studs the mean sheet keeps 97.7% but the 99th percentile only 88.5%, because a misjudged stud is missed at a run of heights. Nailing the joint edges at 4 in keeps 99.98% of sheets at or above the intact 6 in sheet with every miss left in, for a median $6.98 a sheet against $7.10 with re-drives. The model loses 20% with the bottom foot unnailed where tested walls lost about half (Weiss 2017), so its losses are lower bounds.
Milestone 15 rebuilds that model as a limit analysis of the whole framed sheet (a linear programme over the top plate, studs and sheet), with stud-to-plate end-nail joints at the NDS yield-limit ratio of 2.55 sheathing nails. It reproduces milestone 14 with rigid joints (0.09%), but at code-strength joints the unnailed bottom foot still costs only 20–21%: Weiss’s 40–50% needs joints at 0.46–0.56 of that ratio, or no hold-downs (64–100%), so the gap is not the framing joints at design strength.
A lone sheet whose top plate can tilt loses 7% intact and 16.7% at the 99th-percentile escaped-miss sheet; a sliding top plate restores milestone 14’s figures, and 4 in joint edges carry the misses under every joint model. Design-level ratios, not tested ultimates.
Milestone 16 pushes milestone 14’s sheet over with each nail a spring on the tested backbone of an 8d common nail through 7/16 in OSB into SPF (Shameli Derakshan et al. 2019: 0.97 kN peak, 0.50 kN/mm at 0.4 of it, 0.8 of the peak at 29 mm; the slip at peak swept 6–15 mm). It reduces exactly to the rigid-plastic model, tracks SDPWS’s spacing ratios within 6%, and peaks at 9.8 kN a sheet (551 plf) at 2.3% drift.
With the bottom foot unnailed it still loses 20.1–20.4% at every backbone, and 20.4% for a near-brittle nail, because at the sheet’s peak the perimeter nails all slip 11–12 mm: softening does not explain Weiss’s walls, which leaves the anchorage. The escaped-miss figures stand (2.39% mean, 11.8% at the 99th percentile) and 4 in joint edges carry them. Monotonic loading only.
Milestone 17 reads Weiss’s thesis in full and builds her tested wall: 8×8 ft, two OSB sheets on a single middle stud, hold-downs only on the double end studs, every nail below 12 in omitted (tested: 57% kept with 6 in edges, 34% with 3 in). The wall limit analysis matches the controls’ spacing ratio (1.93 against 1.87) and their absolute capacity to 0.89–0.92 at a tested nail strength, but at NDS-ratio stud-to-plate joints the decayed walls keep 80% and 63%: the anchorage as tested does not explain them.
Post hoc, end nails at 0.47 of the NDS ratio fit both groups (58%, 30%), because the decayed wall’s least-work mechanism becomes the whole wall sliding on its 18 sill nails, which makes its capacity independent of the sheathing schedule, as Weiss found; one fitted parameter, so a tested end-nail ultimate is the check. An escaped miss costs 2.7% of a two-sheet wall on average and 10.4% at the 99th percentile.
Milestone 18 adds the shear path through the hold-downs’ 5/8 in anchor rods (NDS design value 617–1,137 lb a rod): at the 0.47 fit the slide stops once each rod carries about 420 lb, and with a rigid path no end-nail strength fits (best rms 22.8 points; the decayed 3 in wall keeps 64% against 34% tested). A fit within 5 points needs each rod below about 970 lb and the end nails weaker as the rod gets stronger (a quarter of the NDS ratio at the rod’s design value, below the nail’s own design value), so a rigid-plastic model with sourced strengths does not reproduce the 3 in wall; rod slack, stud splitting and cyclic tear-out remain. An escaped miss still costs 2–3% of a wall.
Milestone 19 pushes the same wall over as a displacement model (sourced nail load-slip curves, softening stud joints, an anchor rod that bears after a gap): it matches the rigid-plastic limit to 0.04%, a 1/16 in gap moves the decayed walls by 0.1 point, joints that peak early leave the 3 in wall at 71%, and no joint strength at 0.75 of the NDS ratio or more comes within 10 points of both tested groups. All 30 fits within 5 points (of 960 pushovers) have the end nails at a quarter of the NDS ratio, so a tested end-nail ultimate, or a cyclic model, is what decides the 34%. The end-nail joint’s curve is swept, not sourced.
Milestone 20 put the wall through the CUREE cyclic protocol with each nail on the SAWS pinching hysteresis, fitted to cyclic tests of nails through 7/16 in OSB into Douglas fir-larch (Anderson et al. 2007) and matching OpenSees to 2×10−13: the intact peaks move 0.1%, the decayed 3 in wall keeps 70.6% against 70.5% monotonic (tested 34%), and across 96 combinations the median change is 0.2 point. SAWS has no fatigue or tear-out term, so those, and weak end nails, are what is left.
Milestone 21 let every nail lose strength in proportion to the energy it has dissipated (the c = 1 form of the Rahnama–Krawinkler rule used by Ibarra, Medina and Krawinkler; the rate γ is not sourced and is swept from none to a rate that costs a lone nail 39% of its CUREE peak): the intact walls’ cyclic peaks fall to 63–97% of monotonic, but the decayed shares rise by 0.3–2.9 points at code-strength joints (3 in wall 70.6% to 73.5%), since an intact wall works its sheathing nails harder. Energy-spread damage cannot make the 34%; what is left is damage that singles out the decayed wall (tear-out along its unnailed bottom edge, the middle stud splitting) or weak end nails.
Milestone 27 models the slide that milestones 10–13 credited with 0.5 mm as stick-slip on the ledger: the stock G1 hand is a 63 N/m spring along the wall (a force source in effect), so how far the sheet runs is set by the time from slip to release. Tapping from either edge leaves 0.19–0.64 mm rms at a 20 ms reaction and 2.5–10 mm at 100 ms, and the best single-2x stack goes from 0.771 to 0.79 at 20 ms and 0.32 at 100 ms (μk/μs 0.8; the OSB friction pair is swept, not measured).
Milestone 28 does the same for the shim (lift the low corner until the edge is plumb, also credited with 0.5 mm): a stack of 1/16 in plastic shims leaves 0.458 mm rms and keeps the stack at 0.796, while a 1:32 composite wedge needs a 118.5 N push at μ 0.4 against the stock G1’s 65.6 N at a 0.5 m ledger, and leaves 0.13 mm at a 20 ms release because a hand weighs under 1 kg along the push (shim face frictions swept, not measured).
Milestone 29 carries the tool error to where it acts: the nailer’s nose 0.3 m ahead of the grip, and the shank 44.4 mm into the stud, where a leaning nail has moved again. The stock G1’s 11-bit joint error becomes 1.31 mm along the wall (0.93 at the hand), and the best single-2x stack falls from 0.796 to 0.264 (H1 0.571); 12-bit sensors give 0.934, and an extra nail lean of 0.43° costs 2 points. Doubled joint studs stay at 1.000 (only joint quantisation is modelled; backlash and deflection would add).
Milestone 30 adds gear backlash (no G1 figure is published, so it is swept to 25 arcmin): it randomises only joints that gravity and the press do not hold one way, and at the stud pose those are the G1’s two ankle rolls. 5 arcmin on every joint takes the 12-bit G1’s stack from 0.934 to 0.765; counting held joints 0.913; with the ankle rolls biased one way, no loss up to 25 arcmin.
Milestone 31 carries it across 15 G1 poses (0.4–1.2 m high, ±0.15 m along the wall): calibrated at every pose, 5 arcmin costs 0.3–9.9 points (worst at the top of the reach, where knees and shoulder go loose), but both hip yaws and the waist roll flip the sign of their torque across the grid, so one tool-frame calibration leaves up to 3.4 mm along the wall; 12-bit joints then tolerate 0.44 arcmin calibrated once against 2.5 calibrated per pose (ground-reaction split least-norm, backlash and holding threshold swept).
Milestone 32 makes the feet push through their soles’ contact points with friction cones: upright, the G1 balances only 26–39 N of press above 0.8 m, so it must lean 0.04–0.12 m into the wall; on the leaned poses the hip-yaw and waist flips vanish under every ground-reaction model, a controller cannot hold every leg joint one way, holding the hip rolls, hip yaws and ankle rolls cuts the worst per-pose loss to 3.7 points, and one calibration still leaves 1.1–2.6 mm (statics only; contact points from the model’s collision geoms).
Milestone 33 tests the remedy: predicting each held joint’s side from the static torques and subtracting the bias leaves nothing when the compensator models the robot’s own balance controller, and with a controller that holds the light leg joints a 64-nail magazine left out of the model costs 0.017 mm (+1 kg 0.057 mm; 0.41–0.84 mm under a torque-minimising controller); a wrong ground-reaction model leaves 1.16–1.92 mm, little better than none, and a guard band on predicted torque does not help.
Milestone 34 adds the joints’ elastic windup under the 25–50 N press (transmission stiffness swept, none published): with the nose free to slide, the 12-bit G1 loses 5.5 points at 104 N m/rad and needs about 1.8×104 to stay within 2; held by friction the windup only tilts the tool (0.02 mm at the nail’s tip), and at μ 0.4 the nose sticks at 10 of 15 G1 poses; an unmodelled magazine droops the servo by 32–51 mm under Unitree’s PD gains, which the encoders see.
Milestone 35 treats the air hose as a force: hanging and dragged along the floor (listed hose weights, floor friction and drag swept), a 5 N pull winds the G1’s nose up 1.25 mm at 104 N m/rad and leaves the matched backlash compensator 1.0 mm, so a hose tension of 0–5 N costs 2.2 points and 0–10 N 21; the hose must come off the tool, be measured, or drag lightly.
Milestone 36 measures the wrench with a wrist F/T sensor (Robotiq FT 300 listed noise 0.1 N): the G1’s residual at a 20 N drag falls from 2.34 to 0.29 mm and 0–20 N costs 1.6 points instead of 91.6, with the joints’ stiffness misjudged by 25% costing little; but even a 0.01 N error flips a held joint sitting at the (modelled, unsourced) friction threshold, leaving a floor of 0.3–0.8 mm, and the sensor’s pre-2017 version (1.2 N) is worse than none.
Milestone 37 plans from the measured wrench and lets the robot settle under the true load: without a preload that costs 56.5 points over 0–20 N, because the balance controller holds its chosen leg joints exactly at the threshold (92% of the flips); holding every leg joint 0.1 N m clear of it removes them and returns 1.6 points for 1.7% more leg torque and one of 15 poses, while the 0.3–0.8 mm left comes from waist and arm joints the ground reaction cannot move.
Milestone 38 adds a squeeze between the two hands along the line joining them, which has no net wrench and so moves only the arm joints: within 10 N it clears every arm flip, the G1’s 0–20 N stack loss falls from 1.64 to 0.05 points (the H1’s from 2.19 to 0.02), and what is left (up to 0.50 mm) comes from the waist joints, which only the tool’s wrench sets; the G1’s wrist pitch already exceeds its 5 N m model rating at 7 of 14 poses.
Milestone 39 sets the waist’s backlash: braking it takes the G1’s worst residual at a 5 N drag from 0.485 to 0.012 mm and its 0–10 N stack loss from 5.80 to 0.04 points, a 1 arcmin waist gearbox gives 0.106 mm, and locking only the roll and pitch (as on the base G1) changes nothing because the yaw is the joint that flips.
Milestone 40 brackets the nailer’s recoil from published measurements (0.09–8.7 N s: an 8d nail at chronograph speeds, and a force-plate study of framers), enough to lift the nose: the press that rises again sets none of the 56 loose joints, 36 of which change sign inside the 25–50 N window, so the spec’s 0.012 mm stands (0.203 mm only at a 20 N drag, without a compensator that plans the rise); at 12 of 35 poses, though, the leaned stance balances only above 3.7–11.8 N of press. The rise is quasi-static and the holding threshold unsourced.
Milestone 41 re-leans those 12 stances 10–39 mm back so each stands at zero press and still balances at 50 N; the window’s width is (m + mtool) g × sole length / grip height, so such a stance exists only below a 1.225 m grip on the G1 (1.883 m on the H1), and the G1 at 1.2 m keeps 1.6 N to spare. The braked spec holds (0.011 mm at a 5 N drag), one H1 pose loses its controller, and the G1’s worst no-hose stack falls 1.3 points.
Milestone 42 splits that loss with a two-by-two of leans and controller sides: the lean costs all 1.30 points and the four flipped leg-joint sides none, though the old sides cannot hold one re-leaned pose; compared like for like, the H1 loses 0.09–0.29 points per re-leaned pose.
Milestone 43 recosts it in the spec’s own plan, which re-plans the leg preload and squeeze for each stance: the G1 then gains (worst stack 0.9160 to 0.9257 braked, 0.9081 to 0.9257 stock), so the 1.3-point loss belonged to the controller’s torques, while the H1 keeps a plan at only three of its six re-leaned poses, two because the stance allows only 0.033 and 0.095 N m of leg preload at a 50 N press.
Milestone 44 gives those two poses a per-pose leg preload of 0.9 of what the stance allows (0.030 and 0.085 N m): both get a plan and their stacks do not move with the preload’s size, but with a 0.2 N sensor error the H1’s worst residual along the wall rises from 0.019 to 0.172 mm, all at (1.4 m, −0.15 m), where the right hip roll, a joint the controller holds, flips; (1.6 m, 0) is rescued at 0.003 mm.
Milestone 45 gives the held joints their own margin first: the stance caps that hip roll at 0.0339 N m however much the other legs give up, so the residual stays 0.172 mm; the joint moves 0.163 N m per newton of sensor z error, so the pose meets the spec only with a z error under about 0.19 N (an ATI Mini45-class sensor, not an FT 300).
Milestone 46 corrects that: the per-robot hold side pinned the right hip roll at its threshold, while the controller wants −3.8 N m; held on the other side the stance allows 2.29 N m, the pose takes the full 0.1 N m preload and leaves 0.005 mm (from 0.172) with an FT 300-class 0.2 N error, so the H1 meets the tool-point spec at all 19 poses at no stack cost; the per-robot side is also wrong at four more H1 poses.
Milestone 47 gives every pose its own hold sides (the combination with the largest zero-drag margin): every pose of both robots clears the preload goal (smallest 0.886 N m), two H1 poses with no plan gain one, and the H1's worst residual at 20 N of hose drag falls from 0.032 to 0.010 mm, but the G1's rises from 0.011 to 0.037 mm (a leg joint flips under drag at three poses), so sides must be scored at the loads the robot meets.
Milestone 48 scores them over 17 hose loads: the G1 returns to 0.016 / 0.011 / 0.011 mm at 0 / 5 / 20 N of drag and the H1 keeps 0.010 / 0.014 / 0.010 mm at no stack cost, but at 8 of 34 poses no side has a plan with 20 N pulling backward at the 50 N press, and the G1's worst residuals since milestone 41 cover only the 83% of swept cases that have a plan.
Milestone 49 finds those missing plans are a stance limit, not a solver fault: at a 50 N press with 20 N of hose pulling back, no contact force balances the base (the robot tips back), and the drag lowers the top of the press window by about 15 N (G1 51.6 to 36.4 N, H1 58.7 to 42.8 N).
Milestone 50 finds no forward lean fixes it: the best keeps the worst friction-cone slack at −0.69 to −10.12 N under 17 hose loads, because it trades backward-drag headroom for forward; the G1 at 1.2 m grips tolerates 4.0 N of hose pulling back (about 3 m of rubber hose dragging, floor friction unsourced), so the hose must be fed or carried.
Milestone 51 finds those drag limits are necessary but not sufficient: every swept case past them has no plan, and inside them the G1 plans 93.0% of its cases (83.1% overall) and the H1 99.7%, in spec, but 75 of the G1’s 96 remaining failures come from its one calibration pose under sideways drag, so the calibration should be taken with the hose slack.
Milestone 52 does that: calibrating under the hanging hose’s weight alone lifts the G1 to 98.1% inside the limits with no residual change, because once the compensator subtracts the measured hose load the calibration absorbs only the sensor’s bias.
Milestone 53 finds the last failures are a leg-preload limit (the pose’s hold sides leave 0.004 to 0.092 N m of leg margin against 0.1), and choosing the hold sides per measured hose load gives both robots a plan at every case inside the drag limits (G1 1368 of 1368, H1 1902 of 1902) with no residual change.
Milestone 54 moves off the 8-direction grid: at 48 Latin-hypercube hose loads both robots still plan every in-envelope case (9438), but 4 leave the 0.03 mm spec (worst 0.284 mm), each a joint sitting at the 0.2 N m friction threshold that the measured-load plan holds and the true load frees, in direction windows under 5° wide that the grid stepped over.
Milestone 55 plans only when every joint sits further than its worst settle change under the 0.2 N sensor bound from that threshold: every planned case is then in spec (worst 0.011 mm) with no held/free flip, but coverage falls to 95.9 to 96.8%, mostly at the G1 waist and H1 torso, and the cases it drops were in spec anyway.
Milestone 56 weights that margin by what each flip costs at the nose (the sum of each in-band joint’s backlash shift through the wall row of the Jacobian, held to half the spec): every planned case stays in spec (worst 0.018 mm) across three load sets, coverage returns to 100% on the G1 and 99.6 to 99.9% on the H1, and the bound is never beaten in 28 158 planned cases.
Milestone 57 sweeps that share on a held-out load set: half the spec replicates (worst 0.018 mm, coverage 99.9% G1, 99.5% H1), coverage does not move between a zero budget and half the spec, the whole spec puts one H1 press at 0.0315 mm, and every flip the budget admits is a knee or ankle pitch joint that moves the nose 0.27 to 0.29 mm across the wall, the direction the bound leaves out.
Milestone 58 prices each flip through both in-plane rows of the Jacobian (along and across the wall, half a 0.03 mm spec each): on a second held-out load set every planned press stays inside 0.03 mm in both directions (worst across the wall 0.022 mm, from 0.27 to 0.29 mm under the one-row bound) at unchanged coverage, 100% G1 and 99.63% H1.
Milestone 59 adds the rows of the depth point 44.4 mm along the nail: on a third held-out set the two-row bound already keeps it inside 0.03 mm (worst 0.022 mm, tilt at most 0.034 mrad) and a four-row bound plans the same presses. Statics only.
Show 117 earlier entries
2026-11-19 · Milestone 59 planned: does the two-row bound also keep the nail's depth point (44.4 mm past the nose) in spec, or does the press's angle need its own rows?
- Why. Milestone 58 certified the nose in y and z; its "Not measured" says the depth point, the second point of
compensated (tool_lean.DEPTH_CHECK = sheathing + required penetration, 44.4 mm along the nail), is not budgeted. A flip that rotates the tool about the nose moves the depth point and not the nose, so the two-row bound cannot see it. - Model. Milestone 58's plan unchanged except the bound: each joint's flip cost through four rows of
backlash.tool_rows, the nose (y, z) and the depth point (y<sub>d</sub>, z<sub>d</sub>, offset L0 + 44.4 mm), giving Y<sub>y</sub>, Y<sub>z</sub>, Y<sub>yd</sub>, Y<sub>zd</sub>. The squeeze minimises the largest Y/b; a plan passes if every Y ≤ its b; the side search as milestone 58. Two configs: two-row (b<sub>y</sub> = b<sub>z</sub> = 0.015 mm, depth unbudgeted; milestone 58's "iso") and four-row (all four 0.015 mm). The depth spec 0.03 mm is the nose spec carried over, not sourced. Section F, held out: 48 Latin-hypercube loads (seed 59), 6 random-direction 0.2 N sensor errors (seed 59). Check: the two-row config against milestone 58's own rule_all ("iso") on the first 2 loads of each robot. Also reported: the tool's tilt, (depth − nose)/44.4 mm, in mrad. - Hypotheses, registered before the run. - H232 (the depth gap). Under the two-row rule on seed 59 every planned case is inside 0.03 mm at the nose in y and z, and at least one planned case on at least one robot is over 0.03 mm at the depth point in y or z. Refuted if either part fails. - H233 (four-row bound). Under the four-row rule every planned in-envelope case is inside 0.03 mm at both points in both directions on both robots, and all four bounds hold (each Y at least the change the robot's flips make to that residual, within 1e-9 mm). Refuted if any fails. - H234 (the depth rows are cheap). Four-row coverage is within 1.0 percentage point of two-row coverage on each robot. Refuted if either robot loses more.
2026-11-19 · Milestone 58 result: pricing each flip across the wall as well as along it keeps every planned press inside 0.03 mm in both directions (worst z 0.022 mm, from 0.27 to 0.29 mm) at no coverage cost (H229, H230 and H231 held)
experiments/m58_tworow/ (run, results, figure), reproduce.py (m58 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL. No new library code (preload.flip_cost takes any Jacobian row), so no new tests. Run time 914 s on 4 processes. One 2-load smoke run (M58_N=2) ran after the hypotheses were committed and before the full run; it is not reported.
- Check. With b<sub>z</sub> = ∞, on the first 4 loads of each robot (756 cases), milestone 56's own
weighted_rule and this run agree on every plan/no-plan and how, and |y| differs by 0.0 mm. - The z gap replicates (H231 held). Section E (48 Latin-hypercube loads, seed 58; 6 random-direction 0.2 N errors, seed 58), y-only rule: worst |y| G1 0.0109, H1 0.0117 mm, none over spec; worst |z| G1 0.2713, H1 0.2865 mm, with 4 G1 and 5 H1 planned presses over 0.03 mm across the wall.
- Two rows (H229 held). Isotropic budgets (b<sub>y</sub> = b<sub>z</sub> = 0.015 mm): worst |y| G1 0.0109, H1 0.0100 mm; worst |z| G1 0.0220, H1 0.0176 mm; none over 0.03 mm in either direction; Y<sub>y</sub> and Y<sub>z</sub> each at least the shift the robot's flips made (smallest slack 0.0 mm in both).
- No coverage cost (H230 held). Coverage G1 4002 of 4002 (100%), H1 5392 of 5412 (99.63%) under both rules. The z row moves plans rather than dropping them: per-load side combinations rise from 250 to 311 (G1) and 94 to 129 (H1), and planned presses with a flip fall from 34 to 30 and 20 to 15. The H1's 20 unplanned cases are the same count; under the two-row rule 13 bind on the left knee and 7 on the right ankle.
- The loose z budget (reported, not registered). b<sub>z</sub> = 0.15 mm (z spec 0.3 mm) gives the same plans and residuals as the isotropic budget on every summary line: the squeeze and side search find a plan with no z-costly flip whenever one exists, so the budget level does not bind here.
- What it means. The y-only bound's 0.27 to 0.29 mm across the wall was not a price of coverage: an alternative hold-side or squeeze with no z-costly flip existed in every case, and the one-row objective could not see the difference. With both rows, the bound certifies 0.03 mm in both in-plane directions at the same coverage as y alone.
- Not measured. One held-out seed (9414 in-envelope cases); statics; constant sensor bias; μ 0.5 not measured; the z spec is the y spec carried over, not sourced; depth (the second point of
compensated) not budgeted. - Next. Preload or re-pose for the H1's left knee and right ankle (the 20 cases neither rule plans); or carry the depth point's rows into the bound, so the press's angle is certified as well as its position.
2026-11-19 · Milestone 58 planned: does a two-row flip bound (along the wall y and across it z, each with its own budget) keep every planned press inside spec in both directions on a held-out seed, and what coverage does the z row cost?
- Why. Milestone 57 found the y-only bound admits flips of knee and ankle pitch joints that barely move the nose along the wall and move it 0.27 to 0.29 mm across it (z); its "Next" asked to put both Jacobian rows into the bound.
- Model. Milestone 56's plan unchanged except the bound: each joint's flip cost through the wall row (Y<sub>y</sub>, as before) and through the z row of
backlash.tool_rows at the nose (Y<sub>z</sub>, same reachable held states). The squeeze minimises max(Y<sub>y</sub>/b<sub>y</sub>, Y<sub>z</sub>/b<sub>z</sub>) (ties as milestone 56); a plan passes if Y<sub>y</sub> ≤ b<sub>y</sub> and Y<sub>z</sub> ≤ b<sub>z</sub>; the side search runs when a leg joint has a cost in either row or the pose's sides have no plan, and takes the smallest objective among passing combinations. b<sub>y</sub> = 0.5 × 0.03 mm throughout. Three budgets for z: none (b<sub>z</sub> = ∞, which is milestone 56's rule), isotropic (z spec 0.03 mm, b<sub>z</sub> = 0.015 mm) and loose (z spec 0.3 mm, b<sub>z</sub> = 0.15 mm, reported, not registered). No source gives a z spec for the press; 0.03 mm is the y spec carried over, and 0.3 mm is the size milestone 57 measured unbudgeted. Section E, held out: 48 Latin-hypercube loads (seed 58), 6 random-direction 0.2 N sensor errors (seed 58). Check: with b<sub>z</sub> = ∞, on the first 4 loads of each robot, milestone 56's own weighted_rule must give the same plan/no-plan and |y|. - Hypotheses, registered before the run. - H229 (isotropic two-row bound). Every planned in-envelope pose-case is inside 0.03 mm along the wall and 0.03 mm across it on both robots, and both bounds hold (Y<sub>y</sub> and Y<sub>z</sub> each at least the change the robot's flips make to |y| and |z|, within 1e-9 mm). Refuted if any fails. - H230 (the z row is cheap). Coverage inside milestone 51's envelope under the isotropic rule is within 1.0 percentage point of the y-only rule on each robot. Refuted if either robot loses more. - H231 (the z gap replicates). Under the y-only rule on seed 58 every planned case is inside 0.03 mm along the wall, and at least one planned case on at least one robot is over 0.03 mm across it. Refuted if either part fails.
2026-11-19 · Milestone 57 result: on a held-out seed the β = 0.5 rule replicates (H226 held) and coverage is flat from β = 0 to 0.5 (H228 held), but β = 1 puts one H1 press over spec (0.0315 mm; H227 refuted), and every flip the budget allows is a leg pitch joint that moves the nose 0.27 to 0.29 mm across the wall, the direction no rule budgets
experiments/m57_beta/ (run, results, figure), reproduce.py (m57 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card. No new library code, so no new tests. Run time 768 s on 4 processes. One 2-load smoke run (M57_N=2) ran after the hypotheses were committed and before the full run; it is not reported.
- Check. On the first 4 loads of each robot (780 cases), milestone 56's own
weighted_rule at β = 0.5 and the one-pass evaluation agree on every plan/no-plan and how, and the residual differs by 0.0 mm. - Replication (H226 held). Section D (48 Latin-hypercube loads, seed 57; 6 random-direction 0.2 N errors, seed 57), β = 0.5: worst |y| G1 0.0105 mm, H1 0.0178 mm; coverage G1 4017 of 4020 (99.93%), H1 5375 of 5400 (99.54%); smallest Y − flip shift −6.9e-18 mm (rounding).
- The frontier. Coverage (G1 / H1) and worst |y|: β 0: 99.93% / 99.54%, 0.0105 / 0.0100 mm; 0.25: same coverage, 0.0105 / 0.0175 mm; 0.5: same, 0.0105 / 0.0178 mm; 0.75: 99.93% / 100%, 0.0278 / 0.0178 mm; 1.0: 99.93% / 100%, 0.0278 / 0.0315 mm. The G1's 3 unplanned cases have no plan on any side at any β; the H1's 25 all bind on the left knee and come back only at β = 0.75. Coverage at 0.25 equals coverage at 0.5 on both robots (H228 held), and so does β = 0.
- β = 1 (H227 refuted). One H1 press, pose (1.2 m, −0.15) at 12.21 N, 182.6 degrees, error 1, leaves 0.0315 mm along the wall. The bound held there (the flip shift stayed under Y); the spec was lost because Y used all 0.03 mm and the windup and calibration terms add on top, as the half-and-half split assumed.
- Across the wall (reported, not registered). The rule budgets only y. Worst |z| of a planned press: β 0: G1 0.026, H1 0.019 mm; β 0.25: 0.267, 0.281 mm; 0.5: 0.277, 0.289 mm; 1.0: 0.278, 0.525 mm. Every planned press with |z| over 0.03 mm (β 0.5: 4 G1 and 8 H1 of 9392) has a held or side flip between plan and robot, on a knee or ankle pitch joint (H1's right ankle, G1's knees and ankle pitch): at these poses those joints move the nose little along the wall and a lot in z. The β = 1 H1 worst, 0.525 mm, is two flips with the left knee in band.
- What it means. On this seed the budget buys nothing until 0.75: β = 0 (plan only when no joint whose flip moves the nose along the wall sits in its band) plans exactly as many presses as 0.5, with the same or smaller worst |y| and a tenth of the worst |z|. The flips the budget admits are cheap in y and expensive in z, so a y-only bound reads them as free. The fix is a bound on the full in-plane shift: each joint's cost through both the y and z rows (or their norm), with a z budget, which is milestone 56's construction with one more Jacobian row.
- Not measured. What spec the z direction needs (whether 0.3 mm in z matters depends on what the press is aligning to); a second held-out seed (one seed, 48 loads, 9420 in-envelope cases, so the 25-case H1 gap at β ≤ 0.5 is 0.46% with wide error); statics only; constant sensor bias; μ 0.5 not measured.
- Next. Put both rows into the bound (Y<sub>y</sub> and Y<sub>z</sub>, each with its own budget) and rerun sections A to D at β = 0 and 0.5; or preload the H1 left knee, the only joint that still binds at β ≤ 0.5.
2026-11-19 · Milestone 57 planned: on a held-out seed, where does the budget share β of milestone 56's weighted rule put coverage against the worst residual, and does β = 1 (the whole spec for flips) still keep every press in spec?
- Why. Milestone 56 set β = 0.5 before its run (half the 0.03 mm spec for backlash flips, half for windup and the calibration) and did not tune it; its "Not measured" and "Next" asked for a sweep on a held-out seed to place the coverage-accuracy frontier.
- Known before the run (from reading the code, no number computed). In
m56_weighted/run.py the squeeze is chosen to minimise Y whatever the budget, and among side combinations that pass the smallest Y wins, which is the combination with the smallest Y overall when that one passes. So one pass per case (Y and residual on the pose's sides, and the smallest-Y side combination with its residual) gives the rule's plan at every β: pose sides if Y<sub>0</sub> ≤ β · 0.03 mm, else the best combination if it passes (searched only when the pose-side plan has a leg cost or no plan, as milestone 56 does), else none. Coverage is then non-decreasing in β by construction; the worst residual need not be monotone. - Model. Milestone 56's setup and rule unchanged (its run imported), with β in {0, 0.25, 0.5, 0.75, 1.0}. Section D, held out: 48 new Latin-hypercube loads (seed 57) with 6 random-direction 0.2 N sensor errors (seed 57), not seen by milestones 54 to 56. Check: at β = 0.5 on the first 4 loads of each robot, milestone 56's own
weighted_rule is called and must give the same plan/no-plan and residual. Also reported, not registered: the residual across the wall (z) of the planned cases, which no rule budgets. - Hypotheses, registered before the run. - H226 (held-out replication, β = 0.5). Every planned in-envelope pose-case is inside 0.03 mm along the wall, coverage inside milestone 51's envelope is at least 99% on each robot, and the flip shift never beats Y (by more than 1e-9 mm). Refuted if any of the three fails. - H227 (β = 1). With the whole spec given to flips, every planned in-envelope pose-case is still inside 0.03 mm on both robots. Refuted if any is larger. - H228 (a flat frontier below 0.5). Coverage at β = 0.25 is within 0.5 percentage points of coverage at β = 0.5 on each robot. Refuted if either robot loses more.
2026-11-19 · Milestone 56 result: weighting the held-or-free margin by each flip's cost at the nose keeps every planned press in spec (worst 0.0178 mm; H223 held), wins back every press milestone 55 dropped on the G1 and all but 24, 6 and 12 on the H1 (coverage 99.6 to 100%; H224 held), and the flip bound is never beaten (H225 held)
experiments/m56_weighted/ (run, results, figure), src/buildbot_sheathing/preload.py (reachable_states, flip_cost), tests/test_flipcost.py (6 tests: zero cost far from the threshold, half and full gap, and the bound against corner and 5000 uniform torque draws), reproduce.py (m56 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card. Run time 1710 s on 4 processes. One 2-load smoke run (M56_N=2) ran after the hypotheses were committed and before the full run; it is not reported.
- The rule. Each joint's worst nose shift from a backlash flip, c<sub>j</sub> = |w<sub>j</sub>| × the largest change of its
held_bias term over the held states its torque band [τ̂ − r<sub>j</sub>, τ̂ + r<sub>j</sub>] can reach at the two press ends, with w the wall row of the nose Jacobian; the plan passes when Y = Σ c<sub>j</sub> ≤ 0.015 mm (half the 0.03 mm spec). Milestone 55's rule is the zero-budget case. - Spec (H223 held). Every planned in-envelope pose-case in all three sections is inside 0.03 mm: worst G1 0.0107 / 0.0110 / 0.0125 mm and H1 0.0178 / 0.0147 / 0.0089 mm in sections A / B / C (milestone 55: G1 0.0107 / 0.0110, H1 0.0099 / 0.0091). The worst H1 case grows by 0.008 mm: that is the budget being spent.
- Coverage (H224 held). G1 4002 of 4002, 3924 of 3924 and 3990 of 3990 (100%); H1 5412 of 5436 (99.6%), 5418 of 5424 (99.9%), 5412 of 5424 (99.8%). Paired case by case with milestone 55 on sections A and B, the weighted rule plans every case milestone 55 planned and wins back 637 of the 667 it dropped (G1 308 of 308, H1 329 of 359), worst residual of a won-back case 0.0147 mm.
- The bound (H225 held). Over 28 158 planned cases, 156 have a held or side flip between plan and robot (the rule now allows them where they are cheap), and on every planned case the flip shift (|y| minus |y| with the plan's held states) is at most Y (smallest Y − shift −2.6e-18 mm, rounding). The bound is tight: on single-joint flips the shift equals Y to rounding (largest shift/Y 1.000), and some flips reduce |y| (shift down to −0.008 mm), since the triangle inequality ignores sign.
- What still binds. The 42 H1 cases left: 32 bind on the left knee (one or two joints in band, Y 0.0166 to 0.0246 mm, over the budget) and no side combination of 16 passes; 10 have the right elbow in band at 0.167 mm, which the hold sides cannot move (not side-searched). The G1 waist and H1 torso, milestone 55's binding joints, never bind here: of the 302 planned cases with Y > 0 the largest cost is on a knee or ankle every time.
- What it means. On these robots most joints sit so that half a backlash gap barely moves the nose along the wall; the waist and torso cost nothing at the budget. A robot can compute the weighted rule from the measured load alone (three extra settle solves for G and one Jacobian row), and it gives the spec guarantee for the backlash term with 0.4% or less of presses lost. The budget split (half the spec for flips, half for windup and calibration) was set before the run, not tuned.
- Not measured. Whether β = 0.5 is the best split (a sweep would trade the H1's last 0.4% against the worst residual); the time the side search takes on a robot; a sensor error that is not a constant bias; the residual across the wall (z) was not given a budget, only y; statics only; μ 0.5 not measured for these soles.
- Next. Sweep β on a held-out seed to place the coverage-accuracy frontier, and add the z residual to the bound; or preload the H1 knee and elbow (the only joints still binding) through the stance or the squeeze direction.
2026-11-19 · Milestone 56 planned: weight the held-or-free margin by what a flip costs at the nose: does a budget on the bound of the flips' nose shift keep every press in spec and win back milestone 55's 3 to 4% of dropped presses?
- Why. Milestone 55's robust rule asks every joint's planned torque to sit further than r<sub>j</sub> = B||G<sub>j</sub>|| from the 0.2 N m friction torque. That fixed milestone 54's misses but dropped 3.2 to 4.1% of in-envelope presses, mostly on the G1 waist and H1 torso, and post hoc none of the dropped cases was out of spec under the old rule. Its "Next" asked for a margin weighted by what a flip costs at the nose.
- Known before the run (from reading the code, no number computed). The nose residual along the wall is linear in the joint error (
pose_range.nose_bias: y = w · δ with w the wall row of backlash.tool_rows at the nose), and a joint's backlash term is −sign(τ<sub>hi</sub>) gap/2 when it is held at both window ends, else 0 (pose_range.held_bias). With the settled torque inside [τ̂<sub>w</sub> − r<sub>j,w</sub>, τ̂<sub>w</sub> + r<sub>j,w</sub>] at each end w, the set of held states the robot can reach (held +, held −, free) can be listed, so each joint's worst flip cost c<sub>j</sub> = |w<sub>j</sub>| · max |hb − hb̂| over that set is exact for independent end intervals and conservative otherwise; the sum Σ c<sub>j</sub> bounds the backlash part of the residual by the triangle inequality. Joints outside their band cost 0, so milestone 55's rule is the special case of a zero budget. - Model. Milestone 55's setup unchanged (its run imported). The weighted rule, from the measured load only: (1) on the pose's sides, choose the squeeze on milestone 38's grid over [−S, S] to minimise the flip bound Y = Σ<sub>j</sub> c<sub>j</sub> (ties: the larger smallest robust arm margin, then the smaller |squeeze|); accept if Y ≤ β · 0.03 mm with β = 0.5 (half the spec left for windup and the calibration, whose worst in milestone 55 was 0.011 mm); (2) otherwise, if some leg joint has a non-zero cost, every hold-side combination with its own squeeze, taking the smallest Y among those that pass; (3) otherwise no plan. The calibration plan keeps milestone 55's strict rule (it passed at every error there). Sections A and B: milestone 55's (seed 54 loads on the 6 sensor axes; seed 55 loads with 6 random 0.2 N errors). Section C, fresh: 48 new Latin-hypercube loads (seed 56) with 6 random-direction 0.2 N errors (seed 56).
- Hypotheses, registered before the run. - H223 (spec). Under the weighted rule every planned in-envelope pose-case has a residual along the wall of at most 0.03 mm on both robots in all three sections. Refuted if any is larger. - H224 (coverage). Inside milestone 51's envelope the weighted rule plans at least 99% of pose-cases on each robot in every section (milestone 55: 95.9 to 96.8%). Refuted if any robot in any section falls below 99%. - H225 (the bound holds). At every planned case, the robot's residual minus the residual it would have with the plan's held states (backlash flips removed) is at most the plan's Y, so the bound is never beaten. Refuted at any case where it is (by more than 1e-9 mm).
2026-11-19 · Milestone 55 result: a held-or-free margin covering the 0.2 N sensor error clears milestone 54's spec misses (H220 held) but drops 3.2 to 4.1% of in-envelope cases (H221, H222 refuted), most at the waist, which nothing preloads, and none of the dropped cases was out of spec under the old rule
experiments/m55_robust/ (run, results, figure), reproduce.py (m55 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card. No new library code, so no new tests. Run time 2182 s on 4 processes. Two 2-load smoke runs (M55_N=2) ran after the hypotheses were committed and before the full run; the second checked a faster replica of the plan against the first and matched it; neither is reported.
- Checks. The replica of milestone 53's
plan (one hold QP and leg-preload solve per window end, then linear settles) matches plan to 0.0 N m at 3158 of 3158 cases checked (error 0 of every load and pose), with no plan/no-plan mismatch. The settled torque at −B along x matches −G<sub>x</sub>B to 2.1e-13 N m, so Δτ is linear in the load error as the code implied. The sensor-error radius r<sub>j</sub> = B||G<sub>j</sub>|| reaches 0.27 N m (median of each plan's largest 0.21 to 0.22 N m), more than the 0.2 N m friction torque itself: on those joints the margin can only be met by holding them hard, not by leaving them free. - Spec (H220 held). On milestone 54's 48 loads at its 6 sensor axes, every planned in-envelope case is inside spec: worst G1 0.0107 mm, H1 0.0099 mm (milestone 54: 0.284 and 0.042 mm). No planned case on either section has a held or side flip between plan and robot (0 of 18 119), so the backlash part of the error is gone, as the construction says. All 4 of milestone 54's misses now plan in spec: the G1 one on the pose's sides with a re-chosen squeeze (robust margin 0.018 N m, 0.0091 mm), the 3 H1 ones on per-load sides (0.0029 to 0.0058 mm).
- Coverage (H221 refuted). Inside milestone 51's envelope the robust rule plans G1 3856 of 4002 (96.4%) and H1 5253 of 5436 (96.6%), against 100% for milestone 54's rule. 434 G1 and 302 H1 cases need per-load sides. The slack calibration passes the rule at every sensor error on both robots.
- Fresh loads, off-axis errors (H222 refuted). On 48 new loads (seed 55) with 6 random-direction 0.2 N errors: G1 3762 of 3924 (95.9%), H1 5248 of 5424 (96.8%), worst 0.0110 and 0.0091 mm. Milestone 54's rule on the same loads plans 100% and is also inside spec (worst G1 0.0292 mm, H1 0.0147 mm), so this seed holds no miss for the rule to fix.
- What binds. Of the 667 dropped in-envelope cases, the binding joint is a G1 waist joint (yaw, roll or pitch) at 291 of 308 and the H1 torso at 281 of 359; the rest are leg joints (knees mostly) and 5 H1 elbow cases, and 8 have no plan at all on any side. The hold sides move only the ground reaction and so only the leg torques, and the squeeze has zero net wrench, so neither moves a waist torque: those cases were not side-searched. Every one of the 125 cases that was side-searched had 0 of 16 or 64 combinations passing.
- The rule is conservative (post hoc, not registered). On the fresh loads, every case the robust rule drops was inside spec under milestone 54's rule (worst G1 0.0110 mm, H1 0.0147 mm, none above 0.02 mm). A waist that the plan holds and the robot frees moves the nose less than the spec here, so asking every joint for the full margin costs 3 to 4% of coverage to fix errors that, on these joints, would not have mattered.
- What it means. The fix for milestone 54's misses is a margin on the held-or-free threshold sized from the sensor bound and the settle map, and it works by construction for the backlash term. Applied to every joint it is too strict: the waist and torso, which no preload reaches, sit inside the band about 3% of the time. The rule a robot should run is a margin weighted by what a flip costs at the nose.
- Not measured. The time the side search takes on a robot (up to 64 small QPs per press); a sensor error that is not a constant bias; statics only; μ 0.5 not measured for these soles.
- Next. Weight the margin by the flip's cost: require it only on joints where half a backlash gap, through the wall-direction Jacobian, would move the nose by more than a set share of the spec; or give the waist a preload (a small torso lean or a hold moment through the tool) and re-run both seeds.
2026-11-19 · Milestone 55 planned: does a held-or-free margin that covers the sensor error on every joint, arms included, close milestone 54's spec misses, and what coverage does it cost?
- Why. Milestone 54 found 4 of 9438 planned in-envelope pose-cases over the 0.03 mm spec (worst 0.284 mm), each one joint whose planned torque sat a few hundredths of a N m from the 0.2 N m friction torque, so the plan (from the measured load) and the robot (under the true load) disagreed on whether that joint is held. Its "Next" asked for a margin that covers the sensor error on every joint.
- Known before the run (from reading the code, no number computed).
preload.settle is linear in the load (a pseudo-inverse correction), so the change from planned to settled torque is linear in the sensor error: Δτ<sub>j</sub> = G<sub>j</sub> δ for a load error δ. G can be read from three extra calls of milestone 53's plan at the same measured load with the truth moved along x, y and z. If every joint's planned |τ| at both window ends is further than r<sub>j</sub> = B · ||G<sub>j</sub>|| from the 0.2 N m friction torque (B = 0.2 N, the sensor bound already used), no error in that ball can change any joint's held status or side, so the backlash part of the joint error is zero by construction; the windup part is not touched. What is unknown is how many plans pass, and whether the squeeze and side choice can make the rest pass. - Model. Milestone 54's setup unchanged (its run imported: milestone 48's sides, 'braked' gaps, |e| 0.2 N, S 10 N, slack calibration, per-load side rule, the drag envelope bisected at the sampled direction). The robust rule, run from the measured load only: (1) plan on the pose's sides and re-choose the squeeze on milestone 38's grid over [−S, S] to maximise the smallest robust margin ||τ<sub>j</sub>| − 0.2| − r<sub>j</sub> over the arm joints (inside each joint's torque limit, as milestone 38 does); accept if the robust margin is non-negative on every actuated joint at both window ends; (2) otherwise search every hold-side combination (64 G1, 16 H1) with the same squeeze step and take the one with the largest smallest robust margin among those that pass; (3) otherwise no plan. The calibration plan (slack hose) is held to the same rule. Section A: milestone 54's 48 loads (seed 54) at its 6 sensor axes. Section B, a fresh test: 48 new Latin-hypercube loads (seed 55) with 6 sensor errors of 0.2 N in random directions (seed 55), not on the axes the rule never sees anyway.
- Hypotheses, registered before the run. - H220 (spec, milestone 54's loads). Under the robust rule every planned in-envelope pose-case has a residual along the wall of at most 0.03 mm on both robots. Refuted if any is larger. - H221 (coverage). Inside milestone 51's envelope the robust rule plans at least 99% of pose-cases on each robot (milestone 54: 100%). Refuted if either robot falls below 99%. - H222 (fresh loads, off-axis errors). On the seed-55 loads and errors, every planned in-envelope pose-case is inside 0.03 mm and coverage is at least 99% on each robot. Refuted if either fails.
2026-11-18 · Milestone 54 result: at 48 Latin-hypercube hose loads every in-envelope press still has a plan (H217 held), but 4 of 9438 leave the 0.03 mm spec, up to 0.284 mm, in direction windows 1 to 5 degrees wide the 8-direction grid never sampled (H218 refuted); the grid's drag envelope is up to 2.2 N optimistic between directions (H219 refuted)
experiments/m54_lhs/ (run, results, figure, post hoc), reproduce.py (m54 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card. No new library code, so no new tests. Run time 461 s on 4 processes (post hoc 99 s). A 2-load smoke run (M54_N=2) ran after the hypotheses were committed and before the full run; it is not reported.
- Reproduction. The bisection run at the 8 grid directions gives milestone 50's limits exactly (largest difference 0.00 N, no None mismatch), so the sampled limits come from the same test.
- The grid envelope between directions (H219 refuted). At 21 (pose, direction) pairs the drag limit at the sampled direction is more than 0.5 N below the smaller of its two grid neighbours, all on the G1 at the 0.4 to 0.8 m grips with the hose pulling back and sideways (186 to 250 degrees); the largest shortfall is 2.17 N ((0.4 m, −0.15) at 202.6 degrees: 26.44 N against 28.61 N). None matters at the drags swept here: every dip is at a limit of 23 N or more, above the 20 N the loads reach, and the poses whose limits bind (G1 1.2 m, 4.0 to 4.2 N; H1 1.6 m, 12.0 N) match the grid within 0.05 N, because their weakest direction is straight back, a grid direction. The H1 has no dip over 0.5 N.
- Coverage (H217 held). Inside milestone 51's envelope at the sampled directions the G1 plans 4002 of 4002 pose-cases (3943 on the pose's hold sides, 59 by per-load sides) and the H1 5436 of 5436 (5379 and 57). Inside the pose-only envelope: G1 4056 of 4056, H1 5436 of 5436. Of the 300 pose-cases past their pose's limit (G1 264, H1 36) none plans on the pose's sides (they were not side-searched), as milestone 51 found on the grid. The slack calibration planned at all 288 (load, sensor axis) pairs.
- Accuracy (H218 refuted). 4 planned in-envelope pose-cases exceed 0.03 mm: G1 (0.4 m, −0.15) at 13.23 N, 290.2 degrees, sensor axis 2: 0.284 mm; H1 (1.2 m, −0.15) at 4.75 N, 305.5 degrees: 0.042 and 0.034 mm (axes 0 and 2); H1 (1.2 m, −0.15) at 12.62 N, 210.5 degrees: 0.032 mm. All 4 are planned on the pose's own sides. Worst by drag band, G1 0.016 / 0.023 / 0.284 / 0.011 mm and H1 0.042 / 0.013 / 0.032 / 0.015 mm over [0, 5), [5, 10), [10, 15), [15, 20] N.
- Why (post hoc, not registered;
posthoc.txt). Each excess is one joint whose torque at one window end sits at the 0.2 N m friction torque at which the model counts a joint as held on one side of its backlash, so the plan (made from the measured load) and the robot (under the true load) disagree on whether it is held, and the error is half a backlash gap (7.3e-4 to 7.5e-4 rad on that joint, every other joint unchanged). G1: the right shoulder yaw, planned −0.226 N m at 50 N, settled −0.200 N m (the arm squeeze is −0.05 N). H1: the left hip pitch (planned −0.100 N m at 25 N, settled −0.273) and the left knee (planned +0.300 at 50 N, settled +0.151). A ±6 degree sweep at 0.25 degree steps puts the over-spec points inside spans 2.75 to 4.75 degrees wide (the G1's is broken: 10 of the 16 points in its 3.75 degree span; 0.009 mm outside it); with grid directions 45 degrees apart, a window that narrow is easy to miss. Applying milestone 53's per-load side choice anyway brings all 3 H1 cases inside spec (0.0011 to 0.0058 mm, 3.8 to 5.6 N m of leg margin), but not the G1 one (0.284 mm unchanged: the joint is an arm joint, not a hold or leg joint). - What it means. Coverage survives the move off the grid; the spec does not, quite. The failure is not a plan limit but a threshold the plan sits on: the leg preload asks a margin over the friction torque, but the plan uses the measured load, so a joint the plan puts a few hundredths of a N m past the threshold can land on the other side under the true load. On the joints that tip over, planned and settled torque differ by 0.026 to 0.173 N m, more than the distance from the threshold the plan left them.
- Not measured. How often these windows occur over a continuous load distribution (48 loads give 4 of 9438 in-envelope cases, 0.04%, an estimate with wide error); whether the H1 dips past 0.5 N at directions not sampled. Statics only; constant sensor bias; μ 0.5 not measured.
- Next. Ask the plan for a held-or-free margin that covers the sensor error on every joint, arms included: no joint's planned torque within |Δτ(e)| of the 0.2 N m threshold at either window end (or the squeeze and side choice that keeps it out), then rerun these 48 loads and a fresh LHS seed.
2026-11-18 · Milestone 54 planned: off the 8-direction grid, do the three hose rules still give every in-envelope press a plan inside spec, and is the grid's drag envelope safe between its directions?
- Why. Milestones 50 to 53 sampled hose drag at 0, 5 and 20 N in 8 directions 45 degrees apart, and the drag limit each pose tolerates (milestone 50) was bisected at those 8 directions only. Milestone 53's "Next" asked to drop the grid: sample loads continuously inside the envelope and check the three rules (slack calibration, each press inside its pose's drag limit, hold sides chosen per measured load) hold off the 8 directions.
- Known before the run. Nothing below has been computed. The envelope between grid directions is unknown: the tolerated drag could dip between two grid directions, so an envelope read off the grid could admit loads no plan exists for.
- Model. Milestone 53's setup unchanged (its run imported: milestone 48's sides, 'braked' gaps, |e| 0.2 N on the 6 sensor axes, S 10 N, slack calibration, per-load side rule). Hose loads: 48 Latin-hypercube points over drag H in [0, 20] N (the swept range) and direction φ in [0, 360) degrees, seed 54, the same 48 for both robots, each at the 6 sensor axes and every pose (G1 15, H1 19: 4320 and 5472 pose-cases). At each sampled φ and each pose (and the calibration pose) the drag limit is bisected exactly as milestone 50 did (largest H in [0, 60] N at which both 25 and 50 N balance, current lean, 25 steps). Envelope: milestone 51's (H at most the limit at the pose and at the calibration pose, at the sampled φ), and the pose-only one (the pose's limit only). Rule: plan on the pose's sides; if none, milestone 53's per-load side search. Residual along the wall with the slack calibration.
- Hypotheses, registered before the run. - H217 (coverage). Inside milestone 51's envelope evaluated at the sampled directions, the three rules give a plan at every LHS pose-case on both robots. Refuted if any in-envelope pose-case has none. - H218 (spec). Every planned in-envelope LHS pose-case has a residual along the wall of at most 0.03 mm. Refuted if any is larger. - H219 (grid envelope is safe). At every sampled direction and pose, the drag limit bisected there is at least the smaller of the limits at the two neighbouring grid directions minus 0.5 N (so an envelope read off the grid by taking the smaller neighbour never admits a load more than 0.5 N past the true limit). Refuted if any pose at any sampled direction falls more than 0.5 N below.
2026-11-18 · Milestone 53 result: choosing the hold sides per measured load gives both robots a plan at every pose-case inside milestone 51's envelope (G1 1368 of 1368, H1 1902 of 1902), in spec; the leg preload, not the hold QP, is what failed; H214 and H215 refuted, H216 held
experiments/m53_loadside/ (run, results, figure; imports milestone 52's run, no change to it), reproduce.py (m53 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card. No new library code, so no new tests. Run time 117 s on 4 processes.
- The leg preload binds, not the hold QP (H214 refuted). At 20 of the 26 plan-limit pose-cases the first stage that fails is the leg-preload solve: the hold QP solves on the pose's sides, but the largest common leg margin it allows (
margin_lp) is below the 0.1 N m asked (G1 0.004 to 0.092 N m; H1 0.019 to 0.062 N m against its tiered 0.0864 / 0.1 N m). The binding leg joints at the optimum are ankle rolls and hip rolls and yaws: G1 (1.2 m, 0) at 20 N φ 45 the left ankle roll at the 25 N end (6 of 6), (1.0 m, 0) at 20 N φ 90 the right ankle roll at the 50 N end (3 of 3). Only 6 fail at the hold QP, all at (1.0 m, +0.15) at 20 N φ 225 at the 50 N end, where 32 or 33 of the 64 side combinations still have a hold solution. Of the 7 sensor-error cases, 5 fail at the preload and 2 at the hold QP. - Per-load sides clear every case inside milestone 51's envelope. The rule (keep the pose's sides when they plan, else take the side combination with the largest common leg margin at the measured load among those whose full plan solves) recovers 31 of the 33 drops. Between 8 (H1, of 16) and 52 (G1, of 64) combinations plan at each case, and the chosen one's common leg margin is 1.84 to 5.22 N m on the G1 and 2.47 to 3.27 N m on the H1, against 0.1 N m asked. Most G1 recoveries flip two joints of milestone 48's sides, the left hip yaw and left ankle roll (
--++-+ to -+-+-+, 15 of 26); the H1 flips its right hip roll (4 of 5). - Coverage. Inside milestone 51's envelope the G1 now plans 1368 of 1368 pose-cases (100%, from 98.1%) and the H1 1902 of 1902 (100%, from 99.7%).
- In spec (H216 held). Every recovered case's residual along the wall is at most 0.0106 mm (H1 (1.6 m, 0) at 5 N straight back), so each robot's worst inside the envelope is unchanged: G1 0.016 / 0.011 / 0.011 mm, H1 0.010 / 0.014 / 0.010 mm at 0 / 5 / 20 N.
- What is left (H215 refuted). The 2 G1 sensor-error cases at (0.8 m, −0.15) at 20 N φ 180, which are inside only the pose-only envelope (the drag passes the calibration pose's limit, so they need the slack calibration to count at all): at the 50 N end no side combination of 64 has a hold solution, so no side choice can help there (a hold-QP limit at that measured load; why it fails was not examined).
- What it means. With the three rules together (calibrate with the hose slack; keep each press inside its pose's drag limit; pick the hold sides at each press from the measured load), every swept case inside the drag envelope has a plan on both robots and stays inside the 0.03 mm spec. The per-pose side choice of milestones 46 to 48 was the right idea at too coarse a grain: one side set per pose cannot hold under every hose load, but one per measured load can.
- Not measured. Whether switching hold sides between presses costs time or wear (each switch moves the joints across their backlash); the side search is 16 to 64 small QPs per press, run offline here. Statics only; μ 0.5 not measured for these soles; constant sensor bias only.
- Next. Drop the sweep's grid: sample hose loads continuously inside the envelope (Latin hypercube over drag and direction) and check the three rules hold off the 8 directions; then the cost of side switches (backlash crossings per wall).
2026-11-18 · Milestone 53 planned: at the 33 pose-cases still without a plan under the slack calibration, which stage fails (hold sides or leg preload), and does choosing the hold sides per measured load clear them?
- Why. Milestone 52 left 33 pose-cases with no plan inside an envelope: 26 inside milestone 51's (G1 21 plan limits at four (pose, load) cells, 5 sensor-error cases; H1 5 plan limits at (1.6 m, 0), 5 N straight back) and 2 G1 sensor-error cases inside the pose-only envelope. Its "Next" asked which hold-side or leg-preload row binds and whether milestone 48's side search, run per load rather than per pose, clears them.
- Known before the run (from reading the code, no number computed).
plan (milestone 38, with milestones 44 to 48's patches) fails at one of two stages at either window end (25 or 50 N): the 'hold' QP on the pose's hold sides has no solution, or the leg-preload solve (solve_margin, tiered at milestone 45's poses) has none. The plan is made from the measured load f + e, so a side rule keyed on the measured load is one the robot can run. Nothing below has been computed. - Model. Milestone 52's setup (its run imported unchanged: milestone 48's sides, 'braked' gaps, |e| 0.2 N on 6 sensor axes, S 10 N, slack calibration). Section A: at each dropped (pose, load, sensor axis), the failing stage and window end, and the largest common leg margin
margin_lp reaches on the pose's sides. Section B, the per-load rule: keep the pose's sides when they plan; otherwise search all 2<sup>n</sup> hold-side combinations (64 G1, 16 H1) and take the one with the largest loaded common leg margin at that measured load among those whose full plan solves. Section C: residual along the wall (slack calibration) for every case the rule recovers, and each robot's coverage and worst residual inside milestone 51's envelope. Cases that already plan are untouched by the rule, so their residuals are milestone 52's. - Hypotheses, registered before the run. - H214 (stage). At most of the 26 plan-limit pose-cases (more than 13) the 'hold' QP on the pose's sides is the failing stage, not the leg preload. Refuted if 13 or fewer. - H215 (per-load sides). The per-load rule finds a plan at every one of the 33 dropped pose-cases. Refuted if any stays without one. - H216 (spec). Every recovered case has a residual along the wall of at most 0.03 mm, so each robot's worst inside milestone 51's envelope stays at most 0.03 mm at 0, 5 and 20 N. Refuted if any is larger.
2026-11-18 · Milestone 52 result: calibrating with the hose slack lifts the G1 to 98.1% of in-envelope cases (from 93.0%) and changes no residual; H211, H212 and H213 held
experiments/m52_slackcal/ (run, results, figure; imports milestone 51's run, no change to it), reproduce.py (m52 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card. No new library code, so no new tests. Run time 208 s on 4 processes.
- The slack calibration always plans (H213 held). Under the hanging hose's weight alone the calibration pose plans at all 6 sensor-error axes on both robots, so no case is dropped by the calibration.
- Coverage (H211 held). Inside milestone 51's envelope the G1 plans 1342 of 1368 pose-cases (98.1%; milestone 51's loaded calibration 93.0%), 1394 of 1530 overall (91.1%, from 83.1%). The H1 is unchanged: 1897 of 1902 (99.7%). Inside a pose-only envelope (the calibration no longer carries drag, so only the pose's own limit applies) the G1 plans 1394 of 1422 (98.0%).
- No residual changes (H212 held). Over every pose-case that plans under both calibrations, |y| differs by less than 5e-5 mm (prints as 0.0000), and the worst residuals are milestone 51's: G1 0.016 / 0.011 / 0.011 mm, H1 0.010 / 0.014 / 0.010 mm at 0 / 5 / 20 N. Why, checked after the run:
wrist_ft.joint_error is actual minus predicted error, and the compensator predicts from the measured load, so with braked gaps and linear windup the error left at the calibration pose depends on the sensor error e only. At the G1's calibration pose the slack and 5 N φ 90 calibration errors differ by 4e-19 rad on joints whose error is about 1e-5 rad. - What is left (inside milestone 51's envelope). G1: 21 plan limits at the pose itself (none with an exact sensor either): the calibration pose (1.0 m, 0) as a press at 20 N φ 90 (3 axes), (1.0 m, +0.15) at 20 N φ 225 (6), (1.2 m, 0) at 20 N φ 45 (6), (1.2 m, +0.15) at 5 N φ 225 (6); and 5 sensor-error cases, (1.2 m, −0.15) at 5 N φ 0 (3) and (1.0 m, 0) at 5 N φ 270 (2). H1: 5 plan limits at (1.6 m, 0) at 5 N straight back. Two more G1 sensor-error cases at (0.8 m, −0.15) 20 N φ 180 lie only in the pose-only envelope.
- What it means. The hose rule for both robots is: calibrate once with the hose slack, then keep each press inside its pose's drag limit (milestone 50). In this model that costs no accuracy, because the calibration's job is the sensor bias, not the load.
- Not measured. A sensor gain (scale) error, or a windup that is not linear in torque, would make the calibration error load-dependent; neither is modelled. Statics only; μ 0.5 not measured for these soles.
- Next. The 26 remaining plan limits at the 1.0 to 1.6 m grips: which hold side or leg-preload row binds, and whether a per-load side choice (milestone 48's search, run per load rather than per pose) clears them.
2026-11-18 · Milestone 52 planned: does calibrating with the hose slack recover the G1's in-envelope coverage, and do residuals stay in spec when the calibration load differs from the press load?
- Why. Milestone 51 found 101 dropped pose-cases inside milestone 50's drag envelope; 75 of the G1's 96 are its calibration pose (1.0 m, 0) failing under drag, which drops all 15 poses at once. Its "Next" asked to calibrate with no hose load (the calibration is one move, so the hose can be slack for it) and rerun the sweep.
- Known before the run (from reading the code, no number computed).
residuals (milestone 38) calibrates at the reference pose under the same hose force as each pose's press. Here the calibration move takes the slack-hose force (hose_f(ref, 0, 0), the hanging weight only) with the same sensor error, and each pose keeps its loaded press. The compensation then subtracts a calibration error measured under a different load, so the residual can grow even where coverage improves. Nothing below has been computed. - Model. Milestone 48's sides, preload and sweep (its post hoc's import), 'braked' gaps, |e| 0.2 N on the 6 sensor axes, S 10 N, at 0, 5 and 20 N of drag in 8 directions; milestone 51's envelope (drag at most milestone 50's tolerated drag at the pose and at the calibration pose) and a pose-only envelope (the calibration no longer carries drag). Section A: per case, coverage and worst residual along the wall under the slack calibration. Section B: per pose-case, the residual change from milestone 51's loaded calibration where both plan. Section C: every remaining in-envelope drop, re-run with an exact sensor.
- Hypotheses, registered before the run. - H211 (coverage). Under the slack calibration the G1 plans more than 98% of its pose-cases inside milestone 51's envelope (milestone 51: 93.0%). Refuted if 98% or less. - H212 (spec). Inside milestone 51's envelope each robot's worst residual along the wall is at most 0.03 mm at 0, 5 and 20 N under the slack calibration. Refuted if any is larger. - H213 (no calibration drops). The slack calibration plans at all 6 sensor-error axes on both robots, so no case is dropped by the calibration. Refuted if any axis fails.
2026-11-18 · Milestone 51 result: every swept case outside milestone 50's drag envelope has no plan, but 101 inside it do not either; the largest block is the G1's calibration pose under 20 N of sideways drag; inside the envelope both robots stay in spec; H208 refuted, H209 and H210 held
experiments/m51_envelope/ (run, results, post hoc, figure; imports milestone 48's post hoc and reads milestone 50's results, no change to either), reproduce.py (m51 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card. No new library code, so no new tests. Run time 237 s on 4 processes.
- Same sweep as milestone 48. The run drops 258 of the G1's 1530 pose-cases and 41 of the H1's 1938, exactly milestone 48's post hoc counts.
- The envelope is necessary. A pose-case is inside when the drag is at most milestone 50's tolerated drag (25 and 50 N both balance) at that direction at the pose and at the calibration pose (1.0 m, 0). Every pose-case outside it has no plan: 162 of 162 on the G1, 36 of 36 on the H1. Inside it the G1 plans 1272 of 1368 (93.0%, against 83.1% of all its cases), the H1 1897 of 1902 (99.7%).
- It is not sufficient (H208 refuted). 101 dropped pose-cases are inside it (post hoc,
posthoc.txt, classes chosen after the run): on the G1, 45 from the calibration pose with no plan at 20 N pulling sideways (φ 90, 3 of 6 sensor-error axes, still no plan with an exact sensor) at 35% of the drag it balances; 30 from the calibration pose at 5 N pulling the other way (φ 270), which plan with an exact sensor; 18 where the pose itself has no plan with an exact sensor either ((1.2 m, 0.15) at 5 N φ 225, 76% of its limit; (1.2 m, 0) at 20 N φ 45, 66%; (1.0 m, 0.15) at 20 N φ 225, 78%); 3 from sensor error at (1.2 m, −0.15). On the H1, 5 at (1.6 m, 0) with 5 N straight back (42% of its limit), no plan with an exact sensor. "Plan limit" means the stance balances but the hold sides and leg preload have no solution. - One calibration couples every pose. A calibration-pose failure drops all 15 G1 poses at once: 75 of the 96 in-envelope G1 drops.
- In spec inside the envelope (H209 held). Worst residual along the wall: G1 0.016 / 0.011 / 0.011 mm, H1 0.010 / 0.014 / 0.010 mm at 0 / 5 / 20 N. These are milestone 48's figures: every case that plans is inside the envelope, so restricting the sweep removes no planned case.
- The 5 N calibration drops are sensor error (H210 held). Both 5 N cases whose calibration fails plan there with an exact sensor.
- What it means. Milestone 50's per-pose drag limits are the right hose rule as far as balance goes: no plan exists past them. What they leave is a planning limit, not a balance limit, and most of it sits at one pose, the G1's calibration pose under sideways drag, because one calibration serves every pose. A hose rule for the G1 needs the calibration taken with the hose slack (or at a pose whose plan holds sideways drag) as well as the per-pose drag limits.
- Not measured. μ_f, so the drag limits in metres of hose (milestone 50's post hoc); which hold side or preload row fails at the plan-limit cases; statics only, μ 0.5 not measured for these soles.
- Next. Calibrate with no hose load (the calibration is one move, so the hose can be slack for it) and rerun the sweep: does the G1's in-envelope coverage go from 93.0% to above 98%, and do the residuals stay in spec with the calibration made under a different load from the press? Then the plan-limit cases at the 1.0 to 1.2 m poses: which hold row binds.
2026-11-18 · Milestone 51 planned: which of milestone 48's dropped sweep cases are hose-drag stance limits, and does keeping the hose inside each pose's drag limit restore coverage and spec?
- Why. Milestone 48's post hoc found section D's residual sweep drops 258 of the G1's 1530 pose-cases (41 of the H1's 1938), and milestone 50 found the drag each pose tolerates (its section A). Its "Next" asked for a hose-management rule with section D's sweep restricted to the loads each pose tolerates, coverage reported.
- Known before the run (from reading the code, no number computed).
residuals calibrates once at milestone 30's pose (1.0 m, 0) and returns nothing for the whole case if that pose has no plan, so a calibration-pose failure drops every pose of the robot. Milestone 48's post hoc counted the same 2 drops at every G1 pose at 5 N and 9 at 20 N, which looks like calibration-pose drops; milestone 50 has the calibration pose tolerating 17.0 N or more in every direction, so the 5 N drops are not explained by its drag limit. Nothing below has been computed. - Model. Milestone 48's sides, preload and sweep (its post hoc's import), 'braked' gaps, |e| 0.2 N on the 6 sensor axes, S 10 N, at 0, 5 and 20 N of drag in 8 directions. Section A, every case: whether the calibration pose plans, whether each pose plans, and its residual. Each pose-case is in the envelope when the drag is at most milestone 50's tolerated drag at that direction at both the pose and the calibration pose. Section B, every dropped case re-run with an exact sensor (e = 0) and, for a calibration failure, the reason (
plan at the calibration pose). - Hypotheses, registered before the run. - H208 (drops are stance limits). Every pose-case milestone 48's sweep drops lies outside the envelope. Refuted if any dropped pose-case is inside it. - H209 (in spec inside the envelope). Over the pose-cases inside the envelope, each robot's worst residual along the wall is at most 0.03 mm at 0, 5 and 20 N. Refuted if any is larger. - H210 (the 5 N drops are sensor error). At every case dropped at 5 N, the same load with an exact sensor plans at the calibration pose. Refuted if any does not.
2026-11-18 · Milestone 50 result: no forward lean keeps the eight high-grip poses balanced under 20 N of hose drag; the best lean tolerates 11.4 to 19.5 N in every direction, so the hose, not the lean, is the fix; H204 and H205 refuted, H206 held, H207 not tested
experiments/m50_relean/ (run, results, post hoc, figure; imports milestone 49's run, no change to it), reproduce.py (m50 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card. No new library code, so no new tests. Run time 152 s on 4 processes.
- A bug caught before reading results. The first run's section A tested "balances" as "milestone 49's
cone_slack returns a value"; that LP returns a negative slack, not None, when the cones cannot hold, so every pose read as tolerating 60 N and a 200 N press. Section A now uses milestone 49's balance-and-cones feasibility LP (as its post hoc did). Sections B to D did not use that test and printed the same numbers in both runs. - The drag each pose tolerates (section A, current lean). The largest hose tension at which 25 and 50 N both balance, per direction: the G1 at 1.2 m grips tolerates 4.0 to 4.2 N straight back (6.3 to 6.6 N at φ 135 and 225), the G1 at 1.0 m 12.9 to 17.0 N, the H1 at 1.6 m 12.0 N. Every one of the 26 poses that had a side plan in milestone 48 tolerates at least 20.1 N in every direction, so a 20 N sweep sat just inside them. Forward drag (φ 0) is the next limit: 21.5 to 22.5 N at the G1's 0.8 to 1.2 m grips.
- The rule is linear (H206 held). At φ 180 the tolerated drag is within 9% of (press top with a slack hose − 50 N) / 0.78 at all 8 poses (1 to 9%). The press top here includes the hanging hose's weight (H 0), so it reads 52.9 N where milestone 49's no-hose figure was 51.6 N.
- No lean covers the firing window under drag (H204 refuted). On a 5 mm grid of com_dx from 0 to 150 mm and a ternary refinement, the worst friction-cone slack over the 17 loads at 25 and 50 N peaks below zero at every pose: −0.69 N (G1 1.0 m, dy 0 and +0.15, com_dx 90.7 mm), −2.38 N (G1 1.0 m, −0.15), −3.47 N (H1 1.6 m, 113 mm), −10.12 N (G1 1.2 m, 110 mm). Leaning forward gains backward-drag headroom at about 0.53 N of slack per mm, and past the peak a 20 N hose pulling towards the wall at the 25 N press takes over (all three robots and grips): the robot tips forward.
- Standing at zero press as well is further out (H205 refuted). With 0 N added to the window, the best worst slack is −19.61 N (G1 1.0 m), −22.18 N (H1), −32.34 N (G1 1.2 m): milestone 41's zero-press stance and a 20 N hose cannot coexist at these grips.
- H207 not tested. It was conditional on a lean from H204.
- At the best lean (section B2) the smallest tolerated tension over the 8 directions is 14.9 to 19.5 N (G1 1.0 m), 11.4 to 13.4 N (G1 1.2 m), 17.0 N (H1 1.6 m): the lean roughly triples the G1 1.2 m figure but stays under 20 N.
- As a hose length (post hoc,
posthoc.txt, after the run, not registered). A dragged hose holds at most μ_f w g L of tension (hose.max_drag). For the rubber 3/8 in hose (0.274 kg/m) at μ_f 0.5, the G1's 1.2 m grips at the current lean reach their limit with 3.0 to 3.1 m of hose dragging; at the best lean 8.5 to 10.0 m; the H1 at 1.6 m 8.9 m (current) and 12.7 m (best). The hybrid hose roughly doubles each length. μ_f is milestone 37's unsourced sweep (0.3 / 0.5 / 0.8). - What it means. The eight poses' missing plans are not a lean to tune: no lean holds a 20 N drag in every direction, because what a forward lean gives to backward drag it takes from forward drag. The fix is in the hose: at the G1's 1.2 m grips keep the dragged length under about 3 m of rubber hose (or feed slack, a boom or an overhead reel), or lean per load direction if the controller knows it. The residual tables' G1 coverage caveat stands.
- Not measured. μ_f for any hose on OSB or slab; dynamic loads when the hose snags; leaning per load direction (needs the direction, which milestone 35 says no controller knows); the side search at a re-lean (H207); statics only, μ 0.5 not measured for these soles.
- Next. A hose-management rule in the plan: per pose, the dragged length allowed (or a reel tension set point), and section D's residual sweep restricted to the loads each pose tolerates, with coverage reported; or a lean chosen per hose direction from a direction estimate (the wrist F/T sensor sees the hose force), with its sensing error.
2026-11-18 · Milestone 50 planned: can a re-lean keep the eight high-grip poses balanced under 20 N of hose drag, and how much drag does each pose tolerate?
- Why. Milestone 49 found the poses milestone 48 could not plan are a stance limit: a backward hose lowers the top of the press window by about 0.75 to 0.8 N per newton of drag, and milestone 41 had leaned these poses back (com_dx 0.123 to 0.085 m on the G1 at 1.2 m) so they stand at zero press, leaving the top at 51.6 to 62.2 N. Its "Next" asked for a re-lean under drag, or the largest drag each pose tolerates as a hose-routing rule.
- Known before the run. For a fixed pose and hose load the press forces that balance are an interval (the balance-and-cones set is a polyhedron in press and contact forces), so checking 0, 25 and 50 N decides a window. Nothing below has been computed.
- Model. Milestone 48's import (M1), μ 0.5, statics. Section A, every pose of milestone 48's set at its current lean: for each of the 8 hose directions, the largest hose tension H (bisected on [0, 60] N, assumed monotone, checked at the found H) at which both 25 and 50 N balance. Section B, the 8 poses with no S_H: com_dx on a 5 mm grid from 0 to 0.15 m (IK via
buildbot_payload.model.pose), the worst cone slack over the 17 loads (0 N; 5 and 20 N in 8 directions) at 25 and 50 N, and separately at 0, 25 and 50 N (standing at zero press under every load, milestone 41's requirement). Section C, at each pose's best section B lean for [25, 50] N, milestone 48's side search (every combination, its loaded_margin). - Hypotheses, registered before the run. - H204 (a lean covers the firing window under drag). Each of the 8 poses has a com_dx on the grid at which 25 and 50 N both balance under all 17 loads. Refuted if any has none. - H205 (one lean does both). Each of the 8 has a com_dx at which 0, 25 and 50 N all balance under all 17 loads. Refuted if any has none. - H206 (the drag rule is linear). At φ 180 at the 8 poses, the tolerated drag from section A is within 15% of (top with no hose − 50 N) / 0.78, milestone 49's slope. Refuted if any is outside. - H207 (sides exist at the new lean). At every pose where H204 holds, some side combination has a plan at all 17 loads and both ends with common leg margin at least 0.1/0.9 N m. Refuted if any has none.
2026-11-18 · Milestone 49 result: the backward-hose cells are a stance limit, not a solver false negative; 20 N of backward drag lowers the top of the press window by about 15 N, below the 50 N firing press; H199 and H200 refuted, H201 to H203 not tested
experiments/m49_backload/ (run, results, post hoc, figure; imports milestone 48's run, no change to it), reproduce.py (m49 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card. No new library code, so no new tests. Run time 105 s on 4 processes (a first attempt lost 20 min to BLAS threads oversubscribing the pool: numpy was imported before milestone 48's OMP_NUM_THREADS=1 default; the script now sets it first).
- A correction to milestone 48's post hoc. Its table 1 prints, per load, "hold solutions / margin LP solutions" at the 25 N end, then at the 50 N end. The entry above read the 25 N pair (e.g. 48/48) as the hold QP solving and the "0/0" as the margin LP failing. In fact both solvers agree everywhere: they both solve at the 25 N end and neither has a solution at the 50 N end. There was no disagreement to explain.
- No balancing force at 50 N (H199 refuted). At the 50 N end of all 24 cells, no contact force balances the base inside the μ 0.5 friction cones: the best cone slack is −23.77 to −0.10 N. At the 25 N end all 24 balance (slack at least 6.98 N).
- No false negative (H200 refuted). The hold QP has no solution at any 50 N cell, so there is no hold force to test. At the 25 N end the margin LP solves at all 984 hold solutions, which have base residual at most 1.7e-13 and cone violation at most 2.0e-9. Section B (a fallback for false negatives, then a rerun) was conditional on one, so H201 to H203 were not tested: the dropped section D cases are real stance limits, and milestone 48's coverage caveat stands.
- What limits the stance (post hoc,
posthoc.txt, after section A, not registered). The net horizontal load the soles carry is at most 0.189 of the vertical load (G1; H1 0.127), well inside μ 0.5, so the stance does not slide as a whole. 13 of the G1's 18 cells do not balance with unlimited friction either: the press plus a backward hose pulls the centre of pressure off the heels, so the robot tips back. 3 G1 cells with a diagonal or sideways hose balance at μ 0.503 to 0.563, just above 0.5 (yaw friction). The other 2 G1 cells and all 6 H1 cells balance only with unbounded friction (none up to μ 5), which is also a centre-of-pressure limit for practical purposes. - The press window shrinks under drag. Bisecting the largest press that balances at μ 0.5: G1 51.6 to 62.2 N with no hose, 36.4 to 49.8 N with the cell's drag; H1 58.7 N, 42.8 to 48.6 N. Straight back at 20 N (φ 180) the top falls by 15.2 N (G1 1.2 m grips, 51.6 to 36.4 N), 14.8 N (G1 1.0 m, 62.2 to 47.4) and 15.9 N (H1 1.6 m, 58.7 to 42.8): about 0.75 to 0.8 N of press per newton of backward drag. Milestone 32 leaned each pose to cover [25, 50] N with no hose, and at these grips the top had only 1.6 to 12 N to spare.
- What it means. The poses milestone 48 could not plan are not a numerical artefact to fix in the solver; they are poses whose lean leaves less headroom above 50 N than a backward hose takes away. The fix is in the pose or the hose: lean those poses forward so the window covers [25, 50] N plus the drag, or route the hose so it does not pull backward at the high grips.
- Not measured. The bottom of the press window under drag (a forward lean costs headroom there); the G1's calibration pose, whose drops come from the same check but were not diagnosed here; statics only; μ 0.5 not measured for these soles on OSB or slab.
- Next. Re-lean the 8 poses (milestone 32's lean search) for the window [25, 50] N under the worst backward drag, check the 25 N end still balances at the worst forward drag, and rerun section D with the coverage reported; or find the largest backward drag each pose tolerates, as a hose-routing rule.
2026-11-18 · Milestone 49 planned: is the backward-hose infeasibility a stance limit or a solver false negative?
- Why. Milestone 48's post hoc found 24 (pose, hose load) cells, all at the 50 N press end and mostly 20 N of drag pulling backward, where the 'hold' QP returns a solution for 37 to 56 of the G1's 64 side combinations (15 to 16 of the H1's 16) but the preload margin LP (
preload.margin_lp) has none for any. The plan path (preload.solve_margin) calls the same LP, so section D drops those cases: 258 of the G1's 1530 pose-cases (17%), 41 of the H1's 1938. Its "Next" asked which it is. - What is known before the run (from reading the code, no number computed).
grf_qp.solve('hold') returns a force only after grf_qp.margin_lp (balance plus friction cones plus the hold rows) is feasible, and falls back to that LP's point when SLSQP's is not clean, so the stance can balance at those cells. preload.margin_lp adds only rows t <= kappa tau + d with t free in [−1000, 1000] N m, which a balancing force satisfies unless some leg torque passes 1000 N m. So a true infeasibility there is unexpected; nothing here has been run to check it. - Model. Milestone 48's import and sides. Section A, at every cell of milestone 48's post hoc table 1 and both window ends, for every side combination the hold QP solves: the hold force's base residual and cone violation,
preload.margin_lp's HiGHS status and message, the common leg margin evaluated at the hold force itself (an explicit feasible point of the margin LP when it is finite), and a re-solve of the same LP (scaled rows, t bound widened). Section B, if section A finds a false negative, preload.margin_lp patched to fall back on the best explicit point (never a looser constraint), then milestone 48's side search and section D rerun on both robots with the drop counts beside the worst residuals. - Hypotheses, registered before the run. - H199 (the stance balances). At every one of the 24 cells, a balance-and-cones LP has a solution. Refuted if any lacks one. (A check: the hold QP needed it.) - H200 (a false negative). At every cell, for at least one side combination, the hold QP's force has base residual and cone violation below 1e-6 and a finite common leg margin, so the margin LP was feasible. Refuted if any cell has none. - H201 (coverage returns). With the fix, the G1 drops under 5% of its swept pose-cases (milestone 48: 17%) and the H1 none (2%). Refuted otherwise. - H202 (the G1 stays in spec on the full sweep). The G1's worst residual along the wall, over the recovered sweep, is at most 0.03 mm at 0, 5 and 20 N of drag. Refuted if any is larger. - H203 (the H1 stays in spec on the full sweep). The same for the H1. Refuted if any is larger.
2026-11-18 · Milestone 48 result: sides scored under hose drag put the G1 back in spec (0.016 / 0.011 / 0.011 mm) and keep the H1's gain (0.010 / 0.014 / 0.010), but at 8 of 34 poses no side holds at every load, and the post hoc finds the G1's worst residuals since milestone 41 skip 17% of the swept cases; H193, H194, H195 and H197 held, H196 and H198 refuted
experiments/m48_loadside/ (run, results, post hoc, figure; imports milestone 47's run with its side rule replaced), reproduce.py (m48 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card. No new library code, so no new tests. Run time 248 s on 4 processes (a branch-and-bound prune stops scoring a side once it falls below the best so far).
- The rule. At each pose, S_H is the side combination with the largest common leg margin over both window ends and 17 hose loads (0 N, 5 and 20 N in 8 directions), exact sensor; ties go to the larger zero-force margin (amended before the run, above). It is milestone 47's S* at 3 of 9 G1 poses and 14 of 17 H1 poses where it is defined.
- The G1 is back in spec (H193, H194 held). Worst residual along the wall ('braked', |e| 0.2 N, S 10 N): 0.016 / 0.011 / 0.011 mm at 0 / 5 / 20 N drag (milestone 47: 0.011 / 0.025 / 0.037; milestone 41: 0.016 / 0.011 / 0.011). Milestone 47's three bad poses ((0.6 m, −0.15 m), (1.2 m, 0.15 m), (1.2 m, 0)) are fixed: (0.6 m, −0.15 m) by S_H, the two 1.2 m poses because they fall back to the per-robot side (below). The 0 N figure gives back milestone 47's gain at (1.2 m, 0.15 m) for the same reason.
- The H1 keeps its gain (H195 held). 0.010 / 0.014 / 0.010 mm, the same as milestone 47.
- Not every pose has a side (H196 refuted). At 8 poses (G1 at grip heights 1.0 and 1.2 m, all three lateral offsets; H1 at 1.6 m, 0 and +0.15 m) no combination of the 64 (16) has a plan at every load, so S_H is undefined and the pose keeps milestone 46's sides. Every pose that has S_H has a loaded margin of at least 0.2000 N m (the τ_f saturation), above the 0.111 goal.
- The stack does not move (H197 held). Worst 'braked' G1 0.9257, H1 0.9655, as milestone 47; every changed pose within 0.01 points.
- Lost cases return (H198 refuted). H1 cases at milestone 44's two poses with no plan: 23 of 204 (milestone 47: 18), all five new ones at (1.6 m, 0) at 5 N, which reverts to milestone 46's side.
- Post hoc (
posthoc.txt, after the run, not registered). (1) At the 8 poses the loads no side can hold are 20 N with the hose pulling backward (φ 135 to 270, 1 to 4 directions; and 5 N at φ 180 at three G1 1.2 m poses), always at the 50 N press end: the 'hold' QP returns a solution for 37 to 56 of the 64 G1 combinations (15 to 16 of 16 H1), but the margin LP over the same contact model has none for any. Whether that is a stance limit (press plus backward drag beyond what the feet can react) or a difference between the two solvers' tolerances is not resolved here. (2) A coverage caveat for every G1 worst residual since milestone 41. Section D's sweep drops a case whose plan has no solution. On milestone 48's sides the G1 drops 258 of 1530 pose-cases (17%): 2 of 48 at 5 N and 9 of 48 at 20 N at every pose (the calibration pose has no plan, so the whole case goes), and up to 27 of 48 at 20 N at the 1.2 m poses. The H1 drops 41 of 1938 (2%), all at (1.6 m, 0) and (1.6 m, 0.15 m). The G1's "0.011 mm at 20 N" is the worst over the cases that have a plan, not over the sweep; milestones 41 and 47 printed no G1 miss count, so their G1 figures carry the same caveat. - What it means. Scoring hold sides at the loads the robot meets is the right score where it is defined: it fixes milestone 47's G1 regression without losing the H1's gain, at no stack cost. But a backward-pulling 20 N hose at a 50 N press leaves no plan at the G1's high grips and the H1's 1.6 m grips whatever the sides, and on the G1 also at its calibration pose; the residual table should report coverage beside the worst case.
- Not measured. Statics; sensor error in the side score; mu 0.5 and the 'braked' waist; the two H1 poses milestone 38 skipped; why the hold QP and margin LP disagree at the backward loads.
- Next. Resolve the backward-load infeasibility (solver tolerance or stance limit: check the hold QP's base-balance residual and the friction cone at those loads); report each robot's coverage beside its worst residual; if it is a stance limit, a press force or stance that keeps a plan under 20 N of backward drag, or a hose-routing rule that keeps the drag forward.
2026-11-18 · Milestone 48 planned: score each pose's hold sides by the worst common leg margin over the hose sweep
- Why. Milestone 47 picked each pose's hold sides S* by the common leg margin at zero hose force. That rescued the H1 (worst residual 0.032 to 0.010 mm at 20 N drag) but pushed the G1 over the 0.03 mm spec at 20 N (0.011 to 0.037 mm): at three G1 poses a leg joint, not a hold joint, flips under drag on S*. Its "Next" asked for sides scored at the loads the controller will meet.
- Nothing computed with the new score before registering. Every number below that predicts is a hypothesis; no loaded margin has been computed for any side combination.
- Model. Milestone 47's import (so M1 is the object milestones 41 to 47 built), with its side rule replaced. At every pose of both robots in milestone 41's set, S_H is the combination (16 for the H1, 64 for the G1) whose smallest common leg margin over both window ends and all 17 hose loads (0 N, and 5 and 20 N in each of 8 directions, the same loads section D sweeps), with an exact sensor, is largest; ties (1e-9 N m) keep the per-robot side. A pose takes S_H only if its zero-force margin reaches 0.1/β (β 0.9), so it plans with the spec's own 0.1 N m leg preload as in milestone 47; otherwise it keeps milestone 46's sides and plan. The score uses only the hose loads, not the sensor errors, and only margins, not residuals, so section D is not fitted to. Sections: A, per pose, per-robot sides, S* (milestone 47), S_H, the zero-force and loaded margins of each; B, the spec stack ('stock', 'braked') at every pose whose sides change, beside milestone 47's; C, each robot's worst spec stack; D, each robot's worst residual along the wall ('braked', |e| 0.2 N, S 10 N, H 0 / 5 / 20 N, every hose direction and sensor axis), beside milestone 47's and milestone 41's (G1) figures.
- Amended before the full run (after the code smoke check, 3 H1 poses, H 0 and 5 N only; hypotheses unchanged). The loaded score saturates at τ_f (0.2000 N m): whenever any leg joint's nominal torque sits inside the friction band at some load,
margin_rows gives it the two rows τ_f ∓ τ, so many combinations tie at exactly 0.2. "Ties keep the per-robot side" would then pick by enumeration order. Ties (within 1e-9 N m) now go to the larger zero-force margin (milestone 47's score), then to the per-robot side. - Hypotheses, registered before the run. - H193 (the G1 is back in spec). The G1's worst residual is at most 0.03 mm at 0, 5 and 20 N drag. Refuted if any is larger. - H194 (the G1 loses nothing to the per-robot side). The G1's worst residual is at most milestone 41's 0.016 / 0.011 / 0.011 mm plus 0.005 mm. Refuted if any is larger. - H195 (the H1 keeps milestone 47's gain). The H1's worst residual is at most milestone 47's 0.010 / 0.014 / 0.010 mm plus 0.005 mm. Refuted if any is larger. - H196 (every pose carries the preload under load). Every pose of both robots has an S_H loaded margin of at least 0.1/0.9 N m. Refuted if any is smaller. - H197 (sides cost no stack). 'braked': each robot's worst spec stack is at least milestone 47's (H1 0.9655, G1 0.9257) minus 0.1 points. Refuted if lower. - H198 (no more lost cases). The H1's swept cases with no plan at milestone 44's two poses are at most milestone 47's 18 of 204. Refuted if more.
2026-11-18 · Milestone 47 result: per-pose hold sides lift every pose's margin above the preload goal and add the H1's two skipped poses, and they lower the H1's worst residual at 20 N drag from 0.032 to 0.010 mm, but the G1 goes over the 0.03 mm spec at 20 N (0.011 to 0.037 mm) because the side is chosen at zero hose force; H187, H188, H189, H191 and H192 held, H190 refuted
experiments/m47_poseside/ (run, results, post hoc, figure; imports milestone 46's run with its single flip replaced by a per-pose rule), reproduce.py (m47 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card. No new library code, so no new tests.
- The rule. At each pose, S* is the hold-side combination (16 for the H1, 64 for the G1) with the largest common leg margin at zero hose force and an exact sensor, the smaller window end. It differs from the per-robot side at every pose of both robots (15 of 15 G1, 19 of 19 H1), though at several G1 poses only by a tie-level amount (G1 (0.6 m, 0): 3.6116 to 3.6128 N m).
- Every pose takes the spec preload (H187 held). Smallest S* margin 0.886 N m (H1 (1.4 m, 0)), against the 0.111 goal; the H1's margins rise from 0.033–4.21 to 0.886–12.79 N m, the G1's from 0.141–4.69 to 0.975–5.31. Milestone 45's tier is no longer needed anywhere.
- The two skipped H1 poses gain a plan (H188 held). (0.4 m, −0.15 m) and (1.6 m, −0.15 m), which milestone 38 skipped for having no 'hold-min' solution on the per-robot sides, have S* margins of 12.79 and 6.62 N m. They are not in the swept set yet: adding them needs milestone 38's pose set rebuilt.
- The H1 improves (H189, H192 held). 'braked', |e| 0.2 N, S 10 N: the H1's worst residual along the wall is 0.010 / 0.014 / 0.010 mm at 0 / 5 / 20 N of hose drag (milestone 46: 0.010 / 0.019 / 0.032). The swept cases with no plan at milestone 44's two poses fall from 23 to 18 of 204 (none at 5 N, from 5).
- The stack does not move (H191 held). Every pose's spec stack is within 0.01 points of milestone 46's; worst 'braked' G1 0.9257, H1 0.9655, unchanged.
- The G1 gets worse under hose drag (H190 refuted). The G1's worst residual is 0.011 / 0.025 / 0.037 mm at 0 / 5 / 20 N (milestone 41: 0.016 / 0.011 / 0.011), over the 0.03 mm spec at 20 N. The post hoc (
posthoc.txt, after the run) checks the harness first (the G1 on its per-robot sides reproduces milestone 41 to 0.000 mm) and finds the rise at three of 15 poses, each with one leg joint (not a hold-set joint) flipping at the worst case: (0.6 m, −0.15 m) 0.006 → 0.037 mm at 20 N, (1.2 m, 0.15 m) 0.006 → 0.025 at 5 N and → 0.016 at 20 N, (1.2 m, 0) 0.001 → 0.004 at 20 N. One G1 pose improves (1.2 m, 0.15 m) at 0 N: 0.016 → 0.006. - What it means. A per-pose side is the right idea for the H1: it rescues two poses, lifts every margin and cuts its worst residual by two thirds at 20 N. But a margin measured at zero hose force is the wrong score to pick sides by: on the G1, sides that widen the zero-drag margin narrow some leg's margin under drag. The controller should score sides at the loads it will meet (the hose sweep), or keep the per-robot side where the gain is small.
- Not measured. Statics; mu 0.5 and the 'braked' waist; sides scored under hose drag or sensor error; the two rescued H1 poses in the sweep; the G1's over-limit wrist pitch at low poses.
- Next. Score hold sides by the worst-case margin over the hose and sensor sweep (or the worst residual), per pose, for both robots; rebuild the pose set with the two rescued H1 poses and sweep them.
2026-11-18 · Milestone 47 planned: per-pose hold sides at every pose of both robots
- Why. Milestone 46 found the per-robot hold sides (milestone 32's
SIGNS, one set per robot) were wrong at H1 (1.4 m, −0.15 m), and its section A1 found the torque-minimising side disagrees with them at both window ends at four other H1 poses. Its "Next" asked for per-pose hold sides at every pose of both robots, with the full stack and residual table. - Disclosed before registering (a diagnostic, zero hose force, exact sensor, both window ends). Section A1's rule (take the 'torque' model's side where it agrees at both window ends, else the per-robot side) is not safe as a controller: it raises the common leg margin at some poses (H1 (1.6 m, 0) 0.095 → 1.00 N m, (1.4 m, −0.15 m) 0.033 → 1.56) but cuts it at others (G1 (0.8 m, −0.15 m) 2.89 → 0.20, H1 (0.4 m, 0) 2.32 → 0.78). It also finds a 'hold' solution at the two H1 poses milestone 38 skipped for having none, (0.4 m, −0.15 m) and (1.6 m, −0.15 m) (margins 5.20 and 4.15). No stack, plan with sensor error or residual has been computed with any per-pose side.
- Model. Milestone 46's import (so M1 is the object milestones 41 to 46 built), with its single flip replaced by a per-pose side rule: at every pose of both robots in milestone 41's set, the hold sides S* are the combination (16 for the H1's four hold joints, 64 for the G1's six) with the largest common leg margin at zero hose force and an exact sensor, the smaller of the two window ends; ties keep the per-robot side. A pose whose S* margin reaches 0.1/β (β 0.9) plans with the spec's own 0.1 N m leg preload (milestone 45's tier dropped); otherwise it keeps milestone 46's sides and plan. Sections: A, per pose, the per-robot sides, S*, both margins and A1's sides (and S* at the two skipped H1 poses, a diagnostic; they are not added to the swept set here); B, the spec stack ('stock', 'braked') at every pose whose sides change, beside milestone 46's (milestone 43's, 45's or 46's figure for that pose); C, each robot's worst spec stack; D, each robot's worst residual along the wall ('braked', |e| 0.2 N, S 10 N, H 0 / 5 / 20 N, every hose direction and sensor axis), beside milestone 46's H1 and milestone 41's G1 figures.
- Hypotheses, registered before the run. - H187 (no pose needs a tier). With S*, every pose of both robots has a common margin of at least 0.1/0.9 N m. Refuted if any is smaller. - H188 (the skipped poses have a hold plan). Both H1 poses milestone 38 skipped have an S* common margin of at least 0.1/0.9 N m. Refuted if either does not. (Close to the diagnostic; registered as a check.) - H189 (the H1 keeps its spec). The H1's worst residual over all its poses is at most milestone 46's 0.010 / 0.019 / 0.032 mm plus 0.005 mm at 0 / 5 / 20 N drag. Refuted if any is larger. - H190 (the G1 keeps its spec). The G1's worst residual is at most milestone 41's 0.016 / 0.011 / 0.011 mm plus 0.005 mm. Refuted if any is larger. - H191 (sides cost no stack). 'braked': each robot's worst spec stack is at least milestone 46's (H1 0.9655, G1 0.9257) minus 0.1 points. Refuted if lower. - H192 (no more lost cases). The H1's swept cases with no plan at milestone 44's two poses are at most milestone 46's 23 of 204. Refuted if more.
2026-11-18 · Milestone 46 result: the right hip roll was held on the wrong side; held on the other side at (1.4 m, −0.15 m) it takes the spec's full preload and the pose leaves 0.005 mm, so the H1 meets the tool-point spec at all 19 poses with an FT 300; H181–H185 held, H186 refuted
experiments/m46_holdside/ (run, results, figure; imports milestone 45's run and changes one hold side at one pose), reproduce.py (m46 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card. No new library code, so no new tests.
- A correction to milestone 45. Its conclusion that (1.4 m, −0.15 m) "is decided by the wrist sensor's z accuracy, not by any preload the stance allows" holds only for milestone 32's per-robot hold sides. The stance did not cap the right hip roll; the hold constraint did: on the + side the hold QP pins it at exactly τ_f (+0.2000 N m at both window ends), and the torque-minimising controller wants it at −3.62 / −3.79 N m.
- The pose takes the full preload (H181 held). Held on the − side, the common leg margin the stance allows is 6.15 / 2.29 N m at 25 / 50 N (from 1.61 / 0.033), so the pose drops milestone 45's tier and plans with the spec's own 0.1 N m like the other 17 poses.
- The pose is rescued (H182, H183 held). 'braked', |e| 0.2 N, S 10 N: the worst residual along the wall at the pose is 0.005 mm at 0, 5 and 20 N drag (milestone 45: 0.172), with no joint flipping at any worst case. The H1's worst over all poses returns to milestone 41's 0.010 / 0.019 / 0.032 mm, so the tool-point spec holds at every H1 pose with a plan, at an FT 300-class 0.2 N error. The other poses reproduce milestone 41 to 0.000 mm.
- No cost (H184, H185 held). Stack 0.9716 ('stock' and 'braked'; milestone 45 0.9715), squeeze at its −10 N bound; the H1's worst spec stack stays 0.9655 at (1.6 m, +0.15 m). Cases at the two re-planned poses with no plan: 23 of 204 (milestone 45: 24).
- The per-robot side rule is the problem, not one joint (H186 refuted). Section A1 finds the torque-minimising side disagrees with the per-robot side at both window ends at four other H1 poses: both hip yaws at (0.4 m, 0) and (0.6 m, −0.15 m), both hip rolls at (1.4 m, 0) and (1.6 m, 0) (and at one end only at (1.0 m, −0.15 m), (1.2 m, −0.15 m) and (1.2 m, 0)). Those poses still meet the spec with the per-robot sides (their margins are large enough or milestone 44/45's tier covers them), so here it costs nothing measured, but a controller that picks hold sides per pose is the cleaner rule.
- What it means. The zero-press stance keeps a spec-grade plan at 5 of the H1's 6 re-leaned poses at |e| 0.2 N (the sixth, (1.6 m, −0.15 m), has no 'hold-min' solution, milestone 43), with no sensor better than an FT 300. Milestone 45's sensor bound (|e_z| < 0.19 N) is a property of the wrong side, not of the pose.
- Not measured. Statics; mu 0.5 and the 'braked' waist; per-pose sides at every pose (only this one changed); whether per-pose sides move the stack or residual elsewhere; the G1 (untouched).
- Next. Per-pose hold sides at every pose of both robots (section A1's rule as the controller), with the full stack and residual table; the tool-point spec as one table over both robots and both leans; the G1's over-limit wrist pitch at low poses.
2026-11-18 · Milestone 46 planned: hold the H1's right hip roll on the other side at (1.4 m, −0.15 m)
- Why. Milestone 45 found the stance caps the right hip roll's margin at 0.0339 N m at (1.4 m, −0.15 m), so a 0.2 N sensor z error flips it and leaves 0.172 mm along the wall. Its "Next" asked whether re-choosing the hold set frees that joint. Milestone 32 fixed each hold-set joint's side once per robot (
SIGNS, from the torque controller at the reference pose; milestone 41 re-derived them for the zero-press stance, still one set per robot). A side that is right at the reference pose need not be right at every pose. - Disclosed before registering (a diagnostic, zero hose force, exact sensor, 50 N end). With milestone 32's sides the hold QP pins the right hip roll at exactly +0.2000 N m (τ_f, the constraint binding), common leg margin 0.0333 N m. Dropping it from the hold set, the torque-minimising plan puts it at −3.79 N m (−3.62 at 25 N), held on the other side, and the common margin rises to 2.29 N m; holding it on the negative side gives the same plan. Releasing a second hold joint as well changes nothing useful. No stack, plan with sensor error, or residual has been computed.
- Model. Milestone 45's import (so M1 is the object milestones 41 to 45 built, tiers kept at (1.6 m, 0)). At (1.4 m, −0.15 m) only, the right hip roll's hold side is flipped to −1; its tier is dropped, so the pose plans with the spec's own 0.1 N m leg preload like the other 17 poses. Every other pose, joint and side is unchanged. Sections: A, at every H1 pose, the hold-set joints whose torque-minimising ('torque' model) side at either window end disagrees with the per-robot side, and at the pose the right hip roll torque and common margin under both sides; B, the spec stack at the pose ('stock', 'braked') beside milestone 45's; C, the H1's worst spec stack; D, milestone 45's section D unchanged ('braked', |e| 0.2 N, S 10 N, H 0 / 5 / 20 N, every hose direction and sensor axis).
- Hypotheses, registered before the run. - H181 (the pose takes the full preload). With the flipped side the pose has a spec plan at zero hose force and an exact sensor with the spec's 0.1 N m preload (common margin ≥ 0.1/0.9 at both ends). Refuted if not. (Close to the diagnostic; registered as a check.) - H182 (the pose is rescued). 'braked', |e| 0.2 N: the worst residual along the wall at (1.4 m, −0.15 m) is at most 0.03 mm at 0 and 5 N drag and 0.05 mm at 20 N. Refuted if any is larger. - H183 (the H1's tool-point spec holds). The H1's worst residual over all its poses is at most 0.03 mm at 5 N and 0.05 mm at 20 N (refuted in milestones 44 and 45 at 0.172). Refuted if either is larger. - H184 (no more lost cases). The swept cases at milestone 44's two poses with no plan are at most milestone 45's 24 of 204. Refuted if more. - H185 (the side costs no stack). 'braked': the pose's spec stack is at least milestone 45's 0.9715 minus 0.1 points. Refuted if lower. - H186 (one wrong side). Section A finds at most one other H1 pose where a hold-set joint's torque-minimising side disagrees with the per-robot side at both window ends. Refuted if more (then the per-robot side rule, not this one joint, is the problem).
2026-11-18 · Milestone 45 result: the hold set has no room; the stance caps the right hip roll at 0.0339 N m whatever the other legs give up, so (1.4 m, −0.15 m) stays at 0.172 mm with a 0.2 N sensor error and is a sensor-grade question, not a preload one; H176, H179 and H180 held, H175, H177 and H178 refuted
experiments/m45_tiered/ (run, results, post hoc, figure; imports milestone 44's run with its per-pose preload switched off), src/buildbot_sheathing/tiered.py (new: tier_lp, solve_tiered, row_mask), tests/test_tiered.py (4 tests), reproduce.py (m45 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card.
- A correction to milestone 44's post hoc. It said the joint that flips at (1.4 m, −0.15 m) is "a held joint the preload does not reach". It is
right_hip_roll, in milestone 32's hold set and one of the leg rows milestone 37's preload acts on: planned at 0.2300 N m (τ_f + 0.030) at both window ends, it settles to 0.1974 under a (0, 0, −0.2) N sensor error. The preload reached it; the margin was too small. - Check. The other 17 H1 poses reproduce milestone 41's worst residuals (0.010 / 0.019 / 0.032 mm) to 0.000 mm; the tiered poses are exactly milestone 44's two.
- No room on the hold set (H175 refuted). With every other leg row allowed down to its threshold, the largest common margin on the hold-set rows is T_H = 0.0339 N m at (1.4 m, −0.15 m) (binding: right hip roll, 50 N end), against the common 0.0333: the other legs were not what capped it. At (1.6 m, 0) T_H is 0.0960 against 0.0946 (left hip roll). The other leg rows, by contrast, had room: given the hold set at m_H, T_L is 0.190 and 0.656 N m, so they take the full 0.1 N m.
- The plan survives, the stack does not move (H176, H179, H180 held). Both poses plan; stacks 0.9715 and 0.9668 as milestone 44 (+0.00 points); 24 of 204 swept cases at the two poses have no plan, as in milestone 44.
- The residual does not come back (H177, H178 refuted). 'braked', |e| 0.2 N: (1.4 m, −0.15 m) leaves 0.172 mm at 0, 5 and 20 N drag (one hold-set flip at each worst case). The 20 N figure falls from milestone 44's 0.241 to 0.172 (the other leg rows now carry 0.1 N m), so the H1's worst over all poses is 0.172 mm at every drag. (1.6 m, 0) stays at 0.003–0.014 mm.
- What it needs instead (post hoc,
posthoc.txt). The right hip roll moves 0.163 N m per newton of sensor error along z (0.016 along x; a 1 N error along y leaves no plan, so that axis has no figure) at both window ends. So the pose holds the spec only while 0.163 |e_z| stays under the hold-set margin: with m_H at 0.9 T_H (0.0305) that is |e_z| < 0.19 N, and the post hoc finds 0.001–0.004 mm at |e| 0.05–0.15 N and 0.172 at 0.2 N; with m_H at 0.99 T_H (0.0336) |e| 0.2 N leaves 0.005 mm, a back-off of 1% that no real controller would trust. This pose is decided by the wrist sensor's z accuracy (0.2 N is two FT 300 standard deviations, milestone 36), not by any preload the stance allows: an ATI Mini45-class error (2 × 1/16 N = 0.125 N, milestone 36's sweep) passes, an FT 300's 0.2 N does not. - What it means. The zero-press stance keeps a spec-grade plan at 4 of the H1's 6 re-leaned poses at |e| 0.2 N, and at 5 of 6 with a sensor whose z error stays under about 0.19 N; a designer who cannot promise that should keep milestone 32's lean at (1.4 m, −0.15 m). A hold-first preload is still worth having elsewhere: it costs no stack and gave the other leg rows three to six times more margin than the common rule.
- Not measured. Statics; mu 0.5 and the 'braked' waist; a different hold set (choosing which roll joint to hold at this pose); the G1 (untouched, its smallest common margin 0.141 N m).
- Next. The tool-point spec as one table over both robots and both leans, now with the sensor bound per pose; whether re-choosing the hold set at (1.4 m, −0.15 m) frees the right hip roll; the G1's over-limit wrist pitch at low poses.
2026-11-18 · Milestone 45 planned: a hold-first (two-tier) preload for H1 (1.4 m, −0.15 m)
- Why. Milestone 44's per-pose leg preload left a 0.172 mm residual at H1 (1.4 m, −0.15 m), and its post hoc said the joint that flips there is "a held joint the preload does not reach". A diagnostic before this plan (worst case of milestone 44's no-drag sweep, β 0.9, sensor error (0, 0, −0.2) N) shows that is wrong: the joint is
right_hip_roll, one of milestone 32's hold set, which is one of the leg rows milestone 37's preload acts on (the hold set is a subset of the legs). Its planned torque is 0.2300 N m at both window ends, exactly τ_f + m_k (0.2 + 0.030), and under the true load it settles to 0.1974 N m, 0.0026 below τ_f: a 0.2 N z error moves it 0.033 N m, more than the 0.030 margin. So the preload reached it; the common margin the stance allows every leg row (0.033 N m) is too small. The open question is whether the stance can give the hold-set rows a larger margin if the other leg rows take less. - Model. Milestone 44's import (so M1 is the object milestones 41 to 44 built; its per-pose preload is switched off, so every pose plans with 0.1 N m unless listed here). New library code
src/buildbot_sheathing/tiered.py (tier_lp: the largest common margin on a chosen subset of rows with every other row at or above its own floor; solve_tiered: the controller QP with a per-row margin target), with tests. Per pose whose common margin t_min is below 0.1/β (milestone 44: exactly the two poses), at zero hose force and an exact sensor, at both window ends: T_H = the largest common margin on the hold-set rows with the other leg rows at margin ≥ 0; m_H = min(0.1, β T_H); then T_L = the largest common margin on the other leg rows with the hold-set rows at ≥ m_H; m_L = min(0.1, β T_L) (no plan if T_L < 0). β = 0.9 as milestone 44 registered. Milestone 38's plan is called unchanged except that the preload step solves with target m_H on hold-set rows and m_L on the rest. Sections: A, T_H, m_H, T_L, m_L per pose; B, the spec stack at the two poses ('stock', 'braked') beside milestone 44's; C, the H1's worst spec stack; D, the worst residual along the wall ('braked', |e| 0.2 N, S 10 N, H 0 / 5 / 20 N, every hose direction and sensor axis), all poses, the other poses (check: equal to milestone 41's) and each of the two poses, with cases with no plan and the flipped moving joints at each pose's worst case. - Disclosed before registering. The diagnostic above (one case, β 0.9). No tiered margin, plan or residual has been computed.
- Hypotheses, registered before the run. - H175 (the stance has room on the hold set). At (1.4 m, −0.15 m), T_H ≥ 0.05 N m. Refuted if smaller. - H176 (the plan survives the trade). Both poses have a tiered spec plan at zero hose force and an exact sensor (T_L ≥ 0). Refuted if either has none. - H177 (the pose is rescued). 'braked', |e| 0.2 N: the worst residual along the wall at (1.4 m, −0.15 m) is at most 0.03 mm at 0 and 5 N drag and 0.05 mm at 20 N. Refuted if any is larger. - H178 (the H1's tool-point spec holds). The H1's worst residual over all its poses is at most 0.03 mm at 5 N and 0.05 mm at 20 N (milestone 44's H174, refuted at 0.172 and 0.241). Refuted if either is larger. - H179 (no more lost cases). The swept cases at the two poses with no plan are at most milestone 44's 24 of 204. Refuted if more. - H180 (the stack does not move). 'braked': each pose's spec stack is within 0.1 points of milestone 44's (0.9715 and 0.9668). Refuted if further.
2026-11-17 · Milestone 44 result: a per-pose leg preload gives both dropped H1 poses a plan at no clean-sheet cost, but only (1.6 m, 0) meets the tool-point spec; at (1.4 m, −0.15 m) a hold-set joint flips and the residual is 0.17 mm; H170, H172 and H173 held, H171 and H174 refuted
experiments/m44_legpreload/ (run, results, post hoc, figure; imports milestone 43's run, so M1 is the object milestones 41 to 43 built, and calls milestone 38's plan unchanged with a per-pose preload), reproduce.py (m44 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card. No new library code, so no new tests.
- Check. The other 17 H1 poses keep 0.1 N m and reproduce milestone 41's H1 worst residuals (0.010 / 0.019 / 0.032 mm at 0 / 5 / 20 N) exactly; the rule changes exactly milestone 43's two dropped poses (t_min 0.0333 and 0.0946 N m, so m_k 0.0300 and 0.0851).
- The plan returns, and the stack does not care how big the preload is (H170, H172, H173 held). With β = 0.9 both poses have a spec plan, squeeze at its −10 N bound. Their stacks are 0.9715 and 0.9668 at every β in {0, 0.25, 0.5, 0.9}, 0.08 and 0.27 points under milestone 40's old-lean spec stacks; the H1's worst over its 19 poses stays 0.9655 at (1.6 m, +0.15 m). The stack counts loose joints, and a leg row met at margin 0 holds the joint as surely as one met at 0.1 N m.
- But the sweep breaks it (H171, H174 refuted). At |e| 0.2 N, 24 of the 204 swept (H, φ, e) cases at the two poses have no plan (0 of 12 at no drag, 5 of 96 at 5 N, 19 of 96 at 20 N). The H1's worst residual along the wall rises from 0.019 to 0.172 mm at 5 N drag and from 0.032 to 0.241 mm at 20 N, all of it at (1.4 m, −0.15 m); at no drag it is 0.172 mm already.
- Why (post hoc,
posthoc.txt). Per pose and β, at the worst case: at (1.6 m, 0) β ≥ 0.5 leaves 0.003 mm at no drag and 0.014 mm at 5 N, inside the spec, with no hold-set joint flipping (0.46–0.71 mm at β = 0, where two or three do). At (1.4 m, −0.15 m) one joint of milestone 32's hold set still flips at β = 0.5 and 0.9 and the residual stays 0.172 mm (0.35 mm at β ≤ 0.25, two flips). The leg preload acts on the leg rows; the joint that crosses its friction band under a 0.2 N sensor error there is a held joint the preload does not reach, so no preload the stance allows fixes it. A larger preload also costs cases: at 5 N drag (1.6 m, 0) keeps a plan in 46 of 48 cases at β 0.5 and 43 of 48 at β 0.9. - What it means. The per-pose preload rescues (1.6 m, 0), so the zero-press stance keeps a spec-grade plan at 4 of the H1's 6 re-leaned poses instead of 3. At (1.4 m, −0.15 m) it gives a plan but a residual nine times the spec's, still 18 times inside the ±3.14 mm joint-nail window but not the 0.01–0.02 mm figure the paper quotes. There the designer should keep milestone 32's lean and catch the lift-off another way, or give up the pose.
- Not measured. Statics; mu 0.5 and the 'braked' waist as before; a preload on the hold-set joints (the QP's hold signs) rather than the legs; the G1 is untouched (its smallest margin, 0.141 N m, carries 0.1).
- Next. The tool-point spec as one table over both robots and both leans, with the pose each figure comes from; the G1's over-limit wrist pitch at low poses; whether a margin row on the hold-set joint rescues (1.4 m, −0.15 m).
2026-11-17 · Milestone 44 planned: a per-pose leg preload for the H1's two re-leaned poses that lost the spec plan
- Why. Milestone 43 found that two re-leaned H1 poses, (1.4 m, −0.15 m) and (1.6 m, 0), keep a 'hold-min' solution but have no spec plan, because the stance allows a leg preload margin of only 0.033 and 0.095 N m at the 50 N window end against the spec's fixed 0.1 N m. Both margins are positive, so the spec can carry a smaller preload there. A designer adopting the zero-press stance needs to know what that costs: in clean sheets, in the tool-point residual along the wall, and whether the plan survives hose drag and sensor bias.
- Model. Milestone 43's import (so M0 and M1 are the objects milestones 41 to 43 built). For every pose of the zero-press set, t_min is the smaller of the two largest leg margins (25 and 50 N, zero hose force, exact sensor; milestone 43's post hoc). The per-pose preload is m_k = min(0.1, β t_min) with β = 0.9 registered (a 10% back-off for hose and sensor error); it changes only poses with t_min < 0.111 N m. Milestone 38's
plan is called unchanged with m_k in place of 0.1. Sections: A, m_k per pose; B, the spec stack at the two poses ('stock' and 'braked') at β ∈ {0, 0.25, 0.5, 0.9}, beside milestone 40's spec stack on the old lean and the 'hold-min' stack; C, the H1's worst spec stack over the set; D, the worst residual along the wall, 'braked', |e| 0.2 N, S 10 N, H 0 / 5 / 20 N over every hose direction and sensor axis (milestone 41's section B), all poses and the two poses alone, with the number of (H, φ, e) cases at the two poses that have no plan. Check: the worst residual over the other H1 poses must equal milestone 41's (they keep m = 0.1). - Disclosed before registering. Milestone 43's margins and stacks are known; no plan with a smaller preload has been computed. That a plan exists at zero hose force for β < 1 is close to construction (the margin LP is feasible at t_min); the squeeze step after it is not.
- Hypotheses, registered before the run. - H170 (the plan returns). With β = 0.9 both poses have a spec plan at zero hose force and an exact sensor. Refuted if either has none. - H171 (and survives the sweep). 'braked', |e| 0.2 N, H ∈ {0, 5, 20} N, every hose direction and sensor axis: both poses have a plan in every case. Refuted if any case drops. - H172 (a small preload costs little). 'braked', β = 0.9: each pose's spec stack is within 0.5 points of its milestone 40 spec stack on the old lean (0.9723 and 0.9695). Refuted if either is further below. - H173 (the H1's worst holds). 'braked', β = 0.9: the H1's worst spec stack over all its zero-press poses with a plan is at least 0.9600. Refuted if lower. - H174 (the tool-point spec holds). 'braked', |e| 0.2 N, β = 0.9: the H1's worst residual along the wall over all poses is at most 0.03 mm at 5 N drag and 0.05 mm at 20 N (milestone 41: 0.019 and 0.032). Refuted if either is larger.
2026-11-17 · Milestone 43 result: in the spec's own convention the zero-press stance gains the G1 0.97 points at its worst pose instead of losing 1.3; the cost moves to the H1, which keeps a spec plan at only 3 of its 6 re-leaned poses; H166, H167 and H168 held, H165 and H169 refuted
experiments/m43_specstack/ (run, results, post hoc, figure; imports milestone 42's run, so both setups are the objects milestones 41 and 42 built), reproduce.py (m43 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card. No new library code, so no new tests.
- Checks. Milestone 32's leans reproduce milestone 40's section D stacks to 0.0e+00, and the 'hold-min' torques on the re-leans reproduce milestone 42's 'braked' cells to 0.0e+00.
- The conventions part at the re-leaned worst pose (H165 refuted). At the G1's (1.2 m, 0), 'braked', the spec plan (leg preload 0.1 N m, squeeze −4.00 N) gives 0.9300; 'hold-min' gives 0.9030. They agreed at 0.9160 on milestone 32's lean, so milestone 42's post hoc finding that the conventions agree at each worst pose does not carry over to the new poses: they differ by 2.70 points there.
- The G1 gains (H166 and H168 held, and more). Over the whole set the G1's worst spec stack rises from 0.9160 to 0.9257 ('braked', +0.97 points; now at (1.2 m, −0.15 m)) and from 0.9081 to 0.9257 ('stock', +1.76). The comparison is like for like: all 14 poses milestone 40 could plan are planned again, plus (1.0 m, +0.15 m), which had no spec plan on its old lean and now has one (0.9547). Per re-leaned pose ('braked'): −0.31, +0.79, new, −1.01, +1.40, −1.17 points.
- The H1 pays in poses, not points (H167 held, H169 refuted). Of the six re-leaned H1 poses, (1.6 m, −0.15 m) lost its controller at milestone 41, and two more, (1.4 m, −0.15 m) and (1.6 m, 0), keep a 'hold-min' solution but have no spec plan. The three that keep one lose 0.09–0.22 points; like for like over the 17 poses both leans can plan, the H1's worst falls from 0.9677 to 0.9655. (Its worst over the set "rises" from 0.9564 to 0.9655 only because 0.9564's pose dropped out: not a gain.)
- Why the two H1 poses drop (post hoc,
posthoc.txt). The hold QP solves at both window ends; what fails is the leg preload at 50 N: the largest leg margin the stance allows is 0.033 N m at (1.4 m, −0.15 m) and 0.095 N m at (1.6 m, 0), against milestone 37's 0.1 N m. The same check explains the G1's new pose: at (1.0 m, +0.15 m) the old lean allowed 0.088 N m at 50 N, the re-lean 0.707. - What it means. Milestones 41 and 42's "the zero-press stance costs the G1 1.3 points" is true of the 'hold-min' torques but not of the spec, whose preload and squeeze are re-planned for the new stance and come out ahead on the G1. The real price of the zero-press stance is on the H1 near the top of its reach: half its re-leaned poses cannot carry the 0.1 N m leg preload at a 50 N press. A designer can lower the preload there (both margins are positive), cap the press, or keep the old lean and catch the lift-off another way.
- Not measured. Statics; zero hose force and an exact sensor (the residuals along the wall were milestone 41's); the 'set' column of milestone 40 (the re-press) is not recomputed; mu 0.5 and the 'braked' waist as before.
- Next. A per-pose leg preload (the largest the stance allows up to 0.1 N m) on the H1's two dropped poses and what it costs the residual; the tool-point spec as one table; the G1's over-limit wrist pitch at low poses.
2026-11-17 · Milestone 43 planned: the spec's own stack on the zero-press poses
- Why. Milestones 41 and 42 costed the zero-press stance with milestone 38's 'hold-min' torques (no leg preload, no squeeze), while milestone 40's per-pose stack, the one the spec quotes, takes each pose's torques from the spec's plan (leg preload 0.1 N m and the 10 N squeeze). Milestone 42's post hoc found the two conventions agree at each robot's worst pose but differ at 9 of 34 poses by up to 0.97 points. So the spec's stack on the re-leaned poses, the number a designer adopting the zero-press stance would read, has not been computed.
- Model. Milestone 42's import (milestone 41's run, so M0 on milestone 32's leans and M1 on the re-leans are the objects it built). For each setup, each pose with a 'hold-min' solution and each waist case ('stock', 'braked'): the spec plan at zero hose force and an exact sensor (milestone 38's
plan, S* 10 N, m 0.1 N m), and the no-hose clean-sheet stack with loose joints uniform, exactly milestone 40's section D 'uniform' column. M0 must reproduce milestone 40's column to 0 (check). Per re-leaned pose: the spec-convention change, set beside milestone 42's 'hold-min' change. - Disclosed before registering. Milestones 40–42's published numbers are known (G1 worst 0.9160 braked, 0.9081 stock, both at (1.2 m, 0); 'hold-min' on the re-leans 0.9030 there); no spec plan on a re-leaned pose has been computed. A re-leaned pose with a 'hold-min' solution but no spec plan drops out and is printed.
- Hypotheses, registered before the run. - H165 (the conventions agree at the worst pose). 'braked': the G1's re-leaned spec stack at (1.2 m, 0) is within 0.10 points of milestone 42's 'hold-min' 0.9030. Refuted if further. - H166 (the G1 loses at most 1.5 points). 'braked': the G1's worst spec stack over the re-leaned set is at least 0.9010 (milestone 40: 0.9160). Refuted if lower. - H167 (the H1 pays little). 'braked': every re-leaned H1 pose with a plan loses at most 0.5 points against its own milestone 40 spec stack. Refuted if any loses more. - H168 (the stock waist pays the same). 'stock': the G1's worst spec stack over the re-leaned set loses at most 1.5 points against milestone 40's 0.9081. Refuted if more. - H169 (the plan survives). Every re-leaned pose with a 'hold-min' solution also has a spec plan (preload and squeeze) at zero hose force. Refuted if any does not.
2026-11-17 · Milestone 42 result: the G1's 1.3-point loss is all lean; the controller's four flipped sides move no stack by a digit but are needed to keep one re-leaned pose; and milestone 41's per-pose losses mixed two torque conventions; H161, H162 and H163 held, H164 refuted
experiments/m42_sides/ (run, results, post hoc, figure; imports milestone 41's run, so both setups are the objects it built), src/buildbot_sheathing/factorial.py (new: the 2 × 2 split), tests/test_factorial.py (4 tests), reproduce.py (m42 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card.
- Which sides flipped. Milestone 41's re-leaned pose set flipped four of the G1's six 'hold-min' sides: both hip rolls and both ankle rolls (the hip yaws kept theirs). The H1's four did not change.
- The lean is all of it (H161, H162, H163 held). At the G1's worst pose (1.2 m, 0) the 2 × 2 is 0.9160 / 0.9160 (milestone 32's lean, either sides) and 0.9030 / 0.9030 (zero-press lean, either sides): lean −1.30 points, sides 0.00, interaction 0.00. At every pose where both sides have a solution, on either lean and on both robots, the two sides give the same stack to the printed digit. The sides change which way a held roll joint is held, not whether it is held, and the stack counts only loose joints (post hoc: the held-joint count differs at one pose, G1 (1.2 m, +0.15 m), 21 vs 20, with the same stack).
- But the flip is not free to undo (H164 refuted). With milestone 38's sides, the re-leaned G1 at (1.2 m, +0.15 m) has no 'hold-min' solution; milestone 41's sides keep it (0.9363). So the side choice decides which poses the controller can hold, not how clean the sheets are where it can.
- A correction to milestone 41 (post hoc,
posthoc.txt). Milestone 40's per-pose stack took the torques from the spec's plan (leg preload 0.1 N m and the squeeze); milestone 41's section C took milestone 38's 'hold-min' torques, and printed one against the other. The two conventions agree at each robot's worst pose (G1 0.9160, H1 0.9564), which is why milestone 41's headline stands, but differ at 9 of 34 poses by up to 0.97 points. In one convention the re-lean costs the G1 0.31–1.30 points at four poses and gains 0.60–0.79 at two, and costs the H1 0.09–0.29 at five (milestone 41 printed 0.1–0.9 for the H1, and a 0.2-point gain at G1 (1.0 m, 0) that is a 0.32-point loss). - What it means. The zero-press stance's cost in clean sheets is the stance itself; a designer cannot buy it back by keeping the controller's old sides, and should let the sides be re-chosen over the new pose set because that keeps one more pose holdable. The cost is up to 1.3 points on the G1 near the top of its reach and under 0.3 on the H1.
- Not measured. Statics; the stacks here use the 'hold-min' torques without milestone 37's leg preload and milestone 38's squeeze (milestone 41's convention), so the spec's own stack on the re-leaned poses is not yet computed; the 'braked' waist and mu 0.5 are as before.
- Next. The spec's stack (preload and squeeze plan) on the re-leaned poses, so the zero-press stance is costed in the convention the spec uses; the tool-point spec as one table; the G1's over-limit wrist pitch at low poses.
2026-11-17 · Milestone 42 planned: is the G1's 1.3-point loss the lean or the controller's new sides?
- Why. Milestone 41 re-leaned 12 poses to stand at zero press and rebuilt milestone 38's setup on the new pose set. The G1's worst-pose no-hose stack ('braked') fell from 0.9160 to 0.9030, but two things changed at once: the six re-leaned G1 poses, and the 'hold-min' controller's chosen sides (
choose_signs picks them by the worst LP margin over the whole pose set, so a new pose set can flip them). The H1's sides did not change and its re-leaned poses lost 0.1–0.9 points. A designer adopting the zero-press stance needs to know whether the cost is the stance (unavoidable) or the side choice (a controller setting that can be kept). - Model. A 2 × 2 on the G1, no hose, 'braked' waist, everything else as milestone 41: leans {milestone 32's, re-leaned} × sides {milestone 38's, milestone 41's}. Two cells exist (0.9160 and 0.9030 at the worst pose); the two crossed cells are computed by re-running milestone 38's
torques with the other set's sides (milestone 41's run is imported, so the poses, plans and stack code are the same objects). Per pose and at the worst pose: the lean's effect at fixed sides, the sides' effect at fixed lean, and the interaction. Which leg joints flipped is printed. The H1 is run the same way as a check (its sides did not change, so its crossed cells must equal its own). - Disclosed before registering. Only milestone 41's published numbers are known; no crossed cell has been computed. If the old sides have no 'hold-min' solution at a re-leaned pose (or the new sides at an old pose), that pose drops from the cell and is printed.
- Hypotheses, registered before the run. - H161 (the lean is most of it). With milestone 38's sides kept, the re-lean alone costs the G1's worst pose at least 0.65 points (half of the 1.30). Refuted if less. - H162 (the sides alone are cheap). On milestone 32's leans, swapping to milestone 41's sides moves the G1's worst-pose stack by at most 0.5 points. Refuted if more. - H163 (no large interaction). At the G1's worst pose, |both − lean only − sides only| is at most 0.3 points. Refuted if more. - H164 (the old sides still hold). Milestone 38's G1 sides keep a 'hold-min' solution at all six re-leaned G1 poses. Refuted if any loses it.
2026-11-17 · Milestone 41 result: every leaned stance can be re-leaned to stand at zero press and still reach 50 N, and the 0.01–0.02 mm spec holds; but the G1 at 1.2 m keeps only 1.6 N above the window, one H1 pose loses its controller, and the G1's worst stack drops 1.3 points; H156 and H159 held, H157, H158 and H160 refuted
src/buildbot_sheathing/stance0.py (new: the bisection on the lean, the window test, the rigid-body window width), experiments/m41_stance0/ (run, results, figure, post hoc ceiling; runs milestone 38's import twice, with milestone 32's leans and with the re-leans, so every plan, side choice and calibration is recomputed), tests/test_stance0.py (4 tests), reproduce.py (m41 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card.
- A zero-press lean exists everywhere (H156 held). All 12 poses that needed a press (milestone 40) have a lean that balances over the whole of 0–50 N: the G1 at z 1.0 m tops out at 58.6–62.2 N, at 1.2 m at 51.6–51.7 N; the H1 at 1.4 m at 67.2 N, at 1.6 m at 58.7 N.
- The re-lean is not small (H157 refuted). The balanced centre of mass moves back 9.9–39.1 mm: 18 mm (G1, 1.0 m) and 39 mm (G1, 1.2 m), 10 and 24 mm (H1, 1.4 and 1.6 m).
- Check, and why the window cannot be wider. At every re-leaned pose the simulated window's width equals (m + m_tool) g × sole length / grip height to the printed digit (G1 33.34 kg, H1 51.44 kg, sole 0.170 and 0.175 m, tool 3.4 kg): the centre of pressure crosses the whole sole between 0 N and the top. So no lean gives more. Post hoc, from that formula: a stance that stands at zero press and still balances at 50 N exists only below a grip height of 1.225 m (G1) and 1.883 m (H1); the G1's top poses sit 4 cm under its ceiling, which is why they keep only 1.6 N of margin above a 50 N press. At lower poses friction or the joint geometry binds first (the analytic width overstates the simulated one).
- One pose loses its controller (H158 refuted). 11 of 12 re-leaned poses keep a 'hold-min' solution at both window ends; the H1 at (1.6 m, −0.15 m) does not and drops out, as the H1 at (0.4 m, −0.15 m) already had. The G1 controller's chosen sides changed with the new pose set (they are chosen over the set); the H1's did not.
- The tool-point spec holds (H159 held). Braked waist, |e| 0.2 N, S 10 N, m 0.1 N m: the G1's worst residual along the wall is 0.016 / 0.011 / 0.011 mm at 0 / 5 / 20 N of drag (milestone 40: 0.021 / 0.012 / 0.012), the H1's 0.010 / 0.019 / 0.032 (unchanged). Over the re-leaned poses alone the worst is 0.016 mm (G1) and 0.014 mm (H1).
- The stack pays (H160 refuted). No hose, braked: the G1's worst pose, (1.2 m, 0), falls from 0.9160 to 0.9030 (1.30 points); among the re-leaned poses three G1 poses lose 0.3–1.3 points and two gain 0.2–0.6, and every H1 pose loses 0.1–0.9. The H1's worst pose over the set rises 0.46 points only because its old worst pose, (1.6 m, −0.15 m), lost its controller and left the set: that is not a gain. Which part of the G1's loss is the lean and which the changed sides is not separated here.
- What it means. "A stance that balances at zero press" can join the spec at every pose tried, at a cost of up to 1.3 points of clean sheets and, for the G1 near the top of a wall, almost no margin above the press (1.6 N); the ceiling is a formula a designer can check before buying: grip height < (m + m_tool) g × sole length / F_max.
- Not measured. Statics only (the lift-off and the lean change between nails are not simulated); the 'hold-min' controller and mu 0.5 are milestone 32's; the stack's split between lean and side choice is not separated; the sole length is the Menagerie collision geometry, not a measured foot.
- Next. The paper's tool-point spec as one table (sensor, leg preload, squeeze, waist brake, zero-press stance and its height ceiling); separate the lean from the side choice in the G1's 1.3 points; the G1's over-limit wrist pitch at low poses.
2026-11-17 · Milestone 41 planned: a stance that stands at zero press, and what it costs in the window
- Why. Milestone 40 found that 12 of milestone 32's 35 leaned poses balance only above a press of 3.7–11.8 N (the G1's six at z 1.0 and 1.2 m, milestone 30's calibration pose among them; the H1's six at 1.4 and 1.6 m). Milestone 32 chose each lean so the centre of pressure sits at the middle of the sole under the mid-window press (37.5 N); a nailer whose recoil lifts the nose then leaves those robots leaning on nothing. The fix an engineer would try first is to lean less. Both edges of a stance's balancing window rise with its lean, so the lean that stands at zero press and keeps the most of the window above it is the furthest-forward one that still stands at 0 N.
- Model. For each pose whose window starts above 0 N, the balanced centre of mass ahead of the feet centre (
com_dx) is bisected between 0 and its milestone 32 lean for the largest value at which the stance balances at zero press (same IK, same contact cones, mu 0.5; tolerance 0.5 mm); every other pose keeps its lean. A re-leaned pose whose window covers 0–50 N replaces the original, and milestone 38's whole setup is rebuilt on the new pose set (its import runs again with the lean swapped, so the controller's chosen sides, plans, squeeze, calibration at milestone 30's pose and torques are recomputed, not patched). Then, with milestone 39's 'braked' waist (milestone 40's spec), S* 10 N, m 0.1 N m, rubber hose, k_t 1e4: (B) the worst residual along the wall over all poses and over the re-leaned ones, |e| 0.2 N, H 0, 5 and 20 N, against milestone 40's 'uniform' column; (C) the no-hose clean-sheet stack per pose against milestone 40's. An analytic check: a rigid stance's window is (m + m_tool) g × sole length / grip height wide. - Disclosed before registering. Only milestones 32–40's published numbers and code have been read; no pose has been re-leaned and no window, residual or stack at a new lean computed. If milestone 30's calibration pose has no zero-press lean it keeps its lean, for calibration only, and that is printed. The re-lean is quasi-static like everything since milestone 30; nothing here simulates the lift-off itself.
- Hypotheses, registered before the run. - H156 (a zero-press lean exists). Every one of the 12 poses has a lean that balances over the whole of 0–50 N. Refuted if any does not. - H157 (the re-lean is small). At every re-leaned pose the balanced centre of mass moves back by at most 20 mm. Refuted if more anywhere. - H158 (the controller still holds). Every re-leaned pose that covers 0–50 N keeps a 'hold-min' solution at both window ends. Refuted if any loses it. - H159 (the spec survives). 'braked', |e| 0.2 N, a 5 N drag: the worst residual along the wall over all poses is at most 0.1 mm on both robots (milestone 40: G1 0.012, H1 0.019). Refuted if either exceeds it. - H160 (the stack survives). 'braked', no hose: each robot's worst-pose stack loses at most 0.5 points against milestone 40's (G1 0.9160, H1 0.9564). Refuted if either loses more.
2026-11-17 · Milestone 40 result: the nailer's recoil (0.09–8.7 N s, from published measurements) lifts the nose, and the re-press from zero sets none of the loose joints, so milestone 39's spec stands; but a third of the leaned stances cannot stand without the press; H153 held, H151, H152, H154 and H155 refuted
src/buildbot_sheathing/liftoff.py (new: the recoil bracket, the set test on a rising press, the post hoc entry test), experiments/m40_liftoff/ (run, results, post hoc census, figure; imports milestone 39's run, so every pose, plan and residual is the same code), tests/test_liftoff.py (5 tests), reproduce.py (m40 listed), paper/draft.md (section 5.11 paragraph, two references), paper/check_numbers.py, README, PROPOSAL and project card.
- The recoil, bracketed. An 8d common shank is 4.33 g; at the chronograph's 29.8–36.9 m/s (Buchalter et al. 2002) it carries 0.129–0.160 N s, at Peterson and Dixon's 20–30 m/s 0.087–0.130 N s. That is a floor (the piston and driver add to it). Lowe et al. (2014)'s force-plate unloading, taken at face value, is 8.7 N s per transfer phase, an upper figure for the user and tool together. The bracket 0.087–8.7 N s straddles the 0.25–1 N s band where
buildbot-recoil's nose leaves the wall. - Check. With the 'uniform' compensator every residual equals milestone 39's to 0.0 mm (G1 braked 5 N 0.012, stock 0.485; H1 braked 0.019).
- One post hoc change, disclosed. The smoke run found that a leaned pose balances only above a press: the G1's six poses at z 1.0 and 1.2 m (milestone 30's calibration pose among them) need 6.4 and 11.6–11.8 N, the H1's six at 1.4 and 1.6 m need 3.7 and 8.0 N. Below that the stance tips towards the wall. So the rise starts where the stance first balances under the load at hand (samples that balance under no controller were skipped: 15,024 of 97,762 in section B), not at F_r. That is itself the finding with consequences: at 12 of 35 poses, a recoil that lifts the nose leaves the robot leaning on nothing, so the stance must be planned to balance at zero press (or the controller must catch the lean) before a sequential-trip shot.
- The re-press sets nothing at the census (H151 refuted, H155 refuted). At H 0 with an exact sensor, 56 nose-moving joints are loose in the window (G1 41, H1 15). Only 11 (all G1: 10 arm, 1 waist) sit below tau_f over the whole window, and none of them reaches tau_f on the rise: the largest |tau| any of them reaches is 0.159 N m (F_r 0) against tau_f 0.2. F_r 10 N sets none either (0 vs 0, so 'fewer' is refuted).
- Why the uniform model survives (post hoc census,
posthoc.txt). Of the 56, 36 change sign inside the 25–50 N window and 7 rise through tau_f inside it: where they sit depends on the force at which the tool fires, not on how the nose got there. Only 2 enter the window above tau_f with one sign, which a rising press would set; this test was not registered. - Milestone 39's compensator keeps the spec (H152 refuted, H153 held). Braked waist, |e| 0.2 N, 5 N drag: the G1's worst residual is 0.012 mm with lift-off and either compensator. In one registered cell it does matter: at a 20 N drag, at pose (0.6 m, −0.15 m) with the hose pulling along −y, the rise sets the G1's right wrist yaw (−0.18 to +0.08 N m in the window, −0.44 to −0.23 on the rise), and the unaware compensator is left 0.203 mm where the aware one keeps 0.012 (stock and braked alike).
- The stack does not move (H154 refuted). With no joint set at H 0, the no-hose stack at the G1's worst pose is 0.9160 either way (braked; 0.9081 stock), the H1's 0.9564.
- What it means. The recoil does lift the nose, but the tool-point spec of milestones 36–39 (current F/T sensor, leg preload, 10 N squeeze, braked waist) does not depend on how the nose returns: the joints it leaves loose are loose because the firing force moves them, and only a heavy drag lets the rise set one (a compensator that plans the rise removes that 0.2 mm). What the recoil does change is the stance: a leaned stance that needs the press to stand must not be used with a tool that lifts off.
- Not measured. The rise is quasi-static (no recoil transient is simulated, and a joint the transient sets and the rise does not reach stays loose); the recoil enters only as whether the nose lifts; tau_f 0.2 N m is milestone 37's unsourced figure (a tau_f below 0.16 N m would let the rise set some joints); balance below the window is not modelled beyond saying it fails.
- Next. The paper's headline for a tool-point spec as one table (sensor, leg preload, squeeze, waist brake, a stance that balances at zero press); a stance chosen to balance over 0–50 N and what it costs in the window; the G1's over-limit wrist pitch at low poses.
2026-11-17 · Milestone 40 planned: the recoil's lift-off and the re-press from zero
- Why. Every milestone since 30 scores a joint whose torque stays below the holding threshold tau_f over the 25–50 N press window as loose: anywhere in its gap, uniformly, at every nail. That assumes nothing about how the nose got onto the wall. But the nose does not stay there between nails. The recoil lifts it:
buildbot-recoil found 0.5 N s loses the wall in most arm loops and 1 N s in all of them, and its recoil was swept because no figure was found. A sequential-trip tool must also release its contact arm before the next shot, and the robot moves 6 or 12 in to the next nail. So each nail starts from a lifted nose and a press that rises from (near) zero. A loose joint whose torque passes tau_f somewhere on that rise is left on that side when the window is reached: a repeatable bias, not a uniform scatter. Milestone 39's compensator does not model it. - Sources for the recoil, read before registering. (1) Nail velocity from pneumatic nailers, ballistic chronograph: 97.7–121.1 ft/s, mean 105.5 (29.8–36.9 m/s; Buchalter et al. 2002, Ear Nose Throat J 81:779–783, as quoted in NIOSH's 2013 Science Blog post on nail-gun speed); Peterson and Dixon (1975) 20–30 m/s at 2 m (same post). An 8d common shank (0.131 in by 2.5 in, steel 7850 kg/m³) is 4.33 g, so the nail alone carries 0.087–0.160 N s: a floor, since the piston and driver's momentum also goes into the tool. (2) Lowe, Albers and Hudock (2014, Int J Ind Ergon 44(5):715–722, doi:10.1016/j.ergon.2014.06.007, Appendix B): force-plate ground reaction under two users nailing flat, cumulative load 32.2 N against the tool-and-hose weight's 42.2 N over the 0.87 s transfer phase, read by the authors as recoil unloading 24% of the weight. Taken at face value that is (42.2 − 32.2) × 0.87 = 8.7 N s per nail, an upper figure for the whole user-tool system (the paper's own wording of the baseline is inconsistent, and the user's arm motion is in it). The bracket 0.09–8.7 N s straddles the 0.25–1 N s band where
buildbot-recoil's nose leaves the wall, so lift-off cannot be ruled out, and the release and the move between nails lift it anyway. - Model. Milestone 39's spec (S* = 10 N squeeze, m = 0.1 N m leg preload, settle model, rubber hose, k_t 1e4, one calibration at milestone 30's pose in the same hose state and sensor error), cases 'stock' and 'braked' (waist). New: the press is quasi-static from a release force F_r up to the window, sampled at F_r, F_r + 5, ..., 20 N with the same controller (hold QP and leg preload at each F, falling back to the hold QP alone where the preload has no solution, counted) and the window's squeeze. A joint below tau_f at both window ends is set if its torque reaches tau_f at some sampled F on the rise: it then sits at the side of the last such sample, a bias of −sign(tau) gap / 2. Otherwise it stays loose. The robot's set sides come from its settled torques, the compensator's from its plan. (A) Census at H 0 with an exact sensor: nose-moving joints loose in the window, and how many the re-press sets, by kind, for F_r 0 and 10 N. (B) Worst residual along the wall over poses, eight hose directions and six signed sensor axes, |e| 0.2 N, H 0, 5 and 20 N, F_r 0: with milestone 39's compensator ('unaware', which takes loose joints as zero mean) and one that predicts the set sides from its own plan ('aware'). (D) The no-hose clean-sheet stack per pose with the spec's window torques: loose joints uniform (as milestones 30–39) against set joints carrying no random part.
- Disclosed before registering. Only milestones 37–39's published numbers and code and the sources above have been read; no torque below 25 N of press has been computed. The press is quasi-static (no recoil transient is simulated; a joint the transient leaves set and the re-press does not reach stays loose, which is what it was). The recoil figures are the bracket above, used only to say whether lift-off happens.
- Hypotheses, registered before the run. - H151 (the re-press sets most loose joints). F_r 0, H 0, exact sensor: over both robots' poses, at least half the nose-moving joints loose in the window are set by the re-press. Refuted if fewer. - H152 (milestone 39's compensator loses the spec). 'braked', F_r 0, |e| 0.2 N, a 5 N drag: the G1's worst residual along the wall with the unaware compensator exceeds 0.1 mm (milestone 39: 0.012, with uniform loose joints). Refuted if 0.1 mm or less. - H153 (an aware compensator keeps it). Same cell with the aware compensator: at most 0.1 mm. Refuted if more. - H154 (the stack gains). 'braked': the no-hose stack at the G1's worst pose rises by at least 0.5 points when set joints carry no random part. Refuted if less. - H155 (a partial release sets fewer). H151's census with F_r 10 N counts fewer set joints than with F_r 0. Refuted if as many or more.
2026-11-17 · Milestone 39 result: braking the waist takes the floor from 0.49 to 0.01 mm and a 1 arcmin waist to 0.11; locking roll and pitch, as on the base G1, changes nothing because the yaw is what flips; H146–H150 held
src/buildbot_sheathing/waist.py (new: the per-joint gap of each waist case), experiments/m39_waist/ (run, results, figure; imports milestone 38's run so every pose, plan and residual is the same code), tests/test_waist.py (3 tests), reproduce.py (m39 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card.
- Check. 'stock' reproduces milestone 38's S = 10 N residuals exactly (largest difference 0.0 mm) and its stacks to the printed digit (G1 0–10 N 5.80, 0–20 N 0.05; H1 0.02 and 0.02).
- Braked (H146 held). With the waist joints rigid and the FT 300 (|e| 0.2 N), the G1's worst residual along the wall at a 5 N drag falls from 0.485 to 0.012 mm, at 0 N from 0.479 to 0.021. The H1 at 5 N 0.501 → 0.019, at 20 N 0.423 → 0.032.
- The base G1's lock does not help (H147 held, at full strength). Making waist roll and pitch rigid (the G1 ships with only the yaw; unitree.com/g1) leaves 0.485 mm at 5 N and 5.80 points over 0–10 N: the waist joint that flips is the yaw, the one every G1 keeps.
- A precision waist gearbox nearly does it (H148 held). 1 arcmin on the waist leaves the G1 0.106 mm at 5 N (0.099 at 0 N) and the H1 0.098 and 0.090. The floor scales with the waist gap, as the held-bias model says (gap / 2 on a flipped joint).
- The stacks close (H149 held). G1 0–10 N: stock 5.80, yaw only 5.80, precision 0.11, braked 0.04 points; 0–20 N 0.05 → 0.03. The H1 stays at 0.02 in every case (its stacks were already set by other joints). The no-hose stack the loss is counted from rises slightly at the worst G1 pose (0.9081 → 0.9160) because a rigid waist adds no random error when loose.
- The old sensor (H150 held). With |e| 2.4 N a braked waist still leaves 0.873 mm at 5 N (G1) and 0.818 mm at 0 N; the 2.4 N error flips arm and leg joints by more than any preload here.
- What it means. With a current wrist F/T sensor, the leg preload (milestone 37) and a 10 N squeeze (milestone 38), the last 0.5 mm is the waist yaw's backlash. A spec can close it with a waist yaw holding brake (or lock) while nailing, or a waist gearbox at about 1 arcmin, which gives about 0.1 mm. Dropping the optional roll and pitch joints, as the base G1 does, buys nothing.
- Not measured. The brake is ideal (no gap, no compliance); the lock holds at the posed angle (waist pitch up to ±29.8°, roll ±8°, yaw ±5° over the G1's poses), whereas the base G1's roll and pitch are fixed at zero, so its poses would differ; poses are not re-solved. No waist backlash figure is published for either robot; 1 and 5 arcmin are the low end and middle of the catalogue range already cited. Statics only.
- Next. The paper's headline for a tool-point spec (sensor, leg preload, squeeze, waist brake) as one table; the nailer's recoil; the G1's over-limit wrist pitch at low poses.
2026-11-17 · Milestone 39 planned: the waist's gap, from stock to braked
- Why. Milestone 38's squeeze cleared every arm flip and left a floor of 0.48–0.50 mm (G1 at 0 and 5 N of hose drag, H1 at 5 N) that sits in the waist (G1) and torso (H1) joints and two leg joints. No ground reaction and no squeeze moves a waist torque: the tool's wrench alone sets it. What is left to an engineer is the waist's own hardware: its gearbox backlash, or a lock. Unitree sells the G1 with one waist joint (yaw) and the roll and pitch joints as an option ("Waist Degrees of Freedom 1, 1+(Optional 2 additional waist degrees of freedom)", unitree.com/g1, read 2026-11-17); Menagerie's G1 model has all three.
- Model. Milestone 38's setup at its S* = 10 N squeeze and m = 0.1 N m leg preload (settle model, rubber hose, k_t 1e4, one calibration at milestone 30's pose in the same hose state and sensor error). New: the total gap is set per joint. Waist cases: 'stock' (5 arcmin on every joint, milestone 38), 'yaw only' (G1: waist roll and pitch rigid at the posed angle, as on the one-waist-joint G1; yaw 5 arcmin), 'precision' (1 arcmin on the waist joints, the low end of the 1–7 arcmin catalogue precision-planetary range the paper already cites) and 'braked' (0 on the waist joints: a holding brake or a lock at the posed angle). Every other joint keeps 5 arcmin. Torques do not depend on the gap, so milestone 38's flip census stands; what changes is what a flip costs. (B) the worst residual along the wall and vertically over poses, eight hose directions and six signed sensor axes, |e| 0.2 and 2.4 N, H 0, 5 and 20 N, per case; (D) the 0–10 and 0–20 N stacks with the FT 300 per case, and the no-hose stack the loss is counted from.
- Disclosed before registering. Only milestone 38's published numbers, the code and the waist joint names and ranges in the two MJCFs have been read; no residual or stack with a per-joint gap has been computed. Locking at the posed angle is not the same as the one-waist-joint G1, whose roll and pitch are fixed at zero; the poses are not re-solved.
- Hypotheses, registered before the run. - H146 (a braked waist clears the G1's floor). 'braked', |e| 0.2 N, a 5 N drag: the G1's worst residual along the wall is at most 0.1 mm (milestone 38: 0.485). Refuted if more. - H147 (yaw alone keeps most of it). 'yaw only', same cell: the G1's worst residual is at least half the 'stock' figure. Refuted if less. - H148 (a precision waist gearbox is nearly as good). 'precision', same cell: at most 0.15 mm. Refuted if more. - H149 (the stacks close). 'braked': the G1's 0–10 N stack loses at most 1 point (milestone 38: 5.80) and every other robot and range at most 0.1. Refuted if either fails. - H150 (the old sensor stays unusable). 'braked', |e| 2.4 N, a 5 N drag: the G1 still leaves more than 0.5 mm. Refuted if 0.5 mm or less.
2026-11-17 · Milestone 38 result: a two-hand squeeze clears every arm flip and takes 0–20 N of hose drag from 1.6 points lost to 0.05; the waist joints set the floor that is left (0.48–0.50 mm); H141, H143, H144 and H145 held, H142 refuted
src/buildbot_sheathing/squeeze.py (new: the squeeze's direction and generalised force, the controller's grid choice, the MJCF torque limits), experiments/m38_squeeze/ (run, results, figure), tests/test_squeeze.py (4 tests), reproduce.py (m38 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card.
- Checks. With S = 0 every residual and stack equals milestone 37's m = 0.1 row to the printed digit (G1 5 N 0.830 mm, 0–20 N 1.64 points; H1 2.19). Per newton of squeeze, the base's generalised force is at most 2.2e-16 and every leg and waist torque 1.1e-16 N m, while an arm joint moves by up to 0.47 N m (H141 held): the squeeze is a pure arm preload, as the statics say.
- One post hoc change, disclosed. The registered choice kept every arm joint inside its MJCF torque limit. The smoke run (before the full run) found that the G1's right wrist pitch already carries more than its 5 N m rating with no squeeze at 7 of 14 poses (the low ones, z 0.4–0.8 m; up to 1.21 times the limit), so the registered rule had no solution there. The run instead forbids the squeeze from pushing any arm joint beyond the larger of its limit and its unsqueezed torque. That over-limit wrist is itself a finding: holding the 3.4 kg nailer with its nose 0.3 m ahead of the grip exceeds the G1's listed wrist-pitch torque at low poses.
- The arm flips go (H143 held). At |e| 0.2 N (FT 300) over both robots and H 0, 5 and 20 N, arm flips fall from 14 to 0 at S = 10, 20 and 40 N. The waist count rises from 7 to 9 and the other leg from 1 to 2. The squeeze cannot move those joints' torques; the census takes each pose's worst hose direction and sensor axis, and those change once the arm flips are gone.
- Where it helps and where it does not (H144 held, barely). G1, |e| 0.2 N: 5 N drag 0.830 → 0.485 mm, 20 N 0.285 → 0.012; 0 N stays 0.479. H1: 0 N 0.373 → 0.010; 5 N stays 0.501, 20 N 0.423. Every cell that stays put is a waist (G1 waist, H1 torso) or leg flip. The floor left is 0.48–0.50 mm at worst.
- Stacks. 0–20 N with the FT 300's noise: G1 1.64 → 0.05 points, H1 2.19 → 0.02. 0–10 N: H1 6.92 → 0.02; G1 6.56 → 5.80 (the worst pose moves to z 0.4 m, dy 0.15 m, where the waist sets it). S* = 10 N, the smallest bound that reaches section B's best.
- The margin it buys (H142 refuted). With an exact sensor at H 0 the smallest arm margin over the G1's poses is 0.040 N m at S 10 N, 0.070 at 20 N and 0.129 at 40 N (13 of 14 poses reach 0.1 N m at 20 N). The H1's four moving arm joints reach 0.71 N m at 10 N. The controller takes the full bound every time (it maximises the margin), at up to 0.99 of the H1's arm limits at S 40.
- The old sensor gets worse in some cells. With |e| 2.4 N the squeeze raises the worst residual at H 0 (G1 0.566 → 0.818 mm, H1 0.504 → 0.928) and leaves 1.09–1.68 mm at 20 N. A 2.4 N error moves arm torques by more than the margins here, and the squeeze changes which joints sit on which side. Not diagnosed further; the rule stands that a pre-2017-class sensor is not usable.
- What it means. With a current wrist F/T sensor, the robot should hold its leg joints 0.1 N m clear of the threshold through the ground reaction (milestone 37) and its arm joints clear through a two-hand squeeze of 10 N or more along the grip line. Together they take 0–20 N of hose drag to 0.05 points. What remains is the waist (and the H1's torso), whose torque only the tool's wrench sets. Clearing it would mean a different grip (moving where the hands take the wrench) or less backlash in the waist.
- Not measured. No hand grip force is published for either robot, so S is swept. The squeeze is modelled as an ideal force pair at the grip points (no slip, no tool compliance). The settle model and tau_f (0.2 N m, unsourced) are milestone 37's. Statics only.
- Next. The paper's headline for a tool-point spec (the stack of remedies: sensor, leg preload, squeeze); a grip that changes the waist's share; the nailer's recoil.
2026-11-17 · Milestone 38 planned: a two-hand squeeze as the arm-joint preload
- Why. Milestone 37's leg preload (m = 0.1 N m) took the hold set out of the backlash lottery, but left 0.3–0.8 mm from 7 waist and 14 arm joints that sit near the 0.2 N m holding threshold tau_f, which no choice of ground reaction can move. Its 'Next' named an internal squeeze: the robot holds the tool in two hands, so it can push the hands towards (or away from) each other along the line between them. That force pair has zero net force and zero net moment, so it leaves the tool's wrench, the ground reaction, the leg torques and the waist torques unchanged and moves only the arm joints. It is the arms' counterpart of milestone 37's preload, and it cannot reach the waist.
- Model. Milestone 37's setup and settle model (G1 and H1 leaned poses, 'hold-min' sides, leg preload m = 0.1 N m, rubber hose, k_t 1e4, one calibration at milestone 30's pose in the same hose state and sensor error). New: a squeeze s (N) along the unit vector between the two grip points, +s on the left hand and −s on the right, held constant through the press. The controller picks s from its plan (measured force) by a grid over [−S, S] (0.05 N steps) that maximises the smallest distance ||tau_j| − tau_f| over the arm joints that move the nose, at both press ends, subject to every arm joint staying inside its actuator's torque limit in the Menagerie model; the same s acts on the robot when it settles under the true force. S in {0, 10, 20, 40} N (no grip-force figure is published for either robot's hand, so S is swept, not sourced). (A) the flip census (hold set, other leg, waist, arm) at |e| 0.2 N; (B) the worst residual along the wall and vertically over poses, eight hose directions and six signed sensor axes, |e| 0.2 and 2.4 N, H 0, 5 and 20 N, per S; (B2) the arm margin the squeeze achieves and the |s| it uses at H 0 with an exact sensor; (D) the 0–10 and 0–20 N stacks with the FT 300 at the S that section B picks.
- Disclosed before registering. Only milestone 37's published numbers and the existing code have been read; no squeeze, census or residual with a squeeze has been computed.
- Hypotheses, registered before the run. - H141 (the squeeze reaches only the arms). At every pose the squeeze's generalised force has base components and waist and leg torques below 1e-9 per newton. Refuted if any exceeds it. - H142 (a squeeze clears the arms). With S = 20 N, an exact sensor and H 0, the smallest arm margin is at least 0.1 N m at every G1 pose with a plan. Refuted if any pose falls short. - H143 (the arm flips go). With S = 20 N and m = 0.1 N m, |e| 0.2 N, both robots and H 0, 5 and 20 N, the census counts at most 4 arm flips (milestone 37: 14). Refuted if more. - H144 (the floor drops). With S = 20 N, m = 0.1 N m, |e| 0.2 N and a 5 N drag, the G1's worst residual along the wall is at most 0.5 mm (milestone 37: 0.830). Refuted if more. - H145 (the waist is what is left). In H143's census, at least half of the flips that remain are waist joints. Refuted if fewer.
2026-11-17 · Milestone 37 result: a 0.1 N m preload takes the legs out of the lottery and gives back milestone 36's numbers under a stricter error model; the floor that is left is in the waist and arms; H136 and H139 held, H137, H138 and H140 refuted
src/buildbot_sheathing/preload.py (new: margin rows, the margin QP, the settle model, the hedged bias, flip test), experiments/m37_preload/ (run, results, figure), tests/test_preload.py (4 tests), reproduce.py (m37 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card.
- Checks. With an exact sensor the settled torques equal the plan to 1e-14 N m; the m = 0 plan reproduces milestone 36's 'hold-min' torques exactly; at m = 0.1 every G1 leg joint at the reference pose sits 0.1000 N m or more from tau_f.
- The stricter error model costs a lot without a preload. Planning once from the measured force and settling under the true one (instead of milestone 36's re-solve) leaves the G1 at a 5 N drag with |e| 0.2 N 1.378 mm along the wall (milestone 36: 0.824), and the 0–20 N stack loses 56.5 points (milestone 36: 1.64). The reason (H136 held): 'hold-min' pins its held joints exactly at tau_f, the active constraint, so the unmeasured settling moves them across. 161 of 175 flipped joint-pose cases (92%) are the hold set, 3 other leg joints, 6 waist, 5 arm.
- The preload removes the leg flips (B). At m = 0.1 N m no hold-set joint flips (0 of 22; 1 other leg, 7 waist, 14 arm). G1, |e| 0.2 N: 0 N 1.031 → 0.479 mm, 5 N 1.378 → 0.830, 20 N 0.852 → 0.285; H1 5 N 0.765 → 0.501, 20 N 0.943 → 0.423. These match milestone 36's re-solve figures (0.479, 0.824, 0.293; 0.502, 0.418) to within 0.01 mm: milestone 36's convention silently assumed what the preload buys. The stack comes back too: G1 0–20 N 56.48 → 1.64 points, 0–10 N 61.71 → 6.56; H1 11.96 → 2.19 and 6.92.
- But not below the waist-and-arm floor (H137 refuted: 0.830 mm). What is left is joints whose torque the ground reaction cannot move (the waist and arms carry the tool's wrench alone); m 0.15 changes nothing more.
- Cost (H138 refuted). m = 0.1 raises the G1's largest leg torque at any pose by 1.66%, but loses one of 15 G1 poses (z 1.0, dy 0.15 m: no plan meets every margin); the H1 keeps all 20 and its largest rise is 24% (at a pose with small leg torques).
- The old sensor (H139 held). With |e| 2.4 N, m = 0.1 still leaves 1.350 mm at 5 N (G1); 2.4 N moves torques by more than any margin below tau_f.
- The hedge does not halve it (H140 refuted). A quarter gap on joints within 0.05 N m of tau_f leaves 1.030 mm against 1.378 (G1, 5 N, m 0); across the table it helps in some cells and hurts in others.
- What it means. A spec that uses model-based backlash compensation should have the balance controller hold its held joints at least about 0.1 N m clear of the threshold (half of tau_f here), which costs a few percent of leg torque and one marginal pose. The remaining 0.3–0.8 mm comes from waist and arm joints at the threshold, which only a change of the tool's wrench at the hands can move.
- Not measured. The settle model (min-norm change of contact forces) stands in for joint-level feedback; tau_f is milestone 30's 0.2 N m (unsourced) with a sharp threshold; statics only.
- Next. An internal squeeze between two hands on the tool as the arm-and-waist preload; the paper's headline for a tool-point spec; the nailer's recoil.
2026-11-17 · Milestone 37 planned: a held-joint preload against the sensor's floor
- Why. Milestone 36 left a floor of 0.3–0.5 mm with a good wrist F/T sensor: any measurement error flips a joint whose torque sits 0.001–0.05 N m from the 0.2 N m holding threshold tau_f, and the compensator then subtracts (or leaves) half a gap. Its 'Next' named a preload: a balance controller can move torque between the legs through the internal ground-reaction forces (milestone 32), so it can command each leg joint's torque at least a margin m away from tau_f, on whichever side it already is. The waist and arm torques are set by the tool's wrench alone (no choice of ground reaction changes them,
grf_qp docstring), so a preload cannot reach a flip there. - A more honest error model. Milestone 36 took the robot's torques as the controller's plan under the true force and the compensator's as the plan under the measured force, as if the controller knew the truth. Here the controller (and the compensator, which shares its model) plans once from the measured force f + e; the robot then settles under the true force f, modelled as the smallest change of the planned contact forces that restores the base's balance (min-norm,
pinv(Aeq)), and its joint torques follow. This 'settle' model is a stand-in for the joint-level feedback, not a measurement; milestone 36's 're-solve' convention is reported alongside at m = 0. - Model. Milestone 36's setup (G1 and H1 leaned poses, 'hold-min' sides, rubber hose, k_t 1e4, one calibration at milestone 30's pose in the same hose state and sensor error). Controller 'hold-margin m': milestone 32's 'hold-min' QP, then for every leg joint and each press end, the end's torque pushed to at least tau_f + m if it is at or above tau_f, else to at most tau_f − m in magnitude; m in {0, 0.02, 0.05, 0.1, 0.2, 0.3} N m. (A) which joints flip at m = 0 (hold set, other leg, waist, arm); (B) the worst residual along the wall and vertically over poses, eight hose directions and six signed sensor axes, for the FT 300 (|e| 0.2 N) and the pre-2017 FT 300 (2.4 N), at H 0, 5 and 20 N, per m; what it costs (poses that lose a solution, the rise in the largest leg torque); (C) a hedged compensator that subtracts a quarter gap, not half, on joints whose planned |tau| is within a band g of tau_f (g 0.02–0.1 N m), with and without the preload; (D) the 0–20 N stack with the FT 300 at the m that section B picks.
- Disclosed before registering. Only milestone 36's published numbers and the
grf_qp code have been read; no residual, flip classification or preloaded solution has been computed. - Hypotheses, registered before the run. - H136 (the flips are in the legs). Under the settle model at m = 0, with |e| 0.2 N, over both robots and H 0, 5 and 20 N, at least half of the mispredicted joint-pose cases (wrong, spurious or missed, summed over the worst sensor axis and direction per pose) are leg joints. Refuted if fewer than half. - H137 (a preload removes the floor). With the FT 300 (|e| 0.2 N) and m = 0.1 N m, a 5 N drag leaves the G1's worst pose at most 0.1 mm along the wall. Refuted if more. - H138 (the preload is cheap). m = 0.1 N m loses no G1 pose that m = 0 solves at H 0 and raises the largest leg-joint |tau| at any G1 pose by at most 10%. Refuted if either fails. - H139 (the old sensor needs more). With |e| 2.4 N, m = 0.1 N m still leaves more than 0.3 mm at a 5 N drag (G1). Refuted if 0.3 mm or less. - H140 (hedging halves it). Without a preload, the hedged compensator (g 0.05 N m) leaves the G1 at a 5 N drag with |e| 0.2 N at most half of the plain compensator's worst |y|. Refuted if more than half.
2026-11-16 · Milestone 36 result: a wrist F/T sensor removes the hose's windup but not its backlash lottery; any sensor error flips a held joint sitting at the friction threshold; H132 and H134 refuted, H133 and H135 held
src/buildbot_sheathing/wrist_ft.py (new: the joint error with a measured load and a misjudged stiffness, the listed sensors, span arithmetic), experiments/m36_wrist_ft/ (run, results, figure), tests/test_wrist_ft.py (4 tests), reproduce.py (m36 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card.
- Checks. The no-sensor case reproduces milestone 35's stacks to the printed digit (G1 2.17, 20.66, 91.55; H1 4.39, 34.28, 67.81 points). An exact sensor with k_t known leaves 0 everywhere.
- The sensor removes most of it (A). Worst residual along the wall at the worst G1 pose, no sensor against a 0.2 N error (2 sigma of the current Robotiq FT 300): H 0 0.49 → 0.48 mm, 5 N 0.97 → 0.82, 10 N 2.09 → 0.48, 20 N 2.34 → 0.29. The pre-2017 FT 300 (2.4 N) leaves 0.75–1.33 mm (H133 held). H1: 0.37–0.50 mm with the FT 300, 0.51–0.86 mm with the old one.
- But not below about 0.3–0.5 mm, whatever the sensor (H132 refuted: 0.824 mm at 5 N). At H 0 a 0.125 N error leaves 0.478 mm, a 1 N error 0.497 mm. The post hoc section A2 (added after section A's first rows printed, not registered) shows why: at 5 N a 0.001 N error leaves 0.000 mm, 0.01 N already 0.267 mm (one joint predicted held that is loose), 0.2 N 0.824 mm (one spurious, one missed). The flipped joints sit 0.0014–0.05 N m from the 0.2 N m holding threshold tau_f. A sharp threshold turns any measurement error into a half-gap error on a joint at its edge; that is a property of milestone 30's held/loose model (tau_f unsourced, swept 0.05–0.5 there), and a real joint's transition is probably gradual, so this floor is an upper bound on the model's brittleness, not a measurement.
- The stiffness matters less than registered (H134 refuted). With an exact sensor and k_t misjudged by +25% or −25%, a 5 N drag leaves 0.060 and 0.100 mm along the wall (0.17 and 0.28 mm vertically); 20 N 0.24 and 0.39 mm. One calibration takes out most of the misjudged windup because it is a near-common bias at 5 N.
- Stacks (C). With the drag uniform over 0–20 N, the FT 300 class costs the worst G1 pose 1.64 points (H135 held; 91.55 without the sensor), the Mini45 class 0.02, the pre-2017 FT 300 33.6. Over 0–10 N both current sensors cost 6.57 points (20.66 without); the pre-2017 one 87 (worse than no sensor: 2.4 N of bias costs more than a hose it does not know). The exact sensor with k_t ±25% costs at most 1.22 points (G1, 0–20 N, −25%). The losses are not monotone in H_max, because which joints sit at the threshold changes with the hose state.
- What it means. A wrist F/T sensor is the right remedy for the windup and the controller's balance (with k_t known to 25% it leaves at most 0.4 mm at 20 N), but it does not make model-based backlash compensation robust: a compensator that subtracts half a gap from every joint it predicts held is a lottery on the few joints near the threshold, and those exist at most poses. A spec should either preload those joints away from the threshold (a commanded bias torque) or drop backlash compensation and buy joints with less backlash. A pre-2017-class sensor (1.2 N noise) is worse than none.
- Not measured. Sensor bias drift is swept, not sourced; the Mini45's figure is its resolution, used as a noise proxy (no accuracy is published); moment errors are not modelled; noise-driven flips per shot are not in the stack (the noise enters linearised); tau_f is milestone 30's 0.2 N m; statics only. The H1's 0.4 m, −0.15 m pose has no 'hold-min' solution and is skipped.
- Next. A held-joint preload (a commanded bias torque that keeps each held joint at least a margin above tau_f) and the margin it needs against a given sensor; the paper's headline for a tool-point spec; the nailer's recoil.
2026-11-16 · Milestone 36 planned: a wrist force-torque sensor as the hose's remedy
- Why. Milestone 35 left the matched compensator 1.0 mm along the wall at a 5 N hose drag, and named three remedies: take the hose off the tool, keep the drag small, or measure the tool's wrench. The third is the one a controls engineer reaches for. A wrist F/T sensor turns the unknown hose force into a measured one: the compensator predicts each held joint's side and the transmission windup from the measured wrench instead of the no-hose model. What is left is the sensor's error and any error in the joints' stiffness k_t, which no datasheet gives (milestone 34).
- Sources. Robotiq FT 300 (instruction manual, specifications table): signal noise 0.1 N on Fx, Fy, Fz (standard deviation over 1 s), 0.005 N m on Mx, My; the version sold before 10/2017 listed 1.2 N (Fx, Fy) and 0.5 N (Fz). ATI Mini45, SI-145-5 calibration (ati-ia.com product page, read 2026-11-16): resolution 1/16 N on Fx, Fy and Fz after dropping four counts of noise; no accuracy figure is published ("contact your sales representative"). A sensor's bias drift between tares is not published by either, so the error magnitude is also swept (0.125-2.4 N).
- Model. On milestone 35's setup (G1 and H1 leaned poses on 'hold-min', rubber hose, k_t 1e4), the robot's actual torques are the QP solution under the true hose force f; the compensator's are the QP solution under the measured force f + e. The joint-space error is the backlash misprediction delta(t) - delta(t_hat) plus the windup error (the true windup at k_t minus the predicted one at k_t (1 + eps)), calibrated once at milestone 30's pose in the same hose state with the same sensor error. (A) the residual along the wall and vertically, worst pose, worst of the six signed axis directions of e (|e| = 2 sigma: ATI 0.125 N, FT 300 0.2 N, pre-2017 FT 300 2.4 N; also 0.5 and 1 N) at H 0, 5 (eight directions phi), 10 and 20 N (milestone 35's worst phi); (B) k_t misjudged by +-25% with an exact sensor; (C) the stack with H uniform over [0, H_max]: the sensor's noise per shot as a scatter (linearised through the QP per axis), the k_t error's share of the windup spread as a scatter, and the residual at H_max / 2 as a bias, against milestone 35's unmodelled stack. Noise-driven backlash flips per shot are not modelled.
- Disclosed before registering. Only the two datasheets above and milestones 33-35's published numbers have been read; no residual, stack or sensor case has been computed.
- Hypotheses, registered before the run. - H132 (a sensor fixes the hose). With a 0.2 N error (2 sigma of the current FT 300) and k_t known, a 5 N drag in the worst direction leaves the matched compensator at most 0.1 mm along the wall at the worst G1 pose. Refuted if more. - H133 (the old sensor does not). With a 2.4 N error (2 sigma of the pre-2017 FT 300) and k_t known, the same case leaves more than 0.3 mm. Refuted if 0.3 mm or less. - H134 (the stiffness must be known too). With an exact sensor and k_t misjudged by 25% (either way), a 5 N drag leaves more than 0.3 mm along the wall at the worst G1 pose. Refuted if 0.3 mm or less in both directions of the error. - H135 (the stack recovers). With the current FT 300's noise and k_t known, a drag uniform over 0-20 N costs the G1 at most 2 points of stack at the worst pose (milestone 35, unmodelled: 91.6). Refuted if more.
2026-11-16 · Milestone 35 result: the hose is a force, not a mass; a 5 N drag winds the joints up 1.25 mm and leaves the matched compensator 1.0 mm; H128 and H131 refuted, H129 and H130 held
src/buildbot_sheathing/hose.py (new: catenary hanging length and touchdown, the floor's drag limit, the force at the fitting), experiments/m35_hose/ (run, results, figure), tests/test_hose.py (5 tests), reproduce.py (m35 listed), paper/draft.md (section 5.11 paragraph), paper/check_numbers.py, README, PROPOSAL and project card.
- Checks. A zero hose force reproduces the no-hose torques exactly (0.0e+00 N m); the catenary's closed-form length matches a numerical arc length of the curve to 1e-6 (test). Milestone 34 re-ran character for character in this session's environment (mujoco 4.0.1) before this run.
- The force (A). From the listed weights, a hanging hose pulls 0.57–2.38 N (hybrid) or 1.04–4.30 N (rubber) down at the grip over the poses' grip heights (0.39–1.19 m on the G1), i.e. 0.06–0.44 kg. The floor holds 0.9–11.9 N (hybrid) and 1.6–21.5 N (rubber) of horizontal tension over mu_f 0.3–0.8 and 2–10 m dragged; a typical case (mu_f 0.5, 5 m) is 3.7 and 6.7 N.
- Transmission windup (B; H130 held). At k_t 1e4 N m/rad a 5 N pull in the worst direction moves the G1's free nose 1.25 mm in the wall plane (rubber; hybrid 1.07), 2.09 mm at 10 N, 2.49 mm at 20 N; hanging alone 0.33 mm. H1 1.20 mm at 5 N, 3.01 mm at 20 N. The encoders do not see it.
- The worst direction is towards the wall (H131 refuted). At 5 N the worst G1 pose winds up 1.251 mm for a pull towards the wall (phi 0), 1.131 mm at 315°, and at most 0.921 mm along the wall (270°). A pull along the wall normal acts like a change of press, which milestone 34 found the costliest load.
- The matched compensator (C). With the hose unmodelled and one calibration at milestone 30's pose (hose in the same state there): hanging rubber hose 0.474 mm along the wall at the worst G1 pose, from 3 joint-pose cases missed (H128 refuted; m33's +0.4 kg magazine left 0.017 mm because its mispredicted joints had short levers, and the hose's weight also changes with grip height); 5 N 1.001 mm (rubber) and 1.082 mm (hybrid) (H129 held), with 1/9/13 wrong/spurious/missed; 10 N 1.9 mm. H1: 0.18 mm hanging, 1.17 mm at 5 N, 1.4–1.5 mm at 10–20 N.
- Stacks (D). 12 bits, 5 arcmin, k_t 1e4, the rubber hose's tension uniform over [0, H_max] in C's worst direction, the backlash residual and mean windup at H_max/2 as a bias and the windup's spread as a scatter: the worst pose loses 2.17 points at H_max 5 N (H1 4.39), 20.7 at 10 N (H1 34.3) and 91.6 at 20 N (H1 67.8; G1 on the 13 poses that balance).
- Balance. 'hold-min' has no solution in 19–22 of the G1's 120 pose-direction cases at 20 N and in up to 33 of the H1's: a hard drag is a balance load too.
- Servo droop (E). Under Unitree's PD gains the hanging hose droops the G1's free nose 29–53 mm, a 5 N pull 175–205 mm, 20 N 369–406 mm (H1 18–32, 134–152, 354–382 mm). The encoders see this; re-targeting on them removes it, as in milestone 34.
- What it means. Milestone 33's rule (compensate backlash with the robot's own controller model) survives a magazine but not a hose: the hose is an unknown, pose- and history-dependent force, and above about 2 N of drag it leaves more than the 0.3 mm a compensator was meant to reach. A spec has three options: take the hose off the tool (a boom or balancer that brings it in from above, so H stays near 0 and its weight is carried elsewhere); measure the tool's wrench (a wrist force-torque sensor would turn the hose into a known load; not modelled here); or keep the drag under about 2–5 N (short hose on the floor, a slick floor, a hose reel that pays out). A hose dragged across 10 m of floor can cost 20–90 points.
- Not measured. Hose weights are listings (fittings included); floor friction and dragged length are swept; the hose's bending stiffness and its own torsion are neglected; the fitting is taken at the grip (on a framing nailer it sits at the end of the handle); H is the same at the calibration pose and the nail; backlash flips within the H range are taken at H_max/2 only (the stack's bias); statics only; the H1's 0.4 m, −0.15 m pose has no 'hold-min' solution and is skipped.
- Next. The paper's headline for a tool-point spec (backlash, compensation, windup, hose); a wrist force-torque sensor as the remedy; the nailer's recoil.
2026-11-16 · Milestone 35 planned: the air hose as a force on the tool (its hanging weight and its drag along the floor)
- Why. Milestones 33–34 treated the hose only as mass (a 1.0 kg unsourced allowance). A hose is not a mass: it hangs from the tool's air fitting to the floor and, once the robot has walked along the wall, the length lying on the floor drags behind it. At the fitting it pulls down with the weight of the hanging length and sideways with whatever horizontal tension the floor friction holds, in the direction the hose runs. Neither the compensator (milestone 33) nor a fixed feedforward knows that tension, and it changes from nail to nail as the robot moves.
- Sources. 3/8 in × 50 ft hoses, as listed: hybrid polymer (Flexzilla HFZ3850YW2, 5.1 lb, toolsource.com; Flexair, 5.08 lb, lawsonproducts.com) and rubber (Workforce, 9.21 lb, kbctools.ca; Tekton, 7.30 lb), so w = 0.152 kg/m (hybrid) and 0.274 kg/m (rubber), fittings included. The hose-on-floor friction coefficient and the dragged length are not sourced: mu_f swept 0.3–0.8, length 0–10 m.
- Model. The hanging part is a catenary from the fitting (at the grip, height h above the floor) to a touchdown point where it lies flat: with horizontal tension H and weight per length w g, the hanging length is s = sqrt(h² + 2hH/(wg)) and the fitting carries (H cos phi, H sin phi, −wg s). H runs from 0 (hose lying loose) to mu_f w g L_drag (the floor length about to slide). The force acts at the grip; the nose's press is unchanged. On milestone 32's leaned poses, robot on 'hold-min' (milestone 33's matched case): (A) the force per pose; (B) its transmission windup at k_t 1e4 (free nose, before contact; invisible to the encoders); (C) the matched backlash compensator's residual and mispredictions with the hose unmodelled, H from 0 to 20 N, phi swept; (D) the stack with H uniform over [0, H_max] as a scatter; (E) servo droop under Unitree's gains (visible to the encoders), for reference.
- Disclosed before registering. Only the hose listings above and milestones 32–34's published numbers have been read; no hose force, windup, residual or stack has been computed.
- Hypotheses, registered before the run. - H128 (a hanging hose is a magazine). With H = 0 the rubber hose at the worst G1 pose leaves the matched compensator a residual of at most 0.1 mm along the wall after one calibration. Refuted if more. - H129 (a dragged hose is not). A 5 N horizontal pull in the worst direction leaves the matched compensator more than 0.3 mm along the wall at the worst G1 pose. Refuted if 0.3 mm or less. - H130 (the drag winds up the joints). At k_t 1e4 N m/rad, a 5 N horizontal pull in the worst direction moves the G1's free nose by more than 0.5 mm in the wall plane at the worst pose. Refuted if 0.5 mm or less. - H131 (along the wall is worst). Of the horizontal directions, a pull along the wall (±y) gives the largest windup at the nose at the worst G1 pose. Refuted if another direction (away from or towards the wall, or diagonal) is larger by more than 10%.
2026-11-16 · Milestone 34 result: a nose free to slide needs joints stiffer than about 1.8e4 N m/rad; friction holds it at 10 of 15 G1 poses; an unmodelled magazine droops the servo by 32–51 mm; H124 refuted, H125–H127 held
src/buildbot_sheathing/windup.py (new: free windup, the nose stuck in the wall plane, the friction-limited slide, the stick test, the press scatter added to a ToolError), experiments/m34_windup/ (run, results, figure), tests/test_windup.py (5 tests), reproduce.py (m34 listed), paper/draft.md (section 5.11 paragraph, the tool-point limitation, the conclusion), paper/check_numbers.py, README, PROPOSAL and project card.
- A change after the first run (disclosed). The first run pinned all three of the nose's translations. The pinning reaction then cancelled the press itself (friction ratios up to 5e9): along the wall normal the press is force-controlled, so x cannot be pinned. The model now pins only y and z (the controller closes any x windup with a commanded motion along x) and adds a friction-limited case: the stuck reaction grows with the press from zero, so the ratio |lam|/F is fixed per pose, and above mu the nose slides 1 − mu/ratio of its free windup (reaction taken along the stuck direction). H124 is stated about the in-plane reaction, so it is tested as registered; the first run's numbers are not used.
- Checks. The stuck nose's in-plane windup is 1e-15 mm (zero); the sticking reaction differs by 1.8e-15 N between k_t 3e3 and 1e5 (a common k_t cancels, as derived).
- Windup over the 25–50 N window, free nose. Worst G1 pose (1.2 m, +0.15 m): 10.6 mm at k_t 3e3, 3.19 mm at 1e4, 1.06 mm at 3e4, 0.32 mm at 1e5 (H1 3.47 mm at 1e4). Stuck, the nail's tip at 44.4 mm moves 0.018 mm at 1e4 (H1 0.037 mm): the tool only rotates about the nose.
- Does the nose stick (H124 refuted). The G1 needs a friction coefficient of 0.073–0.789 (median 0.383); at mu 0.4 it sticks at 10 of 15 poses (0.2: 4; 0.8: all 15). The worst is 1.2 m, −0.15 m. H1 0.026–1.166, 10 of 20 at mu 0.4.
- Stacks (12 bits, 5 arcmin, matched backlash compensation). At k_t 1e4 the press window costs the G1 5.47 points with a free nose (H125 held), 0.00 stuck (H126 held), 1.05 friction-limited at mu 0.4 (H1 4.64, 0.01, 1.82). Within 2 points at the worst pose takes k_t ≥ 1.78e4 N m/rad free and ≥ 1.0e4 at mu 0.4 (grid of quarter decades), for both robots. Below about 6e3 the stack collapses (G1 free: 23.5 points lost at 5.6e3, 80.5 at 3.2e3).
- The mean windup is a pose bias. At 37.5 N, one calibration at milestone 30's pose leaves 1.76 mm along the wall and 12.4 mm vertically at k_t 1e4 (G1; z worst at 0.4 m, where the knees and hips carry the most). It is linear in the torques, so a model computes it; with k_t misjudged by 25% a quarter is left (3.1 mm vertically at 1e4).
- Servo droop (H127 held). A full magazine left out of the feedforward moves the G1's free nose by 32.1–51.3 mm under Unitree's gains (worst 0.4 m, −0.15 m, nearly all vertical), mostly through the waist pitch and the ankle pitches (60 and 40 N m/rad); H1 17.3–22.4 mm. The servo's error is what the joint encoders read, so a robot that re-targets on its encoders, or counts the nails it has fired, removes it; one that trusts PD with a fixed feedforward does not.
- What it means. Three rules for a tool-point spec: fire on a known force or let friction hold the nose (a stuck nose turns the press's windup into a 0.02 mm tilt); ask for joints stiffer than about 1e4 N m/rad (2e4 if the nose can slide), behind the encoder; and close the loop on the encoders, since the servo is about a hundred times softer (40–150 N m/rad). A gripping nose tip (teeth, rubber) is the cheapest of the three.
- Not measured. No torsional stiffness is published for these robots, so k_t is swept and the same on every joint; links and the tool mount are rigid; the nose-to-OSB friction is swept; the slide is approximated along the stuck reaction's direction; torques are taken as linear from zero press at contact; the press scatter enters the stack as a Gaussian of the uniform's variance with the y–z correlation dropped; statics only (no recoil). The H1's 0.4 m, −0.15 m pose has no 'hold-min' solution and is skipped.
- Next. The paper's headline for a tool-point spec (now with backlash, compensation and windup); the hose as a force; the nailer's recoil.
2026-11-16 · Milestone 34 planned: deflection under the press (the joints' elastic windup, the servo's droop, and whether the nose sticks)
- Why. "Deflection under the press" has been the open item of milestones 29–33. Every joint so far was rigid apart from its backlash gap. A joint also twists elastically: through its servo (Unitree's published PD gains,
humanoid.GAINS: G1 hips 100, knees 150, ankles 40, waist and arms 60 N m/rad) against any torque the controller does not feed forward, and through its transmission (gearbox and structure, stiffness k_t) against all of it. The press is the load that changes between the nose touching the sheet and the shot (anywhere in 25–50 N, milestone 30), so its windup is not a fixed bias a calibration or milestone 33's model can remove unless the shot's force is known. - Sources looked for. No published torsional stiffness was found for the G1's or H1's joints (a search on 2026-11-16 found catalogue figures for strain-wave gears and one measured value for a 3D-printed planetary prototype, neither for these robots, and the catalogue host does not answer from this session). So k_t is swept, 3e3 to 1e5 N m/rad on every joint, and the result is stated as the stiffness a spec must ask for. The nose-to-OSB friction coefficient is not sourced either; mu swept 0.2–0.8, 0.4 the registered case.
- Plan.
src/buildbot_sheathing/windup.py: free windup -dtau/k; the nose stuck by friction (its three translations pinned: the reaction lam = (A K^-1 A^T)^-1 A K^-1 dtau, the windup survives only as a rotation of the tool about the nose, seen at the 44.4 mm depth point); the stick test |lam_yz| <= mu (F - lam_x). On milestone 32's leaned poses (G1 15, H1 20–21), with the robot balancing on 'hold-min' (milestone 33's matched case): (A) the windup over the 25–50 N press per pose, free and stuck, per k_t; (B) whether the nose sticks (per newton of press, from contact at zero press; with the same k on every joint lam does not depend on k); (C) the stack (single 2x, 12 bits, 5 arcmin backlash on loose joints, held joints compensated by the matched model) with the press uniform over the window, as a Gaussian of the same variance, against milestone 33's matched stack; the k_t that keeps the worst pose within 2 points; (D) the mean windup at 37.5 N as a pose bias: after one calibration, and after model compensation with k_t misjudged by 25%; (E) servo droop: a full 64-nail magazine of 8d nails (0.295 kg, milestone 33) left out of the feedforward, under Unitree's gains, at the nose (free, before contact). - Disclosed before registering. Only the code above and milestones 30–33's published numbers have been read; no windup, reaction, stick ratio, droop or stack has been computed at any pose.
- Hypotheses, registered before the run. - H124 (the nose sticks). At mu 0.4, the reaction that keeps the G1's nose from sliding as the press grows stays inside the friction cone at every leaned G1 pose. Refuted if it leaves the cone at any pose. - H125 (a sliding nose pays for the press). With the nose free and k_t = 1e4 N m/rad on every joint, the press window costs the 12-bit G1 more than 2 points of stack at the worst leaned pose. Refuted if 2 points or less. - H126 (sticking saves it). With the nose stuck and the same k_t, the press window costs at most 1 point at the worst leaned pose. Refuted if more than 1. - H127 (the servo droops). Under Unitree's PD gains with the listed 3.4 kg tool fed forward, an unmodelled full magazine (0.295 kg) moves the G1's free nose by more than 1 mm in the wall plane at the worst leaned pose. Refuted if 1 mm or less.
2026-11-16 · Milestone 33 result: model-based backlash compensation removes the residual only if it models the balance controller; then a full magazine costs 0.02 mm; H121 held, H122 and H123 refuted
src/buildbot_sheathing/compensate.py (new: the predicted bias with an optional guard band, misprediction classes, the residual after compensation and one calibration), experiments/m33_compensate/ (run, results, figure), tests/test_compensate.py (4 tests), reproduce.py (m33 listed), paper/draft.md (section 5.11 paragraph, conclusion), paper/check_numbers.py, README, PROPOSAL and project card.
- Checks. The matched compensator leaves 0.00e+00 mm and the worst-pose stack equals milestone 32's per-pose calibration (G1 3.74 points on 'hold-min', 4.69 on 'torque'); with nothing predicted ('none') the residual reproduces milestone 32's one calibration, −1.401 mm at 1.2 m, −0.15 m.
- The unmodelled load, computed. A 64-nail magazine (Metabo HPT listing) of 8d common nails (2-1/2 × 0.131 in) is 0.295 kg of steel, of 16d (3-1/2 × 0.131 in) 0.410 kg; +1.0 kg is an unsourced allowance for the collation and the hose.
- The controller must be in the model (H121 held). With the G1 balancing on 'hold-min', a compensator that assumes the least-norm split leaves 1.156 mm at 0.4 m, +0.15 m (17 joint-pose cases on the wrong side, 2 spurious, 47 missed over 15 poses); assuming 'force' 1.450 mm, 'torque' 1.237 mm, against 1.401 mm with no compensation at all. With the G1 on 'torque', wrong models leave 1.237–1.924 mm (no compensation 2.632). A wrong ground-reaction model buys at most 0.25 mm on the G1 when the robot holds its light joints. On the H1 the same mismatch leaves 0.475–1.128 mm (12–38 joint-pose cases on the wrong side).
- A full magazine (H122 refuted, on the count). With the controller modelled ('hold-min'), +0.4 kg mispredicts 2 joint-pose cases (1 spurious, 1 missed) and leaves 0.017 mm, so the registered "no misprediction" fails while the 0.3 mm half holds; +1.0 kg leaves 0.057 mm (4 spurious, 6 missed; 3.80 points lost at the worst pose against 3.74 matched). The 'torque' controller is far more sensitive: +0.2 kg already leaves 0.412 mm (7.58 points at that pose) and +1.0 kg 0.836 mm (19.40 points). It leaves light joints wherever the optimum puts them, often near zero, while 'hold-min' keeps them at tau_f or more. H1 matched ('hold-min'): 0.179 mm at +0.2 and +0.4 kg, 0.435 mm at +1.0 kg.
- A guard band does not help (H123 refuted). Compensating only predicted |tau| >= 1 N m leaves 1.374 mm for the least-norm compensator (wrong sides fall from 17 to 11, but missed joints rise from 47 to 79, each leaving half a gap). On the matched compensator a band of 0.5–2 N m turns 0 into 0.961–1.351 mm. A joint left uncompensated costs as much as a guessed one.
- What it means. The remedy milestones 31–32 named works, and works robustly, on one condition: the compensator predicts torques with the robot's own balance controller (the same QP, the same held set). Then an unmodelled magazine is invisible (0.017 mm) and even +1 kg costs 0.057 mm, provided the controller keeps the light leg joints one way ('hold-min'). The design rule is a sign-holding controller paired with compensation computed from that controller; a generic model of the robot's statics is not enough.
- Not measured. Statics only; the "true" robot is still the model (its gains, joint friction and gap asymmetry are not measured); the poses do not re-lean for extra mass; the H1's 0.4 m, −0.15 m pose has no 'hold-min' solution and is skipped (20 of 21 poses). Mass enters at the tool only.
- Next. The paper's headline for a tool-point spec; deflection under the press; the hose as a force rather than a mass.
2026-11-16 · Milestone 33 planned: model-based backlash compensation, and what model error it survives (the controller, the magazine)
- Why. Milestones 31–32 concluded that one tool-point calibration cannot remove the held joints' backlash along a wall (1.1–2.6 mm left on the leaned G1 poses) and named the remedy: predict each held joint's side from the static torques at every pose and subtract the predicted bias in joint space. Neither milestone tested that remedy. A compensator predicts with its model, and two parts of that model are uncertain on a real robot: which ground-reaction split the balance controller picks (milestone 32: leg torques differ by tens of N m between splits, waist and arm torques not at all), and the tool's mass (the 3.4 kg listing is the empty tool; the magazine holds 64 nails, Metabo HPT product page, and the air hose hangs off it).
- Plan.
src/buildbot_sheathing/compensate.py: the predicted bias delta_hat(p) from a model's torques (held set and sides at tau_f 0.2 over the 25–50 N press, optionally only joints whose predicted |tau| >= a guard band tau_g at both press ends), the true bias delta(p) from the robot's torques, and the residual after compensation plus one tool-frame calibration at milestone 30's pose: mu(p) = A(p)[delta(p) − delta_hat(p)] − A(ref)[delta(ref) − delta_hat(ref)]. Counted per pose: joints mispredicted (wrong side, or predicted held but loose, or held but not compensated). On milestone 32's 15 leaned G1 poses (21 H1), 12 bits, 5 arcmin: (A) matched model (check: residual 0, stack equal to per-pose calibration); (B) controller mismatch: robot runs 'hold-min' or 'torque', compensator assumes 'free', 'force' or 'torque'; (C) mass mismatch: the robot carries 0, +0.2, +0.4 and +1.0 kg the compensator does not know (the magazine's nails computed from steel density and nail size, the 1.0 kg a hose allowance, not sourced); (D) guard band tau_g 0.2–2 N m on the mismatched cases; the stack per pose for each. - Disclosed before registering. Only milestone 32's published numbers and the code have been read; no compensated residual, misprediction count or stack has been computed.
- Hypotheses, registered before the run. - H121 (the controller must be in the model). A compensator that assumes the least-norm 'free' split while the G1 runs 'hold-min' leaves more than 1 mm along the wall at the worst grid pose after compensation and one calibration. Refuted if 1 mm or less. - H122 (a full magazine flips nothing). With the controller modelled correctly ('hold-min'), an unmodelled +0.4 kg at the tool mispredicts no G1 joint at any grid pose and leaves at most 0.3 mm. Refuted if any joint is mispredicted or the worst residual exceeds 0.3 mm. - H123 (a guard band rescues a wrong controller model). For H121's mismatched compensator, compensating only joints with predicted |tau| >= 1 N m brings the worst residual to 1 mm or less. Refuted if it stays above 1 mm.
2026-11-16 · Milestone 32 result: through real soles the upright stud poses cannot take the press; leaned into the wall the hip-yaw and waist flips vanish, but one calibration still leaves 1.1–2.6 mm; H118 and H119 refuted, H120 not testable
src/buildbot_sheathing/grf_qp.py (new: sole contact points from the Menagerie collision geoms, friction-pyramid LP and QPs, centre of pressure, press window, leaned pose), experiments/m32_grf_qp/ (run, results, figure), tests/test_grf_qp.py (4 tests), reproduce.py (m32 listed), paper/draft.md (section 5.11 paragraph, the tool-point limitation, the conclusion), paper/check_numbers.py, README, PROPOSAL and project card.
- The press needs a lean (not in the plan; found by the first solve). With each sole pushing only through its contact points (G1: four 5 mm spheres; H1: both ends of three capsules) and mu 0.5, milestones 30–31's upright poses balance at most 79.6 N at a 0.4 m hand height, 39.0 N at 0.8 m, 31.1 N at 1.0 m and 25.9 N at 1.2 m (G1; H1 100.9 to 25.2 N over 0.4–1.6 m). The free 6-D foot wrench had hidden this: above about 0.8 m the 50 N press tips the robot back onto its heels. Leaning the centre of mass 0.040–0.124 m towards the wall (secant on
com_dx until the centre of pressure sits mid-sole at 37.5 N) brings the whole 25–50 N range inside the window at all 36 poses, so every number below is on the leaned poses. The plan did not include this; it changes the poses, so milestone 31's numbers are not directly comparable. - Structural check. Waist and arm torques are identical under all five ground-reaction models (largest difference 0.00e+00 N m): their subtrees hold no foot.
- The flips (H118 refuted). On the leaned poses neither hip yaw nor the waist roll flips under any model, including milestone 31's least-norm one; the only G1 joint held everywhere that flips is the right wrist pitch (−6.04 to +3.00 N m), an arm joint no controller can touch. Milestone 31's flips came from the upright poses, not from the ground-reaction split.
- Holding every leg joint (H119 refuted). Keeping all 12 G1 leg joints on one side by 0.2 N m is infeasible at all 15 poses (worst LP margin −45.5 N m: hip pitches and knees carry tens of N m whose sign the lean sets). H120 asked about the residual under that controller and is not testable.
- Holding the light leg joints (added after the first run, not registered). Holding only the hip rolls, hip yaws and ankle rolls (6 G1 joints) is feasible at every pose (worst margin 3.31 N m). Calibrated per pose, the worst 5-arcmin loss falls from 8.63 points (least-norm) and 4.69 (torque-minimising) to 3.74. With one calibration at the 1.0 m pose the worst offset along the wall is 1.401 mm (least-norm 1.950, force 2.056, torque 2.632) and the worst pose still loses 59.59 points.
- The calibration pose (not registered). The 1.0 m pose is where most G1 joints are loose under the least-norm split (knees, waist roll and pitch). Searching every grid pose as the calibration pose: least-norm 1.949 mm at best (0.8 m), torque 1.910, light-joint hold 1.094 mm (calibrated at 1.2 m), still 36.67 points lost at the worst pose. What remains comes from joints held at some poses and loose at others, and from the wrist pitch.
- Sweeps. mu 0.3–0.8 and tau_f 0.05–0.5: the full hold is infeasible everywhere; the light-joint hold is feasible at 10/15 poses at mu 0.3 and 14–15/15 otherwise, with worst one-calibration offsets 1.29–2.18 mm. At tau_f 0.05 the waist yaw and roll flip again under the force model.
- H1. The torque-minimising controller does best: one calibration loses at most 6.65 points (0.669 mm); least-norm 65.03 points.
- What it means. A balance controller cannot remove the need for pose-by-pose compensation, but it is what makes the press possible at all above chest height, and holding the light leg joints halves the per-pose backlash loss. The engineering rule is unchanged from milestone 31 (predict each held joint's side from the static torques at every pose), and the poses it applies to must be the leaned ones.
- Not measured. Statics only: no balance dynamics or the nailer's recoil. Contact points are the model's collision geoms (a real sole's patch differs); mu swept 0.3–0.8. The QP holds a joint at exactly tau_f, the least a real controller would allow. The lean is set at 37.5 N; a controller that re-leans per press would change the held sets.
- Next. The paper's headline for a tool-point spec; deflection under the press; whether re-leaning with the press (a centre-of-pressure servo) keeps the ankle pitch held.
2026-11-16 · Milestone 32 planned: the sign flips under a balance controller's ground reaction (friction cones, the sole's corners)
- Why. Milestone 31 found three G1 joints held at every pose that flip sign along the wall (both hip yaws and the waist roll), so one tool-point calibration leaves up to 3.4 mm, and the two ankle rolls are loose at every pose. All of it used the least-norm split of a free 6-D wrench at one point per foot, which no robot applies: a real sole pushes only through its contact patch (unilateral normal force, friction cone, centre of pressure inside the sole), and a balance controller picks one force distribution out of the many that balance the body. Milestone 31's "Not measured" named this.
- A structural point, stated before running. Joint torque in a static balance is set by the loads on the joint's subtree. With the pelvis as the floating base, the waist and arm joints' subtrees hold the torso, arms and nailer but no foot, so no choice of ground reaction changes their torque: the waist-roll flip survives every controller. The leg joints (hip yaw among them, and the ankle rolls) carry the ground reaction and can be shaped. The run checks this numerically (waist and arm torques equal across controllers to 1e-9 N m) rather than assuming it.
- Plan.
src/buildbot_sheathing/grf_qp.py: (A) each foot's contact points from the Menagerie model's own collision geoms (G1: four 5 mm spheres per sole; H1: three capsules per sole, both ends), each with a 3-D force, normal force at least 0, a friction pyramid at mu 0.5 (swept 0.3–0.8); (B) three controllers, each a QP over those forces that balances the floating base exactly: force-regularised (min sum of squared contact forces, the usual whole-body-control regulariser), torque-minimising (min squared leg-joint torque, plus a 1e-6 force regulariser), and sign-holding (torque-minimising, plus every leg joint that moves the nose kept on one side, |tau| >= tau_f, with the side per joint chosen from the torque-minimising solution at milestone 30's pose, both sides tried for joints near zero there); (C) for each controller, milestone 31's grid, held sets, sign flips, per-pose stack and one-calibration residual (split into its Jacobian and flip parts), at 12 bits, 5 arcmin, a 25–50 N press, tau_f 0.2 (swept 0.05–0.5); (D) where the sign-holding QP is infeasible, which constraint binds. - Disclosed before registering. Only the Menagerie collision geometry has been looked at (the G1 sole points are at x −0.05 and +0.12 m, y ±0.025–0.03 m). No QP has been solved and no torque, held set or stack number exists under any of these controllers.
- Hypotheses, registered before the run. - H118 (cones alone keep the flips). Under the force-regularised controller with friction cones (mu 0.5), at least one of the G1's hip yaws still flips sign across the grid while held at every pose. Refuted if neither hip yaw does. - H119 (a controller can hold the legs). The sign-holding controller is feasible at every G1 grid pose at both press ends (mu 0.5, tau_f 0.2). Refuted if any pose is infeasible. - H120 (the waist keeps the millimetres). Under the sign-holding controller (where feasible), one tool-point calibration at milestone 30's pose still leaves more than 1 mm along the wall at the worst G1 pose. Refuted if 1 mm or less.
2026-11-16 · Milestone 31 result: three held G1 joints flip sign along the wall, so one calibration leaves up to 3.4 mm; per pose, 5 arcmin costs up to 9.9 points at the top of the reach; H115 refuted, H116 and H117 held
src/buildbot_sheathing/pose_range.py (new: held-joint bias, nose bias, the residual after one calibration and its Jacobian and sign-flip parts), tool_lean.AxisError (a fixed offset mu_s/mu_d, zero by default: the reference pose reproduces milestone 30's 0.9344 and 0.9134), experiments/m31_pose_range/ (run, results, figure), tests/test_pose_range.py (5 tests), reproduce.py (m31 listed), paper/draft.md (section 5.11 paragraph, the tool-point limitation, the conclusion), paper/check_numbers.py (243 of 243, was 233), README, PROPOSAL and project card.
- Grid. G1 15 poses (0.4–1.2 m, dy 0 and ±0.15 m), H1 21 (0.4–1.6 m). The plan said the H1's 0.4 m row reached at dy 0 only; the probe output in fact reached ±0.15 m there too (misread when writing the plan), so the run includes it. No hypothesis concerns the H1.
- Held sets. G1 13–20 of 22 nose-moving joints held at tau_f 0.2 (the ankle rolls loose everywhere; at 1.2 m the knees, a shoulder joint and wrist joints join them). H1 12–15 of 15.
- Per-pose calibration (H115 refuted). 5 arcmin at 12 bits costs the G1 0.33–9.93 points; worst at 1.2 m, dy 0: 0.9131 to 0.8138. H1 at most 1.42 points.
- Sign flips (H116 held). Six G1 joints are held at every pose; three of them flip: left hip yaw (−10.63 to +3.41 N m), right hip yaw (−11.33 to +2.31), waist roll (−27.83 to +19.97). H1: right shoulder yaw.
- One calibration (H117 held). At 1.0 m, dy 0, the residual is zero by construction; elsewhere the G1's nose sits up to +3.394 mm off along the wall (0.4 m, dy +0.15; stack 0.0000), the sign-flip part +3.795 mm and the Jacobian part −0.401. At dy 0 the residual is 0.11–0.44 mm from 0.4 to 0.8 m and 0.86 mm at 1.2 m; the shifted poses carry the flips. H1 up to 1.65 mm (1.6 m, dy −0.15: 0.2010).
- Spec (within 2 points of each pose's own stack, worst pose). 12 bits: 2.49 arcmin calibrated per pose, 0.44 once; 13 bits: 4.13 and 0.79.
- tau_f sweep. 0.05–0.5 N m: worst per-pose loss 9.93–10.04 points, worst one-calibration offset 3.39–3.68 mm. The conclusion does not depend on it.
- What it means. Milestone 30's "held joints calibrate out" needs the bias known at every pose. The static torques give the side of each gap, so a model-based compensation (or a calibration per stance and height band) restores the per-pose numbers; a single tool offset does not. Near the top of the G1's reach the held set thins, so the per-pose spec tightens to 2.5 arcmin at 12 bits.
- Not measured. The ground-reaction split is least-norm, and it is what sets the hip-yaw and waist torques that flip; a balance controller's split could move or remove the flips. Backlash and tau_f swept; the bias is taken as exactly half the gap (no friction or compliance inside the gap).
- Next. The flips under a balance controller's ground reaction (a QP with friction cones); deflection under the press; the paper's headline for a tool-point spec.
2026-11-16 · Milestone 31 planned: backlash across the poses along a wall (does the held set hold, and does one calibration do)
- Why. Milestone 30 found backlash costs only on joints nothing holds, at one pose (hand reach 0.4 m, height 1.0 m), and treated every held joint's backlash as a fixed bias a tool-point calibration removes. A robot nailing a stud works from about 0.4 m to the top of its reach, and a little to either side of its stance. Two things can change with the pose: which joints are loose (the stack's random part), and the sign of each held joint's torque, which sets the side of the gap it rests on. Even with no sign change, the same joint-space bias lands at a different nose offset as the Jacobian changes, so a single calibration at one pose leaves a pose-dependent bias.
- Grid, from an IK probe. G1: heights 0.4–1.2 m in 0.2 m steps, hand targets shifted 0 and ±0.15 m along the wall (15 poses; at ±0.3 m and at 1.4 m the IK misses by 1.8–37 cm, so those are out of the stock G1's stance reach). H1: 0.6–1.6 m and the same shifts (the 0.4 m row reaches at dy 0 only). The probe ran the IK only; no torque, held set or stack number has been computed at any new pose.
- Plan.
src/buildbot_sheathing/pose_range.py: (A) at each pose, milestone 30's static torques (gravity, the 3.4 kg nailer, a 25–50 N press at a 0.3 m nose, min-norm ground reaction) and held set (tau_f 0.2 N m, swept 0.05–0.5); (B) each pose's best single-2x stack at 12 bits, 5 arcmin, preloaded, with per-pose calibration (milestone 30's model at that pose's Jacobian); (C) the held joints' bias, each resting on the side its external load pushes it (half the gap, against the sign of the motor torque), carried to the nose and the depth point; with one tool-point calibration at milestone 30's pose, the residual bias at every other pose, split into the part from sign flips and the part from the Jacobian alone; (D) the stack at each pose with that residual bias as a fixed offset of every nail (a mean added to ToolError, zero by default so every earlier number reproduces); (E) a model-based calibration (the sign predicted from the torque model, the bias subtracted in joint space) as the remedy, and the largest backlash one calibration tolerates for a 2-point loss at the worst pose. - Hypotheses, registered before the run. - H115 (the held set is stable). At every pose of the G1 grid, the preloaded 12-bit G1 with 5 arcmin and per-pose calibration stays within 3 points of that pose's no-backlash stack. Refuted if any pose loses more than 3. - H116 (some held joint flips). At least one G1 joint that moves the nose and is held at every grid pose changes the sign of its torque across the grid. Refuted if none does. - H117 (one calibration is not enough). With one tool-point calibration at milestone 30's pose and 5 arcmin, the residual held-joint bias moves the G1's nose along the wall by more than 0.3 mm at the worst grid pose. Refuted if 0.3 mm or less.
2026-11-16 · Milestone 30 result: backlash costs only on joints nothing holds; for a G1 at the stud pose those are the two ankle rolls; H112 held, H113 and H114 refuted
src/buildbot_sheathing/backlash.py (new: a general tool offset, per-joint sigmas, MuJoCo static torques with a press at the nose, held/loose joints), experiments/m30_backlash/ (run, results, figure), tests/test_backlash.py (5 tests), reproduce.py (m30 listed), paper/draft.md (section 5.11 paragraph, the tool-point limitation rewritten, a conclusion sentence), paper/check_numbers.py (12 new checks), README, PROPOSAL and project card.
- Checks. At zero lateral offset the general rows match milestone 29's to 4e-19 m and the stack reproduces 0.2637. Uniform per-joint backlash draws match the Gaussian tool model for one nail to 0.0015 (three windows, 200,000 draws).
- Nose length, sourced. The NR90AES1 is 12-11/16 in tall (Metabo HPT), so the grip sits 0.29–0.32 m behind the nose: the 11-bit G1 stack is 0.279–0.235 across that bracket, 12 bits 0.937–0.929. The 0.3 m of milestone 29 stands.
- Handle offset (H114 refuted). Only the vertical offset matters, through the hand's roll about the push axis: a grip 0.15 m above the drive axis costs the 11-bit G1 8.02 points, below it gains 4.08 (not registered); for the H1 the signs flip (+8.40 up, −11.37 down). Sideways offsets change nothing along the wall. At 12 bits the effects shrink to −1.56/+0.68 (G1).
- Which joints are held. At the stud pose with the 3.4 kg nailer and a 25–50 N press, 16–19 of the G1's 22 joints that move the nose keep one sign above 0.2 N m; the two ankle rolls carry no torque under the min-norm ground reaction and the wrist roll almost none. The H1 holds all 15 at 50 N.
- The stack (12-bit G1, L 0.3 m). Unpreloaded 5 arcmin: 0.9344 to 0.7646 (H112 held, −16.98 points); 10 arcmin 0.1759. Preloaded: 0.9134 at 5 arcmin, a 2.10-point loss at both press ranges (H113 refuted, bar 2), 0.8169 at 10. With every leg joint loose: 0.8870 at 5. With the ankle rolls biased to one side (a case added after the first run, not registered): 0.9344 at every gap to 25 arcmin.
- The spec. Largest backlash for 0.90 / 0.95 clean: unpreloaded 2.67 / – at 12 bits, 5.30 / 4.13 at 13; preloaded 6.24 / – and 12.28 / 9.64; ankle rolls biased, not binding below 60 arcmin at 13 bits. 12 bits alone never reach 0.95 (0.9344).
- What it means. Backlash is a gravity question: the arm joints are held by the nailer's weight and the press, and the error that reaches the nail comes from joints a standing robot leaves unloaded. A balance controller that keeps the ankle-roll torque one-signed (a small lateral shift of the centre of pressure) takes backlash out of the tool-point budget; one that does not makes a two-stage precision gearbox (3–7 arcmin) cost up to 13 points at 12 bits. The H1 has no ankle roll and holds everything.
- Not measured. No G1 backlash figure is published (swept 0–25 arcmin against catalogue classes); the ground-reaction split between the feet is the least-norm one, not a balance controller's (the legs-loose case bounds it); the holding threshold is swept 0.05–0.5 N m; the handle's offset from the drive axis is swept, not measured; with output-side encoders backlash would show as hunting rather than a static error (see
backlash-twin). - Next. The tool-point spec in the paper's abstract and highlights; link and mount deflection under the press (a deterministic part calibrates out, the press's scatter does not); the pose range along a wall, where held joints may change sign.
2026-11-16 · Milestone 30 planned: backlash at the tool point, and which joints gravity and the press preload
- Why. Milestone 29 carried the G1's 11-bit quantisation to a 0.3 m nailer nose and found a stock G1 misses the single-2x stack (0.264), with 12-bit joints restoring 0.934. Quantisation is not the only joint error: planetary gear trains have backlash, and no G1 backlash figure is published (a search found teardown notes of two-stage planetaries at about 15:1 with output-side encoders, no number). Catalogue precision planetaries are rated 1, 3 or 5 arcmin (one stage) and 3, 5 or 7 arcmin (two stages); a search excerpt credits an unmodified quasi-direct-drive actuator with about 17 arcmin (not verified here: the thesis PDF host refused the proxy). Backlash only randomises a joint that is not held against one side of its gap: gravity on the arm and the 3.4 kg nailer, plus the press of the nose into the wall, load most joints one way at every nail, which turns their backlash into a fixed, calibratable bias. Milestone 29's nose was also a point on the hand's push axis; a real nailer's handle sits off the drive axis.
- Nose geometry, sourced where possible. Metabo HPT's product page gives the NR90AES1 (buildbot-fasten's nailer, 7.5 lb) as 20-3/4 x 4-15/16 x 12-11/16 in (L x W x H). The drive axis runs the tool's height, so a grip on the handle sits about 0.29–0.32 m behind the nose along the nail (milestone 29's 0.3 m is now inside a sourced bracket). The handle's offset from the drive axis is not on the sheet: swept 0–0.2 m, up and sideways.
- Plan.
src/buildbot_sheathing/backlash.py: (A) per-joint tool-point rows for a general tool offset t (nose = hand + omega x t; depth point t + 44.4 mm along the nail), with per-joint sigmas, reducing to milestone 29's model at zero lateral offset; (B) the right arm's and whole body's static joint torques at milestone 29's pose (MuJoCo Menagerie G1 and H1, feet planted, gravity, the nailer's weight and a press force F on the right hand, ground reaction re-solved), F swept: a joint is loose if its torque changes sign as the press varies over [F/2, F] or stays below a holding threshold tau_f (swept 0.05–0.5 N m, not sourced); loose joints get a uniform error over the gap (sigma b/sqrt(12)), held joints a fixed bias (removed by calibration); (C) a Monte Carlo of uniform per-joint draws checked against the Gaussian tool model for one nail; (D) the best single-2x stack (milestone 29's, 0.7964 at the 0.9 mm point model) with each robot's error at 11–14 bits, backlash 0–25 arcmin, unpreloaded (every joint loose) and preloaded; (E) the tool-point spec: the largest backlash at each resolution that keeps the stack at 0.90 and 0.95. - Disclosed before registering. Back-of-envelope only: 5 arcmin uniform is a 1-sigma of 0.024 deg against 12-bit quantisation's 0.025 deg; the hand's 0.1 deg roll at a 0.15 m offset is about 0.26 mm. No stack number with backlash or a lateral offset has been computed, and the torques have not been looked at.
- Hypotheses, registered before the run. - H112 (unpreloaded backlash costs as much as a bit). 5 arcmin of backlash on every G1 joint, none preloaded, costs the 12-bit G1 at L 0.3 m more than 5 points of the single-2x stack. Refuted if 5 or less. - H113 (gravity and the press preload it away). With the press swept 25–50 N and tau_f 0.2 N m, the same 5 arcmin costs the preloaded 12-bit G1 at most 2 points. Refuted if more than 2. - H114 (the handle's offset is small). A 0.15 m offset of the grip from the drive axis, in the worse of the up and sideways directions, moves the stock 11-bit G1's single-2x stack at L 0.3 m by less than 2 points. Refuted if 2 or more.
2026-11-16 · Milestone 29 result: carried to a 0.3 m nailer nose and 44 mm into the stud, the G1's 11-bit error is 1.31 mm and the best single-2x stack falls from 0.796 to 0.264; 12-bit joints restore 0.934; H109, H110 and H111 held
src/buildbot_sheathing/tool_lean.py (new), tack.py and ceiling.py (a tool= error pair; the old path unchanged, 0.7964 reproduces), experiments/m29_tool_lean/ (run, results, figure), tests/test_tool_lean.py (6 tests), reproduce.py (m29 listed), paper/draft.md (section 5.11 paragraph, a limitations bullet), paper/check_numbers.py (221 of 221, was 211), README, PROPOSAL and project card.
- Checks. L 0 reproduces
buildbot-fasten's 0.93/0.87 mm; one-nail Monte Carlo matches the bivariate normal to 0.0016; the 0.9 mm point model through the new path gives 0.7969 (old 0.7964). - The robots. 11-bit hand rotation about the vertical: G1 0.095°, H1 0.080° (1σ). Along-wall scatter at a 0.3 m nose: G1 1.312 mm (surface) / 1.375 mm (44.4 mm deep), H1 1.064 / 1.122.
- The stack. G1 0.7346 / 0.5118 / 0.2637 / 0.0993 at L 0 / 0.15 / 0.3 / 0.45 m; H1 0.9117 / 0.7830 / 0.5706 / 0.3247. The point model alone is as steep (0.9595 at 0.6 mm, 0.3152 at 1.3). H109 held (−53.3 points).
- Resolution. G1 at 0.3 m: 12 bits 0.9344, 13 bits 0.9910, 14 bits 0.9966.
- Lean. 1° (1σ) on top costs 9.28 points (H110 held); the stack tolerates 0.434° for 2 points. Doubled joints 1.0000 at 2° (H111 held), 0.9998 at 3°; at four times the quantisation and L 0.3 they drop to 0.3952.
- What it means. The single-2x numbers in the paper need the tool point, not the hand, held to about 0.9 mm: a stock G1 with a nailer misses that by half. The tool spec, not any placement remedy, now decides the single-2x case.
- Next. Backlash, link and mount deflection under the push carried to the nose (they only add); a measured nose length; restate the paper's abstract and conclusions for a tool-point spec.
2026-11-16 · Milestone 29 planned: the nailer's length and the nail's lean (what the 0.9 mm tool scatter leaves out)
- Why. Every clean-sheet figure since milestone 1 takes the tool error as
buildbot-fasten's 0.9 mm: the G1's 11-bit joint quantisation carried to the hand point, with the nailer's length ignored (an open gap in that project's audit), and the nail treated as a point on the sheathing surface. A nailer's nose sits about 0.3 m ahead of the grip (buildbot-fasten's wall distance, not sourced), so the hand's rotation error moves the nose; and the nail must stay inside the stud for 10 diameters past the sheathing (milestone 2's band), 44 mm below the surface, so a leaning nail leaves the stud even when its head is in its window. Milestone 10 found the tool carries 45–47% of the single-2x misses, so the tool model matters more than any remedy. - Plan.
src/buildbot_sheathing/tool_lean.py: (A) the G1 and H1 hand's translation and rotation Jacobians (feet planted, MuJoCo Menagerie, buildbot-fasten's pose), carried to the nose at a tool length L (0, 0.15, 0.3, 0.45 m) and to the shank 44.4 mm (7/16 in OSB plus 10 nail diameters) into the wall, giving a correlated pair (surface, depth) per nail and axis; (B) tack.simulate takes that pair: the panel-edge bound sees the surface point and the stud or plate bounds see the depth point (the old path unchanged when it is not given); a one-nail Monte Carlo is checked against the bivariate normal; (C) the best single-2x stack (measured studs, per-nail reads, mid push, milestone 27's 20 ms slide, milestone 28's 1/16 in shim stack: 0.7964) and the doubled-joint pipeline re-run with the robot's tool error at each L and joint error 1, 2 and 4 times the 11-bit quantisation; (D) an independent nail lean (the tool's own aim, a nail deflecting in the wood) swept 0–3° (1σ), and the lean the stack tolerates for a 2-point loss (bisection). - Disclosed before registering. A probe of (A) at the G1's stud-height pose ran first: the hand rotates about 0.1° (1σ) from quantisation, and at L 0.3 m the nose's along-wall scatter is 1.31 mm against the hand's 0.93. No clean-sheet number had been computed.
- Hypotheses, registered before the run. - H109 (the nailer's length matters). With the G1's 11-bit error carried to a 0.3 m nose and the 44 mm depth, the best single-2x stack loses more than 2 points against the 0.9 mm hand-point model. Refuted if it loses 2 points or less. - H110 (a 1° lean breaks the stack). An independent nail lean of 1° (1σ, both axes) on top of the robot's error costs the best stack more than 5 points. Refuted if it costs 5 or less. - H111 (doubled joints shrug it off). The doubled-joint pipeline stays at or above 0.99 clean at a 2° lean. Refuted if below.
2026-11-16 · Milestone 28 result: a 1/16 in shim stack leaves 0.458 mm and keeps the stack at 0.796; a stock G1 cannot drive a wedge (66 N against 119 N); H106 held, H107 and H108 refuted
src/buildbot_sheathing/shim_mech.py (new), src/buildbot_sheathing/ceiling.py (a callable shim: a residual set by each sheet's own tilt; the old path unchanged, m12 reproduces), experiments/m28_shim_mech/ (run, results, figure), tests/test_shim_mech.py (6 tests), reproduce.py (m28 listed; --check m28_shim_mech reproduces, 98 lines), paper/draft.md (section 5.11 paragraph, a limitations bullet), paper/check_numbers.py (211 of 211, was 201), README, PROPOSAL and project card.
- Sources. Composite taper shims 8 in long, 1/16 to 5/16 in (slope 1:32), and plastic horseshoe shims at 1/16, 1/8 and 1/4 in, from supplier listings (no manufacturer spec sheet found). Face frictions swept 0.2–0.6.
- Lift needed: median 0.946 mm, p99 3.018 mm (rock draws); 43% of sheets need under half a 1/16 in step.
- Stack: 0.229 / 0.458 / 0.934 mm rms at 1/32, 1/16, 1/8 in. With milestone 27's 20 ms slide the best stack gives 0.8042 / 0.7964 / 0.7653, against 0.7912 with a Gaussian 0.5 mm. H106 held (+0.52 points).
- Wedge: 108–129 N at μ 0.4 (press position swept), self-locking everywhere. Stock G1, both hands into the wall: 40.6 / 65.6 / 118.0 N at z 0.35 / 0.5 / 0.7 m; it can drive faces up to μ 0.16 / 0.25 / 0.40. H1 130–135 N. H107 refuted (118.5 against 65.6 N at the model's 0.5 m ledger).
- Wedge stick-slip: a hand's effective mass along the push is 0.78–1.17 kg (MuJoCo, feet planted, joints free), so a tap runs 12.7 mm at 20 ms and 337 mm (past the 8 in wedge) at 100 ms. Loop rms 0.013 mm at 2 ms (70 taps), 0.041 at 10 ms, 0.132 at 20 ms. H108 refuted (bar 0.1 mm).
- What it means. The paper's shim is a 1/16 in stack, placed, not driven: no force or timing requirement, and as good as the 0.5 mm assumed. The wedge pays a point only for a robot that pushes about 120 N low on the wall with a 10–20 ms release.
- Next. Measure the composite-shim and OSB friction pairs; a shim-dispensing end effector and its cycle time in milestone 13's cost model.
2026-11-16 · Milestone 28 planned: the shim as a mechanism (a shim stack or a driven wedge under the low corner)
- Why. Milestones 10, 12, 13 and 27 credit the shim (lift the sheet's low bottom corner until its edge is plumb) with a 0.5 mm corner-height residual, and milestone 27 left it as the last remedy without a mechanism. Two ways to set the height exist on a site: a stack of fixed-thickness plastic shims (1/16, 1/8 and 1/4 in horseshoe shims, supplier listings) and a tapered composite shim driven under the corner (8 in long, 1/16 to 5/16 in thick, a slope of 1:32, supplier listings).
- Plan.
src/buildbot_sheathing/shim_mech.py: (A) the lift each sheet needs, from milestone 5's ledger and panel tilt draws; (B) a shim stack's residual, the lift rounded to the step; (C) the quasi-static wedge: push to drive it and pull to back it out, with the sheet pivoting on its far corner (m g / 2 on the wedge plus wall friction from the press), friction on both faces swept 0.2 to 0.6, checked for self-locking and against the stock G1's largest push into the wall at the ledger height (milestone 3b's whole-body check, MuJoCo Menagerie); (D) driving the wedge as milestone 27's stick-slip tap, with the hand's effective mass along the push from MuJoCo, from either side until the height is within tolerance; (E) the residuals put back into milestone 12's best single-2x stack with milestone 27's slide residual. - Hypotheses, registered before the run. - H106 (a 1/16 in stack is as good as assumed). With milestone 27's slide residual (20 ms, μ_k/μ_s 0.8), a 1/16 in shim stack keeps the best single-2x stack within 2 points of what it gives with a 0.5 mm shim. Refuted if it loses more than 2 points. - H107 (a stock G1 can drive the wedge). At μ_s 0.4 on every face and the press half way along the sheet, the wedge's breakaway push is below the stock G1's largest push into the wall at the ledger height. Refuted if it is above. - H108 (the taper divides the stick-slip). With the G1 hand's effective mass, a 20 ms release and μ_k/μ_s 0.8, the wedge loop's rms height residual is below 0.1 mm. Refuted if it is 0.1 mm or more.
2026-11-15 · Milestone 27 result: the slide leaves 0.19–0.64 mm if the push drops within 20 ms of the sheet moving, 2.5–10 mm at 100 ms; H103 refuted, H104 and H105 held
src/buildbot_sheathing/slide_mech.py (new), experiments/m27_slide_mech/ (run, results, figure), tests/test_slide_mech.py (5 tests), reproduce.py (m27 listed; --check m27_slide_mech reproduces, 110 lines), paper/draft.md (section 5.11 paragraph, a limitations bullet), paper/check_numbers.py (201 of 201, was 194), README, PROPOSAL and project card.
- Hand stiffness along the wall (MuJoCo 4.0.1, Menagerie G1/H1, milestone 3b's press pose, stock kp, point grip): one hand G1 35–102 N/m, H1 286–590; both hands G1 128–134, H1 596–1,394. Breakaway is about 105 N, so at 63 N/m the set-point would sit 1.7 m behind the hand: the stock hand is a force source, and what stops the sheet is letting go.
- One tap. The time-stepped model matches the force-source closed form to 0.3% or better. At μ_s 0.4, μ_k/μ_s 0.8: 0.17, 0.39, 0.81, 3.2 and 12.3 mm at 2, 10, 20, 50 and 100 ms. Across μ_s 0.3–0.7 and ratio 0.5–0.9 at 20 ms: 0.51–3.2 mm.
- H103 refuted. G1 poses stay within 1.9% of a force source; both H1 hands at z 1.3 m travel 17.4% less at 100 ms (bar 10%). The stiffer the hand the less it runs, but no stock pose is stiff enough to matter below 50 ms.
- The loop (2,000 sheets, from either edge, stop within max(0.1 mm, half the last tap), breakaway ±10% a tap): 0.054–0.090 mm rms at 2 ms but 5–10 taps; 0.19–0.64 mm at 20 ms in 1.5–3 taps; 0.66–2.9 mm at 50 ms; 2.5–10 mm at 100 ms. H104 held: 3.855 mm at 100 ms, ratio 0.8.
- H105 held. Milestone 12's best stack with the loop's rms in place of 0.5 mm: 0.7912 at 20 ms, ratio 0.8 (0.7707 with 0.5 mm, reproduced exactly). At 50 ms 0.40–0.75, at 100 ms 0.13–0.45.
- Not measured. The OSB-on-lumber friction pair (search excerpts only: static 0.25–0.7, kinetic 0.15–0.4 for timber, a review in Preprints 202404.0050; static above kinetic and more variable for OSB, Materials and Structures 51:2, 2018, full text unreachable here), the tap-to-tap variation, the release latency (instantaneous release after τ, no arm inertia). The residual is put into the tolerance model as a Gaussian at its rms, though the loop's residual is closer to uniform.
- Next. The same treatment for the shim; a friction test or data set for OSB on SPF; a G1's measured slip-to-release latency.
2026-11-15 · Milestone 27 planned: the slide as a mechanism (stick-slip on the ledger)
- Why. Milestones 10, 12 and 13 credit the slide (centre the sheet's gap along the ledger before tacking) with a 0.5 mm residual, and the paper's best single-2x stack (0.771 clean on measured studs) rests on it. Nobody has asked what leaves 0.5 mm. A seated sheet starts on static friction and slides on kinetic; a push-only hand overshoots by however far the sheet runs before the hand lets go.
- Plan.
src/buildbot_sheathing/slide_mech.py: a one-axis stick-slip model of the sheet on the ledger and wall (milestone 3b's P = mu_l m g + mu_w F_n, split into static and kinetic), pushed by a hand of the stiffness the stock G1/H1 gains give along the wall (MuJoCo Menagerie models, milestone 3b's press pose), which ramps its force, detects the slip and releases after a latency. A closed form for a force-source hand checks it. Then a correct-from-either-side loop gives the slide residual, which goes back into milestone 12's grid. Friction: static 0.3–0.7 and mu_k/mu_s 0.5–0.9 swept (no OSB-on-lumber pair measured in a source this session could open). - Hypotheses, registered before the run. - H103 (the stock hand is a force source). Along the wall the stock G1 and H1 hands are so soft that a single tap's travel is within 10% of a pure force source's at every reaction time swept. Refuted if any robot pose differs by more than 10%. - H104 (reaction time sets the slide). With mu_k/mu_s 0.8 and mu_s 0.4, the loop's rms residual exceeds 0.5 mm at a 100 ms reaction time. Refuted if it is 0.5 mm or less there. - H105 (what the stack keeps). At a 500 Hz loop with 20 ms latency, milestone 12's best single-2x stack (0.771) stays within 2 points with the slide residual from the mechanism in place of 0.5 mm, at mu_k/mu_s 0.8. Refuted if it loses more than 2 points.
2026-11-15 · Milestone 26 result: every experiment now reproduces; the last ten re-runs found one more drifted table (m6), two rounding or bookkeeping lines, and a Drake upgrade that breaks m6
reproduce.py (m6's per-seat timing column masked, wall time N s lines skipped, Drake 1.51.1 named for m6), tests/test_reproduce.py (5 tests, was 4), src/buildbot_sheathing/drake_seat.py (docstring), experiments/m6_drake and m9_straighten (re-run), experiments/m21_degrading/check_diag.txt, paper/draft.md, paper/check_numbers.py (194 of 194), README, PROPOSAL and project card.
- What was run.
reproduce.py --check on the ten experiments milestone 25 left: m3, m3b, m6, m6b, m9 (MuJoCo 4.0.1; Drake 1.58.0, then 1.51.1 for m6) and m15, m16, m19, m20, m21 (numpy 2.5.3, scipy 1.18.1). Four groups ran side by side on four CPUs, about 2 h 20 min wall clock; m20 alone took 8,397 s under that load, against 2,099 s committed. - Reproduce exactly (timing lines aside): m3 (133 lines), m3b (151), m6b (157), m15 (65), m16 (57), m19 (76), m20 (71) and m21 (52 in
results.txt). - m6 failed on the current Drake. Drake 1.58.0 rejects
discrete_contact_approximation="tamsi", which section D's solver comparison uses. It was committed under 1.51.1 (its results.json records the version); in a separate venv with drake==1.51.1 it runs. The Drake tests and m6b do not use TAMSI and pass on 1.58. - m6 drifted, for milestone 25's reason. Section F2 re-runs milestone 5's 24 ledger pipelines with each engine's seated residual. Its "no residual" column now equals m5's refreshed table line for line (0.4170, 0.5814, 0.2434, 0.9404, …), so the drift is milestone 9's reordered tilt draws, already traced in milestone 25; the MuJoCo and Drake columns moved with it. The worst |Drake − MuJoCo| over the 24 pipelines is now 0.05 points (was 0.06); H35 (bar 10) still holds by the same margin. Re-run in place on an otherwise idle machine: Drake 0.080 s and MuJoCo 0.020 s a seat (were 0.114 and 0.025). Sections A–E and G are unchanged.
- m9: one value at a rounding edge. "2x, three, nails fixed" missed nails 0.48 → 0.47.
results.json shows the committed mean was 0.47500000000000164, floating-point error from milestone 14's old (1 − mean hit) × 62 formula; the per-sheet sum gives exactly 0.475, which prints as 0.47. Same cause as m12's third-decimal drift. The paper does not quote it. - m21: a stray line. The committed
check_diag.txt ended in DONE, a wait marker a session's shell had echoed into the file; check_diag.py does not print it. Replaced with the re-run, otherwise identical. - What the check needed. m6 ends two rows with seconds a seat and m6, m6b, m9 end with
wall time N s; neither matched milestone 25's timing patterns, so both are now masked (tests added). Diffs print up to 80 lines (40 cut off m6's). - Paper. One number moved: section 5.6's 0.06 points is now 0.05.
check_numbers.py 194 of 194. - Still open before submission: authorship and affiliations; the NDS and SDPWS editions; a tested 16d end-nail ultimate. Drake 1.51.1 should be named in the paper's data-availability statement.
2026-11-15 · Milestone 26 planned: re-run the experiments milestone 25 left unchecked
- Why. Milestone 25 checked twelve experiments under
reproduce.py --check; m15, m16, m19, m20, m21 (the long scipy ones, ~150 min on four CPUs together) and the MuJoCo/Drake ones (m3, m3b, m6, m6b, m9) were not. The paper quotes numbers from all of them, so a drifted one would put a stale number in the submission. - Plan. Install
mujoco and drake in this session, run reproduce.py --check on each remaining experiment in background jobs, and for any that differ, bisect the cause as milestone 25 did, refresh the results and update paper/draft.md and paper/check_numbers.py. No new hypotheses.
2026-11-15 · Milestone 25 result: two experiments no longer reproduced their committed numbers; both drifted from later edits to shared code, now refreshed, no verdict changed
reproduce.py (new: one command per experiment; --check re-runs in a temporary copy and diffs each results.txt, wall-clock timings masked), tests/test_reproduce.py (4 tests), experiments/m5_cycle, m7_cost, m12_remedies, m13_price (results and figures re-run), paper/draft.md and paper/check_numbers.py (moved numbers updated; 194 of 194, was 190).
- Why. The paper is near submission and the programme's bar asks for a reproducible artefact.
check_numbers.py ties the paper to the committed results files, but nothing tied the results files to the code that is committed now. Shared modules (tack.py, cycle.py) have been edited by eight later milestones. - What was re-run. m1, m2, m4, m5, m7 and m10–m14 (numpy, ~6 min together) and m17 and m18 (scipy, ~15 min); m19 and m15 were started after them and stopped at the session's 30-minute limit for a background job, unchecked. The MuJoCo/Drake experiments (m3, m3b, m6, m6b, m9) and the long cyclic ones (m16, m20, m21) were not.
- m5 drifted. H22's best single-2x pipeline came back 0.5814, not 0.5833. Bisecting over the five commits that touched
tack.py: milestone 5's own commit reproduces 0.5833 exactly, and every commit from milestone 9 (1d598701) on gives 0.5814. Milestone 9 rewrote cycle.ledger_draw to add the bowed-edge option and, in doing so, drew the panel tilt before the ledger tilt from the same generator. The distributions are the same and the draws are not. Not a bug; the committed table was simply from older code. - m12 drifted in the third decimal. Milestone 14 changed the mean missed-nail count from (1 − mean hit) × 62 to a per-sheet sum, the same quantity: three values move 0.001 (0.549 to 0.548). m7 and m13 read m5's and m12's results, so they moved with them.
- What moved, all inside the intervals quoted with them. H22 0.5833 → 0.5814 [0.574, 0.588]; H23 0.9997 → 0.9994; H26 7.64 → 7.39 points (largest loss now at three reads, not dense); single-2x penalty 17.8–65.2 → 17.9–65.5 s; break-even $6.51 [5.21, 7.88] → $6.47 [5.17, 7.83]. H22–H26 keep their verdicts.
- A gap in the number check.
check_numbers.py bound "0.9997" to its results line but not the same number written as "99.97%", so the stale figure sat in the abstract, a highlight, the discussion and the conclusion without failing anything. Four checks now bind the percentage form. - After the refresh all ten quick experiments reproduce, and m17 and m18 reproduce as committed (80 and 66 lines).
- Still open before submission: authorship and affiliations; the NDS and SDPWS editions; re-running m15, m16, m19–m21 and the MuJoCo/Drake experiments (needs
pip install mujoco drake) under --check.
2026-11-15 · Milestone 24 result: the abstract fits the journal's 150 words, and the build now fails on a long abstract, a single-author citation left as text, or an unused reference
paper/draft.md (abstract rewritten, 469 to 150 words; Ghazinader 2024 cited in section 4.10), paper/build_latex.py (abstract check, wider leftover-citation pattern, unused .bib entries now fail), paper/check_numbers.py (one check re-pointed), paper/references.bib (SDPWS note), tests/test_latex.py (8 tests, was 6). No new computation; check_numbers.py still 190 of 190; build_latex.py --pdf compiles to 26 pages with 24 works cited and nothing undefined.
- The limit. A web search's excerpt of Automation in Construction's guide for authors says the abstract "must not exceed 150 words". The guide page itself (sciencedirect.com, elsevier.com) is refused by this session's proxy, so the limit is as excerpted, not read on the page; it is a constant in
build_latex.py to change if the page says otherwise. Milestone 23's note had guessed "near 250". - What the abstract keeps. The task (34 of 62 nails on a 1.5 in face), the chain (tolerance model, MuJoCo and Drake, cost), and one number per finding: 87% cap on single-2x joints with perfect sensing, 99.97% with doubled joint studs, neither humanoid carries the 21.8 kg sheet, a G1 on a lift base beats a crew in 37% of draws at 100 houses a year, an escaped miss costs about 2% of capacity, and the decayed walls fit only at end nails of a quarter of code strength. Dropped (all still in the results and conclusion): the 0.41 mm and 0.03° requirements, the ledger and Drake figures, the $6.50 break-even, the 4 in joint-edge remedy and the 63–74% range of the five models.
- What the checks caught. One number check (
beats a walking H1 in every draw) lived only in the old abstract; it now points at the same result in section 5.7 (costs less than the walking H1 in every draw), so 190 checks still bind 190 numbers. Citing Ghazinader (2024) as Ghazinader (2024) passed milestone 23's leftover check silently, because that pattern caught only "et al.", "A and B" and Weiss: a single-author citation missing from the key table would have printed as plain text. The pattern now catches any capitalised name before a year; it raised no false alarm on the draft. - The editions, still open. Search records (AWC errata for the 2021 SDPWS; STRUCTURE magazine) show section 4.3.7.1 and its adjoining-panel-edge framing rule in the 2021 edition and the two-2x alternative in every edition since 2008, so the paper's use holds from 2008 on; which edition Table 4.3A's values came from is still not confirmed (awc.org refused by the proxy).
- Still open before submission: authorship and affiliations; the NDS and SDPWS editions; a tested 16d end-nail ultimate.
2026-11-15 · Milestone 23 result: the paper builds into Elsevier's template, 27 pages with every citation resolved
paper/build_latex.py (new), tests/test_latex.py (6 tests), paper/references.bib (two entries completed, four standards added), paper/draft.md (sections 4.7–4.8 reordered, five new highlights, two citations renamed). No new computation; paper/check_numbers.py still 190 of 190.
- What it does. Turns
draft.md into paper/latex/manuscript.tex in the elsarticle class (author-year, elsarticle-harv) through pandoc, with the highlights, abstract and keywords in the front matter. Author-year citations in the draft's three forms ([A 2016; B 2017], A et al. [2016], A (2017) and bare ones inside parentheses) become natbib commands keyed to references.bib: 23 works cited, 4 standards listed with \nocite. Compiled with pdflatex and bibtex: 27 pages, no undefined citation or reference, 3 tables and 3 figures numbered by LaTeX. - What the checks caught in the draft. Section 4.8 came before 4.7, so LaTeX's numbering would have made every "section 4.7" in the text point at the wrong one; the two blocks are swapped (text unchanged). All eight highlights ran past Elsevier's 85 characters (84 to 110) and Elsevier asks for 3 to 5: now five, 74 to 84 characters, every number from the abstract or results. The script fails on any of these from now on, and on a citation left as plain text or a key missing from the .bib file.
- References completed. "Construction Robotics (2024)" had no authors: it is Cote, Tish, Koehle, Koga and Chitta, Construction Robotics 8, 8 (catalogue record and Autodesk Research listing); the text now cites Cote et al. Lachmayer et al. (2022) had no title: "Autonomous sensing and localization of a mobile robot for multi-step additive manufacturing in construction", pages 453–458, doi:10.5194/isprs-archives-XLIII-B1-2022-453-2022 (Copernicus page, TUM record). Both seen through web search; Crossref and OpenAlex were refused by this session's proxy.
- Still open before submission: authorship; the abstract is 469 words where Elsevier journals commonly cap near 250 (the journal's own guide was not reachable, so not confirmed); the NDS and SDPWS editions;
ghazinader2024 is in the .bib but not cited by the draft.
2026-11-15 · Milestone 22 result: the paper now carries the capacity half; five models of Weiss's decayed walls in one table
paper/draft.md (sections 4.9–4.10, 5.13–5.14, Table 3; abstract, highlights, discussion, limitations, conclusion and references extended); paper/check_numbers.py (60 new checks, 190 of 190 ok; its README said 70 before this step, the file had 130). No new computation.
- What the paper now says. An escaped joint-edge nail costs a mean 2.08% of its sheet (milestone 14); three reads with every miss left in lose 2.34% mean and 11.49% at p99, so misses should not escape on 6 in joint edges; 4 in joint edges carry them (0.9998 of sheets beat an intact 6 in sheet; $6.98 a sheet against $7.10 with re-drives). Joints of NDS strength and softening nails leave this standing.
- Table 3. At code-strength end nails the decayed 3 in wall keeps 63.4% (limit analysis), 73.5% (rigid rod path), 72.8% (pushover), 70.6% (CUREE) and 73.2% (energy damage at γ 10), against 33.9% tested; the spring models' fits all sit at m 0.25 (all 30 of the 960 pushovers within 5 points rms; checked from
m19/results.json). - Writing checks that changed the text. The first draft said every fit within 5 points rms had m 0.25; milestones 17 and 18 fit at 0.47 and 0.17, so the claim is now limited to the spring models. "4 in joint edges carry every escaped miss" was corrected to 0.9998 of sheets. H63's bar is stated as registered (it fails on p99 and the wall term, not the mean).
- Searched again for a tested 16d end-nail ultimate. None reachable: the nearest hit (an NAHB roof-to-wall uplift report, two 16d end nails 211 lb) is uplift, not lateral, and nahb.org and huduser.gov are refused by this session's proxy.
- Still open: the tested end-nail ultimate, a local tear-out model for the nails above the unnailed strip, the Elsevier template and authorship.
2026-11-15 · Milestone 22 planned: the paper's capacity sections (milestones 14–21)
- Why.
paper/draft.md stops at milestone 13. Milestones 14–21 answer the paper's structural question (what an escaped missed nail costs a shear wall) and check the capacity model against Weiss's (2017) decayed-wall tests through eight models. None of it is in the paper, so a reviewer would see the robot results with no capacity check behind them. - Plan. Add methods 4.9–4.10 and results 5.13–5.14 to the draft, carry the finding into the abstract, discussion, limitations and conclusion, and tie every quoted number to its results file in
paper/check_numbers.py. No new computation; every number is read from experiments/m14–m21 results files.
2026-11-15 · Milestone 21 result: nails that lose strength with dissipated energy lower every wall alike, and lift the decayed shares; the 3 in wall stays above 70% at code strength
experiments/m21_degrading/ (run.py, results.txt, results.json, check_diag.py, check_diag.txt, figure.py, figure.png); wall_cyclic.Saws(gamma=...) and Saws.degrading; tests/test_wall_degrading.py (5 tests). About 35 min on four CPUs (72 CUREE runs of 3,000–9,300 steps, 8 monotonic pushes).
- H98 refuted as registered, by a mechanism, not a bug. γ = ∞ reproduces milestone 20 exactly (0 difference on the single-spring CUREE trace, the OpenSees trace and the registered-point wall peaks). But a monotonic push at γ 5 moves the wall peaks by up to 1.6e-5 FU, not ≤ 1e-9: in a wall, 203–381 of the 218–454 springs reverse at least once after yielding during a one-way push (the cross spring of each oriented pair, and springs that unload after a neighbour snaps), so the rule does start counting their work. It does so after the peak (
check_diag.txt: s falls to 0.69–0.83 by 150 mm), so the monotonic references stand. - H99 refuted as scripted, by the check. The hand recomputation in
run.py left out the rule's "s never rises" clamp, so it disagreed by 9e-3 FU while a spring unloaded. With the clamp (check_diag.py) the force on the envelope equals s × E(d) to 4e-16 FU and s matches the hand value exactly. - H100 held. At code-strength joints (m 1, λ 1, 1/16 in gap) the decayed 3-12 wall keeps 70.6, 70.9, 71.1, 72.4 and 73.2% at γ ∞, 100, 50, 20 and 10 (73.5% at 5).
- H101 held, in the direction nobody wanted. Degradation moves each decayed share by +0.3 to +2.9 points: up, not down. At m 0.25 the 6-12 share rises 10 points (57.9% to 68.1%) by γ 5.
- H102 held. No point with m ≥ 0.75 within 10 points of both shares at any γ.
- Why the shares rise. Damage follows the energy each nail dissipates, and an intact wall dissipates more in its sheathing nails than a decayed one whose capacity is set by the frame sliding on its end nails. At m 0.25, γ 5, the intact walls' sheathing nails keep 71–77% of their strength and the decayed walls' 86–96%. At m 1 all four walls lose about alike (s 0.69–0.74), and the shares still rise 1.5–2.9 points.
- The cost to the intact walls. Their CUREE peaks fall to about 97, 94, 85, 75 and 63% of their monotonic peaks at γ 100, 50, 20, 10 and 5 (within a point at every m), and the intact ratio 3-12/6-12 stays at 1.90–1.94 (tested 1.87). Rates that would matter at all therefore also predict cyclic peaks well below monotonic for the intact walls.
- Best fits get worse with damage. The best point at every γ is m 0.25; its rms grows from 2.1 points (no degradation) to 2.5, 3.1, 4.3, 6.4 and 8.7.
- What it means for the programme. Damage spread by dissipated energy, the standard cumulative-damage term in hysteretic wall models, cannot produce the 34%: it scales intact and decayed walls together. What is left must single out the decayed wall: tear-out of the few nails left along its unnailed bottom edge, or the middle stud splitting (both reported by Weiss), or end nails that are genuinely weak. The first two need a local failure model (a nail that fails by edge distance and cycle count, from tested curves of nails at an unsupported panel edge); the third a tested 16d end-nail ultimate. For the robot nothing moves.
- Inputs not sourced: γ (swept; no calibrated value for nails in OSB was found, and the Stanford repository holding the CUREE reports did not resolve from this session); the end-nail joint's curve (as milestones 19–20). One rule for every spring.
2026-11-14 · Milestone 21 planned: Weiss's wall with nails that lose strength as they dissipate energy
Written before any milestone-21 wall load was computed (only the single-nail calibration below was run). Inputs: milestone 20's wall, nails, rods, CUREE history and Δ (experiments/m20_cyclic/results.json).
- Why. Milestone 20 found that SAWS pinching alone leaves the decayed 3 in wall at 70.6% kept under the CUREE history (tested 34%). SAWS has no term for damage that grows with cycles (nail fatigue, tear-out), which Anderson et al. (2007) and Gatto and Uang (2003) saw. This step adds the standard energy-based strength degradation and asks how much of it the 34% would need, and whether that much is compatible with the intact walls.
- Model.
wall_cyclic.Saws(gamma=...): every spring's envelope and pinching lines are scaled by s = 1 − E/E_t, E the work done on the spring from its first reversal after yielding (a monotonic push is never degraded), E_t = γ FU Δu of that spring. This is the c = 1 continuous form of Rahnama and Krawinkler (1993)'s rule as used in the Ibarra, Medina and Krawinkler (2005) model. Unloading and reloading stiffness are not degraded. γ is not sourced (no calibrated value for nails in OSB was found this session): it is swept over ∞, 100, 50, 20, 10, 5. Calibration on one Anderson nail through the CUREE history at Δ = 2Δu (work 13.8 FU Δu in all): its envelope peak falls to 0.96, 0.91, 0.81, 0.73 and 0.61 of the undegraded peak at γ 100, 50, 20, 10, 5. - H98 (code check: reduces to milestone 20). γ = ∞ reproduces milestone 20's
Saws exactly (0 difference) on the single-spring CUREE trace and milestone 20's registered-point wall peaks to 1e-9; at any γ a monotonic push is unchanged. - H99 (code check: the rule). On a single spring, whenever it is on its envelope, its force equals s × the undegraded envelope with s recomputed from its own force–displacement history, to 1e-9 FU.
- H100 (moderate damage does not take the 3 in wall to 34%). At the registered point (m 1, λ 1, gap 1/16 in, cap 2.55 FU), the decayed 3-12 wall keeps at least 50% for every γ ≥ 10.
- H101 (damage scales intact and decayed walls alike). At the registered point, for every γ in the grid, each decayed share moves by at most 10 points from its γ = ∞ value.
- H102 (no fit near code strength). Over m ∈ {0.25, 0.47, 1.0}, λ 1, gap 1/16 in, and the γ grid, no point with m ≥ 0.75 puts both decayed shares within 10 points of the tested 56.5% and 33.9%.
- Also reported, not scored: the intact walls' cyclic peaks against their monotonic peaks at each γ (an intact wall far below its monotonic peak is a sign γ is too low for Weiss's materials), the intact ratio 3-12/6-12 (tested 1.87), the best fit at each γ.
- Limits known in advance. γ is swept, not sourced; one rule for every spring (sheathing nails and end-nail joints alike); damage is spread over the spring, not tied to a failure mode (tear-out at the unsupported bottom edge would hit the bottom nails first); Anderson's nail envelope was itself fitted to cyclic tests, so part of a nail's cyclic damage is already in it and any γ here is extra damage.
2026-11-14 · Milestone 20 result: under the CUREE protocol with pinching nails Weiss's wall keeps what it keeps when pushed once; cyclic loading does not explain the 3 in wall
experiments/m20_cyclic/ (run.py, ops_check.py, results.txt, results.json, loops.json, figure.py, figure.png); src/buildbot_sheathing/wall_cyclic.py; tests/test_wall_cyclic.py (7 tests, one against an OpenSees trace in tests/data/). About 70 min on four CPUs (112 monotonic pushes, 8 radial, 120 CUREE runs of 2,500–13,300 steps).
- H93 held, exactly.
wall_cyclic.Saws reproduces OpenSees' SAWSMaterial on the CUREE history to 2e-13 FU rms (largest 6e-13) for Anderson's four rows and Folz and Filiatrault's set, at two step sizes. Getting there took two rule fixes that the registered description left implicit: a spring on the envelope stays on it while it keeps moving away from zero, and after unloading from the envelope it reloads on the same r3 K0 line until its force reaches that side's pinching line. - H94 refuted, by a known bias. Pushed once, the oriented pair carries 3.1–5.2% more than milestone 19's radial spring with the same envelope (6-12 intact 18.84 against 17.95 FU), as Judd and Fonseca report for uncoupled pairs. The shares kept agree to 0.2 point (78.6%|70.2% against 78.4%|70.1%), so the comparison the step is about, cyclic against monotonic in the same model, is unaffected. The pair also holds its load longer after the peak (Δm 57–75 mm against 61–82 mm in mixed directions).
- H95 held. Intact walls under the CUREE history (Δ 37.2 and 39.9 mm) peak at −0.1% and −0.0% of their monotonic peaks.
- H96 held. At code-strength joints, the rods at their design value and a 1/16 in gap, the decayed 3-12 wall keeps 70.6% cyclic against 70.5% monotonic (+0.1 point); the decayed 6-12 wall 77.8% against 78.3%. Tested: 34% and 57%.
- H97 held. No grid point with m ≥ 0.75 is within 10 points of both tested shares under cyclic loading; the closest (m 0.75, λ 0.25, 1/4 in gap) keeps 76.3% and 43.3% (rms 15.5 points).
- Across the grid cyclic loading moves a decayed share by a median 0.2 point and at most 3.0 points (m 0.25, λ 0.25, 1/4 in gap). The best cyclic fit is m 0.25, λ 1, no gap: 57.6% and 31.4% (rms 1.9 points), decay ratio 1.04 (tested 1.12), the end nails at a quarter of NDS again. Post hoc, with this envelope one point at m 0.47 also fits (λ 0.25, 1/4 in gap: 55.5% and 30.0%, rms 2.9), where milestone 19's nail gave 63.4% and 38.3%; that comes from the nail's envelope, not the cycles (the monotonic push there gives 54.8% and 30.8%).
- Sensitivity (not scored). Anderson's other three rows move the registered point's shares by under a point; Folz and Filiatrault's set (milder pinching, r3 1.4) keeps 74.3% and 66.2% cyclic against 75.4% and 66.9% monotonic; Δ from the radial model's push and 1 mm steps change nothing to 0.1 point. Largest residual over every run 1e-7 FU.
- Where the cycles do act: the rods. In the decayed 3-12 wall the anchor rods reach their cap and the holes crush, to 25–28 mm on one side by the end of the history (intact 3-12: 0.4 mm; 6-12: none). The extra slack comes after the peak, so it does not lower it.
- What it means for the programme. With the hysteresis model Weiss's own MCASHEW2 belongs to, cyclic loading does not take the 3 in decayed wall from 70% to 34%: its envelope peak is the monotonic peak. Two explanations are left. The end nails are weak (a quarter of NDS fits every model since milestone 18), or the walls lose capacity through damage this class of model has no term for: nail fatigue and tear-out along the unsupported bottom edge, which Anderson et al. saw grow with cycles, and the middle stud splitting. A tested 16d end-nail ultimate decides the first; a degrading envelope (cumulative-damage SAWS variants, or a tested decayed-wall nail curve) tests the second. For the robot nothing moves.
- Inputs not sourced: the end-nail joint's curve (shape borrowed from the sheathing nail, strength m and slip scale λ swept); Anderson's nail is 0.113 in against Weiss's 0.131 in 8d common (only shares are compared); Δ from this model's own push.
2026-11-14 · Milestone 20 planned: Weiss's wall under the CUREE protocol, with pinching nails
Written before any milestone-20 load was computed. Inputs: milestone 19's wall (wall_push.PushWall bodies: 8 by 8 ft, two sheets, seven studs, hold-downs on the double end studs, the anchor rods), milestone 17's NDS end-nail ratio (1.297 F) and milestone 18's rod path (2.55 F).
- Why. Milestone 19 left two explanations for Weiss's 3 in decayed wall (34% kept): end nails at about a quarter of their NDS ratio, or what a monotonic push leaves out. Weiss tested cyclically (ASTM E2126, CUREE). This step loads the same wall with the CUREE basic history and hysteretic fasteners, so the second explanation is tested directly.
- New sourced inputs. (1) The CUREE basic loading history, exactly as Krawinkler et al. (2001, CUREE W-02, Sec. 1.1) define it: six initiation cycles at 0.05Δ; primaries at 0.075, 0.1, 0.2, 0.3, 0.4, 0.7 and 1.0Δ followed by 6, 6, 3, 3, 2, 2 and 2 trailing cycles at 75% of the primary; then primaries rising by 0.5Δ, each with two trailing cycles; Δ = 0.6 Δm, with Δm the monotonic drift where the load first falls below 0.8 of its peak. (2) The ten-parameter SAWS hysteresis (Folz and Filiatrault 2001: Foschi envelope, unloading at r3 K0, pinching lines through ±FI at r4 K0, reloading at K0 (F0/(K0 δmax))^α towards β times the last excursion) with the parameters Anderson et al. (2007, Wood and Fiber Science 39(4), Table 4) fitted to cyclic single-nail tests of 0.113 in nails through 7/16 in OSB into Douglas fir-larch, Weiss's sheathing and framing species: at fyb 115 ksi K0 2,851 lbf/in, r1 0.037, r2 −0.069, r3 3.91, r4 0.028, F0 249 lbf, FI 32.4 lbf, Δu 0.425 in, α 0.265, β 1.15. Folz and Filiatrault's own set (8d in 9.5 mm OSB, as tabulated by Ghazinader 2024, Table 2) is the alternate.
- Model (
src/buildbot_sheathing/wall_cyclic.py). Milestone 19's bodies and penalties. Each sheathing nail is an oriented spring pair (Judd and Fonseca 2005): two uncoupled SAWS springs along and across the direction the nail first slips. Each stud-to-plate joint laterally is a SAWS spring with forces scaled by m × 1.297 per end nail and displacements by λ (shape borrowed from the sheathing nail, as milestone 19). Each anchor rod bears after a gap g in a hole that crushes: past its cap c the hole lengthens, so later cycles have more slack. Displacement control of the top plate (no rotation), Newton on the out-of-balance force at each 0.5 mm step from the committed fastener states. Peak = the mean of the largest positive and negative loads over the history (the envelope's peaks). Loads in units of FU, the nail's envelope peak. - H93 (code check: the hysteresis). On a single spring driven through the CUREE history (Δ = 2 Δu),
wall_cyclic's SAWS matches OpenSees' SAWSMaterial (the port of Folz's code) to an rms difference of at most 2% of FU over the whole history, for both parameter sets. Refuted otherwise (a bug, or a documented difference in a rule). - H94 (code check: the pair against the radial spring). Pushed monotonically, the hysteretic wall's peak is within 3% of milestone 19's path-independent
PushWall with the same Foschi envelope as a radial spring, for the four walls at m 1, λ 1, cap 2.55 F, no gap. - H95 (cyclic loading barely moves the intact walls). Under the CUREE history the intact 6-12 and 3-12 walls' peaks are within 10% of their monotonic peaks at m 1, λ 1, cap 2.55 F, gap 1/16 in.
- H96 (cyclic loading does not explain the 3 in wall). At the same point, the decayed 3-12 wall's share kept under the CUREE history is within 10 points of its monotonic share (so cyclic loading does not bring 70% down to 34%).
- H97 (no fit near code strength, cyclic). Over m ∈ {0.25, 0.47, 0.75, 1.0}, λ ∈ {0.25, 1}, g ∈ {0, 1/16, 1/4} in (cap 2.55 F), no point with m ≥ 0.75 puts both cyclic decayed shares within 10 points of the tested 56.5% and 33.9%. Expected held; refuted would mean cyclic loading reopens code-strength joints as an explanation.
- Also reported, not scored: the cyclic and monotonic shares and decay ratios over the grid; the best cyclic fit; the step size (0.5 against 1 mm) at one point; the rod hole's growth; Δ for each schedule; Anderson's other fyb rows and Folz and Filiatrault's set at the registered point.
- Limits known in advance. The SAWS envelope does not degrade with cycles beyond what β and the pinching give (no low-cycle fatigue or tear-out, which Weiss and Anderson saw); Anderson's nail is 0.113 in, Weiss's 8d common 0.131 in, so only shares are compared; rigid studs and sheets; the end-nail joint's shape is borrowed; Δ comes from this model's own monotonic push, not from Weiss's protocol records.
2026-11-14 · Milestone 19 result: slack at the rods and early-peaking joints do not explain Weiss's 3 in wall; every fit needs end nails at a quarter of NDS
experiments/m19_wall_pushover/ (run.py, results.txt, results.json, figure.py, figure.png); src/buildbot_sheathing/wall_push.py; tests/test_wall_push.py (6 tests). About 10 min on four CPUs (960 pushovers).
- H87 held. Elastic-perfectly-plastic springs (0.1 mm yield) reach
Wall.capacity within 0.044% in all 16 cases. - H88 refuted, as a modelling convention, not a bug. On milestone 15's studs the pushover's peak is 1.24% (intact) and 1.32% (Weiss pattern) above milestone 16's. A rigid stud pinned on its centre line moves the edge nails, which sit 3/8 in off that line, vertically by θ × 3/8 in; milestone 16's parallelogram moves them only along the wall. With the pins on the nail lines the two agree to 1e-7 (now a test). The LP of milestones 15–18 already used the rigid-body convention.
- H89 held. At code-strength joints (m 1, λ 1), the rod at its design value and no gap, the pushover keeps 79.5% and 72.8%; the LP keeps 79.5% and 73.5%. Softening does not change the wall, as milestone 16 found for one sheet.
- H90 held. At m 0.47, a 1/16 in gap lowers the decayed shares by 0.1 point. Across the grid the largest effect of any gap up to 1/4 in on the 3 in wall is about 9 points (m 1, λ 0.25: 71.3% to 62.7%).
- H91 refuted. Joints that peak at a quarter of the nails' slip leave the decayed 3 in wall at 71.3% (6 in: 79.6%) at code strength.
- H92 refuted, as expected. No point with m ≥ 0.75 puts both shares within 10 points; the closest (m 0.75, λ 0.25, 1/4 in gap) keeps 78.0% and 52.0%.
- Post hoc (not registered). All 30 grid points within 5 points rms have m = 0.25; the best keeps 57.8% and 32.5% (tested 56.5%, 33.9%), decay ratio 1.05 (tested 1.12), intact ratio 1.87 (tested 1.87). That is milestone 18's rigid-plastic trade-off (0.25 of the NDS ratio at the rod's design value) found again by a different model. At that point 79% of the decayed 6-12 wall's stored energy at peak is in sheathing nails and 21% in end-nail joints; 56% of the sheathing nails' share is on the double end studs (Weiss's failure line), 26% on the middle stud.
- Numerics. Every peak settled to 1e-7 Fp; after the peak some walls snap and Newton cannot settle every step (largest residual 3.6e2 Fp). Median share of nails unloading before the peak 0%, 90th percentile 5.2%, largest 100% (a few runs peak after a snap; the path-independent reading counts them on their backbone).
- What it means for the programme. Slack at the rods and the order in which fasteners peak are ruled out under monotonic loading. Two explanations are left: the 16d end-nailed joints really carry about a quarter of their NDS ratio at ultimate in these walls, or the 34% comes from what a monotonic model leaves out (cyclic tear-out at the unsupported bottom edge, the middle stud splitting). A tested end-nail ultimate decides the first; a cyclic (CUREE) push with pinching, which
wall_push can now host, tests the second. For the robot nothing moves: milestone 17 and 18's escaped-miss figures stand. - Inputs not sourced: the end-nail joint's backbone (strength m and slip scale λ swept, shape borrowed from the 8d nail); δp (swept 6–12 mm, as milestone 16); the rod's stiffness is the NDS design modulus.
2026-11-14 · Milestone 19 planned: Weiss's wall pushed over with softening nails and slack at the anchor rods
Written before any milestone-19 load was computed. Inputs: milestone 17's wall (wall_lp.Wall geometry: 8 by 8 ft, two sheets, seven studs, hold-downs on the double end studs), milestone 16's sourced nail backbone (8d through 7/16 in OSB, Shameli Derakshan et al. 2019: Fp 0.972 kN, K at 0.4 Fp 0.496 kN/mm, Δu 28.7 mm; δp not tabulated, swept), milestone 17's NDS end-nail ratio (1.297 F per 16d end nail) and milestone 18's rod path (design value 2.55 F).
- Why. Milestones 17 and 18 are rigid-plastic: every fastener reaches its capacity at once and nothing has slack. Milestone 18 left two things that model cannot show: slack in the anchor rod's hole, which lets the end studs slide before the rod bears, and joints that peak at a different slip from the sheathing nails, so one set is past its peak before the other arrives. This step builds the same wall as a displacement model and pushes it over.
- Model (
src/buildbot_sheathing/wall_push.py). The bodies of wall_lp.Wall (top plate, studs, sheets; small rigid motions). Sheathing nails: milestone 16's backbone along each nail's slip. Stud-to-plate joints: a 1-D backbone along the wall with peak m × 1.297 Fp per end nail and slip scale λ (δp_joint = λ δp, Δu_joint = λ Δu, same shape); vertically, stiff one-sided bearing, free opening; at the held-down end-stud bottoms a stiff two-sided tie. Anchor rod: 1-D along the wall at each held-down bottom, zero force over a gap g, then the NDS load/slip modulus for a dowel in wood with a steel side member (2 × 270,000 D^1.5 lb/in, D 5/8 in: about 47 kN/mm) up to a cap c Fp, elastic-plastic. Top plate pushed along the wall without rotation to 150 mm in 0.5 mm steps; at each step the bodies sit at the least-energy state (damped Newton from the last step), as milestone 16. Loads in units of Fp. - H87 (code check: the rigid-plastic limit). With every spring elastic-perfectly-plastic at a 0.1 mm yield slip, no gap, and stiff penalties, the pushover's load at 150 mm is within 2% of
Wall.capacity (same joint strength and hd_shear) for the intact and decayed 6-12 and 3-12 walls at m 0.47 and 1.0, rod cap 0 and 2.55. Refuted otherwise (a bug). - H88 (code check: milestone 16). One sheet on milestone 15's four studs, joints pinned (stiff penalties), hold-downs at both end studs, milestone 16's base backbone (δp 12 mm): the peak is within 1% of
softnail.pushover for the intact sheet and the Weiss pattern. - H89 (softening alone changes little). At δp 12 mm, λ = 1, m = 1 (NDS), rod cap 2.55 F and no gap, both decayed shares are within 5 points of the rigid-plastic LP's at the same joint strength and rod path (milestone 16's finding for one sheet, carried to the wall).
- H90 (slack in a standard hole is not enough). At m = 0.47, λ = 1, cap 2.55 F, opening the gap from 0 to 1/16 in (a standard oversize bolt hole) lowers each decayed share by at most 5 points.
- H91 (stiff joints that peak early). At m = 1, cap 2.55 F, no gap, joints that peak at a quarter of the sheathing nails' slip (λ = 0.25) bring the decayed 3-12 wall below 50% kept.
- H92 (a fit near code strength). Over m ∈ {0.25, 0.47, 0.75, 1.0}, λ ∈ {0.25, 0.5, 1}, g ∈ {0, 1/32, 1/16, 1/8, 1/4} in, cap ∈ {2.55, 5.1} F and δp ∈ {6, 12} mm, some point with m ≥ 0.75 puts both decayed shares within 10 points of the tested 56.5% and 33.9%. Expected refuted.
- Also reported, not scored: the full grid's shares and decay ratio, the best-fitting point, drift at peak against Weiss's reported drifts where the thesis tabulates them, and at the best point the share of joint and nail work by member.
- Limits known in advance. Monotonic push against cyclic tests; nail and joint unloading read back down the backbone (path-independent, counted as in milestone 16); rigid studs and sheets; the end-nail joint backbone's shape is borrowed from the sheathing nail (no tested 16d end-nail curve found yet), so λ and m are swept, not sourced; the rod's stiffness is a design modulus.
2026-11-14 · Milestone 18 result: the hold-downs' anchor rods stop Weiss's decayed wall sliding, and then no joint strength fits her 3 in wall
experiments/m18_holddown_shear/ (run.py, results.txt, results.json, figure.py, figure.png); wall_lp.Wall.capacity(hd_shear=...); 3 new tests in tests/test_wall_lp.py. About 4.5 min on four CPUs.
- Input this step added: the shear path through each HTT4's 5/8 in anchor rod. Its NDS yield-limit design value, bearing in the 1.5 in DF-L sill with a steel side member: 617 lb (1/8 in steel), 729 lb (1/4 in), 1,137 lb (1/2 in), all mode II. At the 242 lb a sheathing nail that the controls imply, 617 lb is 2.55 F.
- H81 held. s = 0 gives milestone 17's four registered capacities exactly; capacity never falls as s rises; s = inf matches s = 10,000 F to 7e-16.
- H82 refuted, as expected. A rigid shear path at NDS joints: the decayed walls keep 79.5% and 73.5% (milestone 17: 79.5%, 63.4%; tested 57%, 34%).
- H83 and H84 refuted. With a rigid path, no end-nail strength from 0.05 to 1.50 of the NDS ratio fits. The best, 0.17, keeps 68.9% and 63.6% (rms 22.8 points), and the decayed 3-12 wall comes out at 1.77 times the decayed 6-12 (tested 1.12). With the end studs held, the decayed 3-12 wall has to shear its dense nails along the end studs, so the schedule matters again, far more than in the tests.
- H85 held, at a fit that misses. At that fit, 69% of the decayed 6-12 wall's sheathing-nail work is on the double end studs (85% for 3-12), which is the failure line Weiss reported. Milestone 17's sliding fit put none there.
- H86 held. At milestone 17's fit (0.47 of the NDS ratio) the end studs stop sliding once each rod carries 1.75 F (about 420 lb). That is below the rod path's own design value, and a tested ultimate would be higher still.
- Post hoc (not registered): a trade-off. The decayed walls' capacity is close to the frame's sliding resistance: 18 end nails plus two rods. Fits within 5 points rms exist for s up to 4 F (about 970 lb a rod), and the end-nail multiple they need falls as s rises: 0.45 at s = 0, 0.25 at the rod's design value (2.55 F), 0.05 at s = 4 F. At 0.25, each 16d end nail carries about 78 lb at the tests' scale, below its own NDS design value (94 lb with C_eg). So milestone 17's explanation, that the decayed wall slides on its sill nails, needs the anchor rods to carry less than their design value and the end nails less than theirs. Code-level fastener strengths do not support both.
- Escaped misses unchanged. On the first 2,000 walls: mean loss 2.74% and 99th percentile 10.77% in milestone 17's model, and 2.13% and 9.02% with a rigid path at the 0.17 fit. Under every anchorage model tried, an escaped miss costs 2–3% of a wall on average and about 9–11% at the 99th percentile.
- What it means for the programme. A rigid-plastic, small-displacement model of Weiss's wall built from sourced strengths does not reproduce her 3 in decayed wall (34% kept). Once the rods' shear is in, it keeps 64–74%. The candidates left are ones this model cannot represent: slack in the anchor rod's hole, which lets the wall slide before the rod bears and is worth more under cyclic loading; the middle stud splitting, which Weiss reports; and sheathing nails tearing out of the unsupported bottom edge under reversed loads. Next: a displacement-based (pushover) wall with a slip gap at the rods, using milestone 16's nail springs. A tested ultimate for 16d end-nailed joints is still wanted.
- Inputs not sourced: the steel side member's thickness (swept 1/8–1/2 in); the rod path is a design equation, not a tested ultimate (only its design value is used, and a tested ultimate would only strengthen H86); the rigid-plastic path engages with no slack.
2026-11-14 · Milestone 18 planned: does the hold-down's shear path stop Weiss's decayed wall sliding, and what joint strength fits then?
Written before any milestone-18 capacity was computed. Inputs: milestone 17's wall (wall_lp.Wall, Weiss's 8 by 8 ft wall), its NDS joint ratio (1.297 F per 16d end nail) and its post hoc fit (0.47 times that ratio).
- Why. Milestone 17's fit has the decayed wall slide on its sill: every stud's end nails shear and nothing else moves. But in Weiss's set-up (Ch. 4.3.3–4.3.4) each double end stud is nailed to an HTT4 hold-down whose 5/8 in anchor rod passes through the sill into the steel test beam, so the end studs cannot slide along the sill without bending that rod and crushing the sill around it. Milestone 17 modelled the hold-downs as vertical ties only and left this "real load path" out, and its fit matched Weiss's spacing finding but not her reported failure (sheathing nails pulling out along the double end studs). This step adds it.
- Model change.
Wall.capacity(..., hd_shear=s): at each held-down stud bottom the lateral joint carries s F on top of its end nails (rigid-plastic; s=inf: no slide at all). s=0 is milestone 17's model. - Sourced scale for s. The NDS yield-limit (dowel) equations for a 5/8 in rod (F_yb 45,000 psi) bearing parallel to grain in a 1.5 in DF-L sill (G 0.50) with a steel side member (the beam flange or the HTT4 seat), single shear, R_d for D > 1/4 in. A design value, so the tested ultimate is larger by an unknown factor: reported as 1 to 5 times Z, swept rather than picked, and converted to F with milestone 17's implied 238–246 lb per sheathing nail.
- H81 (code check).
hd_shear=0 reproduces milestone 17's registered capacities (intact and decayed, both schedules, NDS joints) to 1e-9 relative; capacity never falls as s rises; s=inf gives the same capacity as s set above the frame's total nail capacity. Refuted otherwise (a bug). - H82 (the shear path alone). At NDS-ratio joints with a rigid shear path, the decayed walls keep 47–67% (6-12) and 24–44% (3-12). Expected refuted: adding a load path cannot lower a decayed capacity, and milestone 17 already kept 79.5% and 63.4%.
- H83 (one joint strength still fits). With a rigid shear path, some single multiple of the NDS joint ratio in 0.05–1.50 fits both tested shares (56.5%, 33.9%) to an rms of at most 5 points.
- H84 (the schedule matters again). At H83's best fit, decay 3-12 over decay 6-12 lies in 1.02–1.25 (tested 1.12; milestone 17's fit gave 1.00).
- H85 (where it fails). At H83's best fit, the decayed 6-12 wall's mechanism spends at least half its sheathing-nail dissipation in nails on the two double end studs (Weiss's reported failure line).
- H86 (the real rod locks the slide). The shear capacity at which, at 0.47 times the NDS ratio, the decayed 6-12 wall stops sliding as a whole (the end studs' slide falls below half the top plate's) is at most the NDS design value Z of the anchor-rod path, in F.
- Also reported, not scored: the share kept and decay ratio over s (0, 1, 2, 4, 8, 16 F and rigid) times the joint multiple (0.05–1.50); the mechanism at each fit; the tested per-wall spread (Table 5-1) beside the fit; escaped misses (milestone 17's H80) rerun at the new fit on a 2,000-wall subsample.
- Limits known in advance. As milestone 17: rigid-plastic, monotonic, rigid studs, cyclic tests of two or three walls a group. The rod's shear path is a dowel design equation, not a tested ultimate; the steel side member's thickness is not given in the thesis text we read, so the steel's bearing is taken as not governing (checked by sweeping its thickness).
2026-11-14 · Milestone 17 result: Weiss's own wall at code-strength joints still keeps 80%; with end nails at about half the NDS ratio the decayed wall slides on its sill and matches both tested groups
experiments/m17_weiss_wall/ (run.py, results.txt, results.json, figure.py, figure.png); src/buildbot_sheathing/wall_lp.py (milestone 15's limit analysis for a wall of several sheets on shared studs); tests/test_wall_lp.py (5 tests). About 5 min on four CPUs (about 9,300 LPs).
- Input this step added: Weiss's thesis full text (open access at WSU), so the wall is built as tested: 8 by 8 ft, two sheets on a single middle stud, hold-downs only on the double end studs, every nail below 12 in omitted; Table 5-2's means (control 6-12 4,688 lb, decay 2,649 lb, 57% kept; control 3-12 8,775 lb, decay 2,974 lb, 34% kept).
- H75 held. One sheet in the wall LP matches milestone 15's
Frame to 0.0 in 8 cases. One deviation, noted in results.txt: Frame has no rotation-free top plate, so the check used its free and slide-only modes. - H76 held. Intact 3-12 over 6-12: 1.931 against the tested 1.872 (+3.2%). Absolute, at the tested 8d-in-OSB peak (0.972 kN, SPF framing): 4,168 lb against 4,688 lb (0.89) and 8,049 against 8,775 lb (0.92). The tests imply 238–246 lb a sheathing nail in DF-L.
- H77, H78, H79 refuted. At NDS-ratio joints (1.297 sheathing nails per 16d end nail, 7/16 in OSB) the decayed 6-12 wall keeps 79.5% and the 3-12 wall 63.4% (tested 57% and 34%), and decay 3-12 over 6-12 is 1.54 (tested 1.12). So Weiss's anchorage, with the shared stud held down by two end nails alone, does not explain her walls at code-strength joints: the lone sheet with its own hold-downs kept 78.9%. No hold-down variant gets there either: with none the intact walls fall to 13.6 and 28.3 F (71% and 77% of the registered values) and the decayed wall keeps 50–51%. Finite hold-downs, withdrawal, sheet-to-sheet bearing, the top-plate boundary and the inclusive or perimeter-only reading of the decay pattern all keep 60–80%.
- H80 held. With every miss left in (three reads, measured studs, two sheets a wall, 10,000 walls) the wall's mean loss is 2.73% and its 99th-percentile loss 10.41%, against milestone 14's 2.34% and 11.49% a sheet. At the fitted joints (below): 2.67% and 10.19%. The escaped-miss figures the programme reports do not depend on the anchorage or the joint strength.
- Post hoc (section F, not registered): one joint strength fits both tested groups. Below 0.6 of the NDS ratio, the LP's least-work mechanism for the decayed wall is the whole wall sliding on its sill: every stud's end nails shear and nothing else moves (section E), so the capacity is 18 end nails' worth whatever the sheathing schedule. That reproduces Weiss's finding that nail spacing "does not have a significant influence" once the bottom nails are gone. A least-squares fit to the two tested shares gives 0.47 times the NDS ratio: 58.2% and 30.2% kept (tested 56.5% and 33.9%; rms 2.9 points), with a decay ratio of 1.00 against the tested 1.12. The intact walls barely notice (18.9 against 19.07 F). Read the other way, the decayed walls carried 147–165 lb per 16d end nail, about the nail's NDS design value without the end-grain factor, while the sheathing nails reached about 3.3 times theirs. That is one fitted parameter against two tested numbers plus the spacing trend, not a validation. It also leaves out a real load path, shear through the hold-down and the anchor bolts, so the joints may be weaker still. And it fits Weiss's report that the middle stud slid off its sill nails, not her report that the failure ran along the double studs.
- What it means for the programme. The candidate milestone 15 left as "joints or fasteners far weaker at ultimate than the NDS ratio implies" is now the one that fits a tested wall, and for a specific joint: the stud-to-sill end nails. For the robot, the escaped-miss losses stand under every joint model tried (H80). The weak link that decides how much a wall with a dead bottom foot keeps is a framing joint the framer nails, not the sheathing schedule. Next: a tested ultimate for 16d end-nailed stud-to-plate joints, to check the 0.47 independently, and the hold-down's and anchor bolts' shear path.
- Inputs not sourced: joint ratios are NDS design (yield-limit) values; the 0.47 is fitted. Rigid-plastic, monotonic, rigid studs (Weiss's middle studs split), against cyclic tests of two or three walls a group.
2026-11-14 · Milestone 17 planned: does Weiss's own wall, with its anchorage, explain the decay loss?
Written before any milestone-17 capacity was computed. New input: Weiss's thesis full text (B. D. Weiss, MS thesis, Washington State University, 2017; open access on the WSU repository, read this session; earlier runs had only the abstract).
- What the thesis says, quoted or tabulated there. Walls 8 by 8 ft, two vertical 4 by 8 ft 7/16 in OSB panels meeting on a single middle stud; Douglas Fir-Larch No. 2, studs 16 in on centre, double end studs, a single sill plate and a double top plate (Ch. 4.3.2). Studs end-nailed with two 16d common nails, four at each double-stud corner; 8d common sheathing nails at 3/8 in edge distance; Simpson HTT4 hold-downs on the double end studs only; the sill bolted to a steel beam (Ch. 4.3.3–4.3.4). Cyclic tests (ASTM E2126, CUREE). Decay simulated by omitting "the bottom perimeter nails and any nails below 12 inches" (Ch. 5.2.3). Peak loads (Table 5-2, group means): Control 6-12 4,688 lb, Decay 6-12 2,649 lb (57% kept), Control 3-12 8,775 lb, Decay 3-12 2,974 lb (34% kept); MCASHEW2 (with hold-downs and anchor bolts as stiff springs) predicted 53% and 46% (Table 5-3). The decayed walls failed by lateral pull-out of the sheathing nails along the double end studs, the panels "acted together as one", and the middle stud split or slid off its two end nails (Ch. 5.2.3). Nail schedule barely mattered once the bottom nails were gone (3-12 decay only 1.12 times 6-12 decay, against 1.87 for the controls).
- Why this changes the plan. Milestones 14–16 modelled one 4 by 8 sheet with a hold-down at each of its end studs. Weiss's wall has hold-downs only at the wall's ends, so the joint stud the two panels share is held to the sill by two end nails, and with the bottom foot unnailed nothing else ties the panels' joint down. Milestone 15 found that without hold-downs the lone sheet lost 64–100%. This is the anchorage candidate milestone 16 left, built as tested instead of guessed.
- Model (
wall_lp.py, milestone 15's rigid-body limit analysis generalised to a wall). Bodies: the sill (fixed, bolted), the double top plate (one body), seven studs (double end studs as one body each, centre lines 1.5 and 94.5 in; single studs at 16, 32, 48, 64 and 80 in) and two sheets. Sheathing nails: milestone 14's layout per sheet (6 in edges: 62 nails; 3 in edges: layout_spacing(3)), the right sheet shifted 48 in, each on the nearest stud or a plate, rigid-plastic F along the slip (64-gon). Stud-to-plate joints: lateral ρ F per 16d end nail (two per single stud, four per double stud) from the NDS yield-limit equations with 7/16 in OSB and G 0.50, end-grain factor 0.67; bearing in compression, opening against ω = 0. HTT4 hold-downs: rigid vertical ties at the double end studs' bottoms. Top plate: translation only along the wall and up, no rotation (MCASHEW2's boundary condition in Weiss Ch. 3.2). No bearing between the two sheets (1/8 in gap). Decay: every nail with z < 12 in omitted. Capacity per unit top displacement, in units of F. - H75 (code check). One sheet, milestone 15's four studs and hold-downs, rotation-free top plate: the new LP matches
frame_lp.Frame.capacity within 1e-6 (relative) for the intact sheet and the Weiss pattern, at rigid and at NDS joints. Refuted otherwise (a bug, not a finding). - H76 (controls). Intact wall at the registered model: 3-12 over 6-12 within 10% of Weiss's tested 1.87.
- H77 (Decay 6-12). The decayed 6-12 wall keeps 47–67% of the intact 6-12 wall (tested 57%, ±10 points).
- H78 (Decay 3-12). The decayed 3-12 wall keeps 24–44% of the intact 3-12 wall (tested 34%).
- H79 (the frame caps a decayed wall). Decay 3-12 over Decay 6-12 at most 1.25 (tested 1.12; the intact ratio is about 1.9).
- H80 (escaped misses on Weiss's wall). Milestone 14's three-read miss patterns on measured studs (20,000 sheets, seed 7), two sheets drawn independently per wall, every miss escaping: the wall's mean loss within 1 point and its 99th-percentile loss within 2 points of milestone 14's per-sheet figures (2.34%, 11.5%). Refuted otherwise.
- Also reported, not scored: the mechanism the LP picks; hold-downs removed and of finite capacity; ω and ρ swept; bearing between sheets; a free and a slide-only top plate; the edge-only reading of the decay pattern (only perimeter nails below 12 in); an absolute capacity at a sourced nail strength beside the tested 4,688 lb (SPF backbone against DF-L framing, not like-for-like).
- Limits known in advance. Rigid-plastic, small-displacement, monotonic; Weiss's tests were cyclic. Studs are rigid bodies (no bending, no splitting, which her middle studs did). Design-level (NDS) joint ratios, not tested ultimates. Two or three walls per group in the tests.
2026-11-14 · Milestone 16 result: softening nails do not explain Weiss's walls either; a racking sheet slips its perimeter nails almost equally
experiments/m16_soft_nails/ (run.py, results.txt, results.json, figure.py, figure.png); src/buildbot_sheathing/softnail.py (pushover of milestone 14's sheet with each nail a nonlinear spring along its slip); tests/test_softnail.py (5 tests). About 25 min on one CPU (about 7,000 pushovers of 600 steps).
- One change from the plan, made before any capacity was computed: the rise of the backbone is P = Fp(1 − (1 − δ/δp)^n) instead of Foschi's exponential, because it peaks smoothly at δp with three numbers that the test paper gives or the sweep sets (Fp, the secant stiffness at 0.4 Fp, δp); the fall is as planned. n = 3.65–9.52 over the sweep.
- H70 held. Elastic-perfectly-plastic nails yielding at 0.01 mm give milestone 14's capacity to 0.000% on the intact sheet, the Weiss pattern and 200 three-read miss patterns.
- H71 held. At the tested backbone (δp 12 mm) the 4, 3 and 2 in sheets carry 1.444, 1.888 and 2.775 of the 6 in sheet, against SDPWS's 1.536, 1.964 and 2.607 (−6%, −4%, +6%), milestone 14's ratios to within 0.006. Every spacing peaks at 55–56 mm of drift (2.3%). The 6 in sheet peaks at 9.80 kN, 551 plf at the tested Fp, beside 560 plf nominal for the heavier 15/32 in Structural I panel (not a like-for-like check, but the right size).
- H72 refuted, firmly. With every nail below 12 in omitted the sheet loses 20.1–20.4% at every δp from 6 to 15 mm and both the monotonic and cyclic Δu (23.4–23.8% at z ≤ 12 in), against 20.4% rigid-plastic. Post hoc (section G), a near-brittle nail that falls to 0.8 Fp 1 mm past its peak gives 20.4% too. The reason: at the sheet's peak the perimeter nails all slip 11–12 mm (median 11.0, max 12.3 mm) and only the field nails near the centre of rotation lag (3–4 mm), so the nails reach their peaks together and the ratio is the rigid-plastic one. No nail unloads before the sheet's peak, so the path-independent spring is exact up to it.
- H73 held. Three reads on measured studs, every miss escaping: mean loss 2.39%, 99th percentile 11.80% (milestone 14: 2.34%, 11.49%); a 4-sheet wall's 1st percentile 0.932 (0.934). The cyclic Δu gives 2.40% and 11.81%.
- H74 held. 4 in joint edges: intact 1.320 of the 6 in sheet; with misses left in, 99.98% of sheets at or above the intact 6 in sheet.
- What it means for the programme. Milestone 15 left three candidates for why the model loses 20% where Weiss's walls lost about half: the framing joints, the nails' softening and the anchorage. Code-strength framing joints (milestone 15) and softening nails (here) are now both ruled out with sourced inputs, and in milestone 15 removing hold-downs moved the loss to 64–100%. So the anchorage is the remaining candidate: the next step is a model with hold-down and sill-plate stiffness and uplift, or Weiss's full text to read how her walls were held down. Milestone 14's escaped-miss figures and the 4 in joint-edge schedule stand under a tested nail curve, which is what a panel-line integrator would act on.
- Inputs not sourced: the slip at peak δp (swept 6–15 mm; not tabulated in the test paper); the fall beyond 0.8 Fp (straight to zero at δp + 5(Δu − δp)). Monotonic loading only.
2026-11-14 · Milestone 16 planned: do softening nails explain why the model understated Weiss's walls?
Written before any milestone-16 capacity was computed. Inputs: milestone 14's nail layout, kinematics (pinned framing, rigid sheet, full anchorage) and miss patterns (20,000 sheets, seed 7, measured studs).
- Why. Milestones 14 and 15 scored missing nails with rigid-plastic fasteners: every nail reaches the same capacity F at any slip, so the sheet's capacity is the sum over nails of F times its share of the slip. Real nails peak at a slip of some millimetres and then soften, so at the sheet's peak the corner nails, which slip most, are already past their peak while the nails near the sheet's centre of rotation have not reached theirs. Removing the bottom foot's nails moves the sheet's centre of rotation up and makes the remaining nails' slips more unequal, which a rigid-plastic model cannot see. Milestone 15 ruled out the stud-to-plate joints at code strength; this is the second of the three candidates it left (anchorage, weaker joints or fasteners, softening nails).
- Model. Milestone 14's kinematics with each nail a nonlinear spring along its slip (the oriented-spring idealisation behind CASHEW, Folz and Filiatrault 2001): load P(δ) = (F0 + r1 K0 δ)(1 − exp(−K0 δ / F0)) up to the peak slip δp, then a straight fall through 0.8 Fp at Δu to zero. Displacement-controlled pushover of the top plate to 150 mm; at each step the sheet's (u, v, θ) minimises the springs' energy (Newton from the last step, monotonic loading taken as path-independent; the fraction of nails that unload is reported). Capacity is the peak racking load. Backbone from tests of 8d common nails through 7/16 in OSB into SPF (Shameli Derakshan, Zhou and Ni 2019, Table 2; six replicates a group): Fp 0.90–1.06 kN monotonic (1.01–1.26 cyclic), secant stiffness at 0.4 Fp 0.40–0.57 kN/mm (cyclic; the monotonic K is unreliable, the authors say), Δu 26.7–31.2 mm monotonic and 18.9–21.4 mm cyclic. The slip at peak δp is not tabulated: swept 6–15 mm (r1 set so the curve peaks there). Capacity ratios do not depend on Fp.
- H70 (the model reduces to milestone 14). With an elastic-perfectly-plastic backbone yielding at 0.01 mm, the peak load matches
racking.capacity within 0.5% for the intact sheet, the Weiss pattern and 200 random miss patterns. Refuted otherwise (a bug, not a finding). - H71 (spacing table). At the measured backbone (monotonic, δp 12 mm), the intact sheet's capacity at 4, 3 and 2 in edge spacing relative to 6 in lies within 10% of SDPWS's ratios (1.536, 1.964, 2.607). Refuted otherwise.
- H72 (softening explains Weiss). Omitting every nail below 12 in loses at least 40% of the intact sheet's peak at every backbone in the sweep (δp 6–15 mm, monotonic and cyclic Δu). Refuted otherwise.
- H73 (escaped misses unchanged). Three reads on measured studs, every miss escaping, at the measured backbone: the mean sheet's loss and its 99th-percentile loss are each within 1 point of milestone 14's (2.34% and 11.5%). Refuted otherwise.
- H74 (4 in joint edges still carry the misses). With the measured backbone, the 4 in joint-edge schedule with every miss escaping keeps at least the intact 6 in sheet's peak in 99% of sheets. Refuted otherwise.
- Also reported, not scored: the drift at peak; the Weiss loss against δp and Δu; how far the sheet's centre of rotation moves; the fraction of nails past their peak at the sheet's peak; the absolute peak in kN/m at the tested Fp, beside SDPWS's tabulated value for a heavier panel (not a like-for-like check); the single-nail losses.
- Limits known in advance. Monotonic loading only (no cyclic degradation, which is how most shear-wall tests are run); pinned framing, rigid sheet, full anchorage, as in milestone 14; one nail-joint study, SPF framing (Weiss's framing species not read); the slip at peak is swept, not measured.
2026-11-14 · Milestone 15 result: stud-to-plate joints at NDS strength do not explain Weiss's walls; they would need half that strength, or no hold-downs
experiments/m15_frame_joints/ (run.py, results.txt, results.json, figure.py, figure.png); src/buildbot_sheathing/frame_lp.py (the framed sheet as a linear programme over rigid-body motions of the top plate, four studs and the sheet; the NDS yield-limit equations); tests/test_frame_lp.py (5 tests). 11 min on one CPU (about 4,000 LPs).
- The NDS ratio. An 8d common through 15/32 in OSB into G 0.50 framing yields at Z = 74.0 lb (mode IIIs); a 16d common through a 1.5 in plate at 140.7 lb (mode IV, which matches the 141 lb of NDS Table 12N); two end nails at the end-grain factor 0.67 give 188.5 lb, so ρ = 2.55 sheathing nails (2.34 in SPF).
- H65 held. With rigid joints the LP's capacity ratios match milestone 14's within 0.09% over 200 miss patterns and the Weiss pattern (0.7970 against 0.7963). Its absolute intact capacity is 1.2% above milestone 14's because the joint-edge nails sit 3/8 in off the stud centreline, so a rotating stud also slips them vertically; with the studs moved under the nails it agrees to 0.04%.
- H66 refuted, for a reason the plan did not foresee. At the NDS joints the intact 6 in sheet loses 7.2% (3.9–7.2% across 6–2 in spacing), and the spacing check fails (2 in: 2.87 against SDPWS's 2.61). The mechanism is not the stud joints sliding: a lone sheet's top plate tilts (0.18 rad per unit slide) while an interior stud's bottom lifts off its plate, which costs nothing when end-grain withdrawal is zero. Withdrawal of 2 F per joint, or (post hoc, section H) a top plate that only slides, as a continuous double top plate over a wall line would hold it, removes the loss (0.00% at 6 and 4 in, −1.5% at 2 in).
- H67 refuted, robustly. At the NDS joints the Weiss pattern loses 21.1% (19.7–21.1% for any withdrawal 0–2 F; 20.3% with the sliding top plate), essentially milestone 14's 20.4%: the joints are strong enough that the studs never slide. A 40% loss needs ρ ≤ 1.42 (0.56 of the NDS ratio) and 50% needs ρ ≤ 1.18 (0.46). Without hold-downs (not scored) the same sheet loses 64% at a withdrawal of 1 F and all of it at 0; the intact sheet without hold-downs keeps only 18–29%.
- H68 refuted as registered; holds with a sliding top plate. Three reads on measured studs, every miss escaping, at the registered joints: mean loss 2.73% (milestone 14: 2.34%), 99th-percentile 16.7% (11.5%), because the tilting-top-plate mechanism loads the joint edges harder (a joint-edge nail is worth 2.38% instead of 2.08%, a field nail 0.27% instead of 0.85%). With the top plate held to a slide: 2.34% and 11.51%, within 0.02 points of milestone 14.
- H69 held. 4 in joint edges keep at least the intact 6 in sheet's capacity in 99.89% of sheets (intact 1.365 relative to the registered intact 6 in sheet, minimum 0.710), and 99.98% with the sliding top plate.
- What it means for the programme. Milestone 14's rigid-plastic model is not unconservative because of the stud-to-plate joints at code strength; the gap to Weiss's walls is somewhere else: the anchorage (the model loses 64–100% without hold-downs), joints or fasteners far weaker at ultimate than the NDS ratio implies, or softening nails, none of which this model can choose between without Weiss's test set-up. So milestone 14's post-hoc 2.45 scaling stays a bracket, not a correction. Two things are firm: an escaped miss's cost depends on the top plate's restraint (a sheet whose top plate can tilt loses half as much again at the tail), and the 4 in joint-edge schedule carries the misses under every joint model tried.
- Inputs not sourced: the NDS ratio is a design-level (yield-limit) ratio, not tested ultimates; joints are rigid-plastic; friction at stud bearing and hold-down shear are ignored. Weiss's wall size, anchorage and framing were not read (abstract only; the thesis host does not resolve from this session).
2026-11-14 · Milestone 15 planned: do the framing joints explain why the model understated Weiss's walls?
Written before any milestone-15 capacity was computed. Inputs: milestone 14's nail layout, miss patterns (20,000 sheets, seed 7) and its intact-sheet reference.
- Why. Milestone 14's rigid-plastic model lost 20% of a sheet's capacity with every nail below 12 in omitted, where Weiss's (2017) tested walls lost "close to or more than half" (H61 refuted), so every escaped-miss loss it reported is a lower bound of unknown slack, and its post-hoc fix scaled every loss by 2.45. One reason the model cannot see: it pins the studs to the plates with joints of unlimited strength. With the bottom-plate sheathing nails gone, the racking load has to leave the sheet through the studs and reach the bottom plate through each stud's two end nails, which are weak in shear (end grain). If that is the reason, the loss is a property of which nails are missing (the bottom-plate row), not a uniform factor, and scattered joint-edge misses keep milestone 14's small losses.
- Model. A rigid-body limit analysis of the whole framed sheet, solved as a linear programme (upper bound: the least dissipation over kinematically admissible small motions with the top plate moved one unit along the wall). Bodies: the bottom plate (fixed), the top plate, each stud and the sheet, each with (u, v, θ). Sheathing nails: rigid-plastic, capacity F along their slip (the Euclidean norm, a 64-sided polygon). Stud-to-plate joints, top and bottom: two 16d common end nails (IRC Table R602.3(1)), lateral capacity ρF with ρ from the NDS yield-limit equations (dowel bearing strength 16,600 G^1.84 psi, nail bending yield by diameter, end-grain factor 0.67 on the end nail; the 8d sheathing nail through 15/32 in OSB at G 0.50 into G 0.50 framing), and swept 0.25–8 because the yield model is a design-level ratio, not a measured ultimate; they bear in compression and open in tension against a withdrawal capacity (end-grain withdrawal is not allowed in NDS: 0, swept to ρF). Hold-downs at the sheet's end studs fix them vertically (fully anchored); a no-hold-down case is reported, not scored.
- H65 (the LP reproduces milestone 14). With rigid (equality) stud-to-plate joints the LP's capacity ratios match
racking.capacity within 0.5% for the intact sheet, the Weiss pattern and 200 random miss patterns from the three-read cell. Refuted otherwise (a bug, not a finding). - H66 (finite joints leave the intact sheet alone). At the NDS ρ, the intact sheet's capacity at 6, 4, 3 and 2 in edge spacing changes by less than 2% each from milestone 14's, so H60 still holds. Refuted otherwise.
- H67 (finite joints explain Weiss). At the NDS ρ, omitting every nail below 12 in loses at least 40% of the intact sheet's capacity. Refuted otherwise.
- H68 (scattered misses stay small). At the NDS ρ, with every miss escaping (three reads, measured studs), the mean sheet's loss and its 99th-percentile loss are within 1 point of milestone 14's (2.34% and 11.5%), so milestone 14's post-hoc 2.45 scaling overstated them. Refuted otherwise.
- H69 (the 4 in joint edges still carry the misses). At the NDS ρ, milestone 14's 4 in joint-edge schedule with every miss escaping keeps at least the intact 6 in sheet's capacity in 99% of sheets. Refuted otherwise.
- Also reported, not scored: the ρ at which the Weiss loss reaches 40% and 50%; the loss of a single missed bottom-plate nail; the mechanism the LP picks (stud sliding on the plate, or racking); no hold-downs; doubled joint studs as two members.
- Limits known in advance. Weiss's walls' size, anchorage and framing were not read (only the abstract; the thesis host is unreachable from this session), so H67 compares against a qualitative "close to or more than half". The yield model gives design-level ratios, not tested ultimates; the joints are rigid-plastic (no stiffness, no softening); friction at the stud bearing and hold-down shear are ignored, which favours larger losses.
2026-11-13 · Milestone 14 result: an escaped miss costs about 2% of the sheet; misses cluster, and 4 in joint nailing covers them for less than a re-drive
experiments/m14_escaped/ (run.py, results.txt, results.json, figure.py, figure.png); src/buildbot_sheathing/racking.py; tack.simulate gains u_joint and return_hits, passed through ceiling.clean; tests/test_racking.py (5 tests). 9 s to run. Milestone 12's cells reproduce (0.4253 clean and 1.095 missed for three reads; 0.7707 for the full stack). The plastic-work minimum is solved by reweighted least squares and matches Nelder–Mead to 1e-13.
- H60 held. Against SDPWS Table 4.3A (15/32 in Structural I, 8d; 280, 430, 550, 730 plf), the model's intact-sheet ratios at 4, 3 and 2 in edge spacing are 1.442, 1.884 and 2.769 against 1.536, 1.964 and 2.607 (−6%, −4%, +6%). The table values were read from secondary copies (a state code reprint, a manufacturer's tech tip quoting SDPWS 2021), not the standard itself.
- H61 refuted, and it matters. Omitting every nail below 12 in (11 of 62) removes 20.4% of the model's capacity (23.7% at ≤ 12 in), against Weiss's (2017) tested walls, which lost "close to or more than half". The rigid-plastic model is unconservative for clustered missing nails by a factor of about 2.5 there. Possible reasons, not tested: the tests' walls may not have been fully anchored, a sheet with no bottom-edge nails rocks and tears out at the next row, and real nails are not rigid-plastic. Every loss below is therefore a lower bound.
- H62 held. One joint-edge nail missed alone costs 2.00–2.21% of the sheet (mean 2.08%; corners most), a field nail 0.70–1.02%, a plate nail 1.05–1.83%. The 62 single losses sum to 1.024, so a nail carries about its share by count, weighted toward the perimeter.
- H63 refuted. Three reads on measured studs, every miss escaping: the mean sheet loses 2.34% (2.13% per miss), but the 99th-percentile sheet loses 11.5% (bar 10%), and a 16 ft wall line's 1st percentile keeps 93.4% (bar 95%). The misses cluster: 13.4% of sheets have three or more joint misses (worst: 11), because a stud the robot misjudges is misjudged for a run of heights. The full stack (0.771 clean) loses 0.63% on average, 6.3% at the 99th percentile; doubled joints lose nothing.
- H64 held. Nailing the two joint edges at 4 in (25 a side; plates 6 in, field 12 in) raises the intact sheet to 1.318 of the 6 in sheet; with 1.572 misses a sheet all escaping, 99.98% of sheets still carry at least the intact 6 in sheet's capacity (1st percentile 1.160, worst 0.868). Post hoc, with every loss scaled by Weiss's 2.45, 97.6% do and the 1st percentile is 0.930.
- Also found. Sixteen extra nails at milestone 7's robot seconds per nail cost less than re-driving 1.1 misses a sheet: at 100 houses a year the 4 in joint edges with misses left cost a median $6.98 per sheet against $7.10 for 6 in with every miss re-driven and $12.17 for doubled joints (cheaper than doubled joints in all 4,000 draws). Nail material is not priced (a few cents). Miss rates by height run 1.8–4.4% a nail, highest at the ends and at 42–54 in, which is where the measured studs' waviness sits.
- What it means for the programme. Milestone 13 asked what a missed nail costs; in capacity, about 2% of its sheet each, more when they cluster, and the model likely understates clustered losses. For a shear wall, a design that expects misses can carry them: 4 in nailing on single-2x joint edges keeps the 6 in schedule's capacity with every miss left in, for no more than re-driving. Whether a code official accepts a schedule that knowingly includes shiners is a separate question this cannot answer; the honest framing is that it is a robot-friendly design margin, not a licence to skip inspection.
- Inputs not sourced: the rigid-plastic fastener (equal capacity, no edge-distance or tear-out effects), full anchorage, sheets acting independently; the seconds of an added nail (t_nail / 62).
2026-11-13 · Milestone 14 planned: what an escaped missed nail does to the sheet's racking capacity
Written before any milestone-14 capacity was computed. Inputs: milestone 12's pipeline (ceiling.clean, measured studs, 110 mm mean, seed 7), now asked for which nails miss, not how many.
- Why. Milestone 13 turned the framing question into a question about missed nails: doubled joints pay only if a missed nail costs about $5.51 more than its re-drive. Every milestone so far assumed each miss is found and re-driven. A miss that escapes (a "shiner" beside the stud) is a fastener missing from a shear wall's schedule, and its cost depends on how much racking capacity it takes away. That is computable without a price, as a fraction of the sheet's capacity, from where the misses fall: single-2x misses sit on the two joint edges, and a fastener's share of a sheet's capacity depends on its place.
- Model. The rigid-plastic model of a fully anchored sheathed wall (Källsner and Girhammar 2009, as we read it): framing pinned at its joints, so a stud nail at height z moves γz along the wall and a top-plate nail γH; the sheet rigid (translation u, v and rotation θ); every fastener rigid-plastic with the same capacity F in the direction of its slip. The racking capacity is the least work over the sheet's rigid motions, H·γ·H_wall = F·min Σ|slip_i|, so a capacity ratio needs only the nail positions, not F. Nails: SDPWS's 6 in edge and 12 in field schedule as the programme already places them (17 per joint edge, 7 per field stud, 7 per plate, 62 in all). A wall line is the sum of its sheets (they share γ). Misses come from
tack.simulate, now returning the hit pattern of each sheet. - H60 (the model reproduces the spacing table). The intact sheet's capacity at 4, 3 and 2 in edge spacing (field 12 in), relative to 6 in, lies within 10% of SDPWS Table 4.3A's ratios for 15/32 in Structural I with 8d common nails (430, 550 and 730 plf against 280: 1.536, 1.964, 2.607). Refuted if any is outside.
- H61 (the model agrees with a test of missing nails). Weiss (2017, MS thesis, WSU) omitted every sheathing nail in the bottom 12 in of tested walls to simulate decay, and the walls lost "close to or more than half" their capacity. The model, omitting every nail below 12 in, loses at least 40%. Refuted otherwise (the model would then be unconservative for clustered misses).
- H62 (a missed nail costs about its share). For each of the 34 joint-edge nails missed alone, the sheet loses at most 4% of its capacity, and the mean over the 34 lies between 1% and 3% (1/62 = 1.6%). Refuted otherwise.
- H63 (escaped misses at the simulated rate are small). Single 2x on measured studs, three reads, ledger (milestone 12: 0.425 clean, 1.095 missed a sheet), every miss escaping: the mean sheet loses under 3% of capacity and the 99th-percentile sheet under 10%, and a four-sheet (16 ft) wall line keeps at least 95% of its intact capacity in 99% of walls. Refuted if any of the three fails.
- H64 (misses can be designed out with nails). The same pipeline with its two joint edges nailed at 4 in (25 nails each; plates 6 in, field 12 in) and every miss escaping keeps at least the intact 6 in sheet's capacity in 99% of sheets. Refuted otherwise.
- Also reported, not scored: the capacity lost by misses on the best single-2x stack (per-nail reads, push, slide and shim; 0.771 clean) and on doubled joints; which joint heights miss most and what each height's nail is worth; the extra nailing seconds and dollars of H64's denser joint edges in milestone 7's cost model, against doubled joints' $5.13 premium; the break-even of milestone 13 restated as a capacity.
- Limits known in advance. A rigid-plastic model with equal fasteners: no stiffness, no sheathing tear-out, no nail-to-edge-distance effects, no partial anchorage; a shiner is assumed to carry nothing; sheet-to-sheet interaction at a shared joint stud is ignored (each sheet has its own edge nails); one wall of six measured studs drives the misses. The model ranks miss patterns and sizes their effect relative to an intact sheet; it does not give a design value.
2026-11-13 · Milestone 13 result: the remedies cost 32 cents a clean sheet; the doubled joint stud costs $5.13 and still loses every draw
experiments/m13_price/ (run.py, results.txt, results.json, figure.py, figure.png); src/buildbot_sheathing/price.py; tests/test_price.py (5 tests). 5 s to run. One new Monte Carlo cell (doubled joints, single read, measured studs: 1.0000 clean, 0 missed); everything else reuses milestone 12's clean sheets and milestone 7's 4,000 cost draws.
- H57 held. At 100 houses a year the per-draw cheapest single-2x pipeline beats doubled joints in 1.000 of draws (bar 0.90), and also at 300 and 1,000; the median saving is $5.13 a sheet. Doubled joints are the fastest pipeline (145 s against 198–304 s) but the stud and a framer's time are 44% of their $12.17.
- H58 held. The full stack (per-nail reads by line laser, mid push, slide and shim; 0.771 clean) costs a median $7.42 per clean sheet against $7.10 for three reads on the ledger (0.425 clean), and is the cheapest pipeline in none of the 4,000 draws. Interpretation note: the plan said "the cheapest single-2x pipeline"; since a per-draw minimum can never be beaten by one of its members, it was scored against the lowest-median pipeline. Three reads on the ledger is cheapest in 52% of draws, per-nail reads by line laser on the ledger in 26%, three reads with slide and shim in 14%.
- H59 refuted. The extra cost a missed nail must carry for doubled joints to tie is a median $5.51 [5–95%: $3.91, $8.55] (8.8 loaded person-minutes), down from milestone 7's $6.51 but above the $5.00 bar; $5.83 with the common part removed. It fell less than the missed nails rose (0.79 to 1.095) because in about half the draws the cheapest single-2x pipeline is one that misses fewer nails (per-nail reads, 0.765; slide and shim, 0.726), which shrinks the denominator: the median saving over 1.095 alone would give $4.68.
- Also found. Below the base's capacity, seconds cost only supervision, so the price of first-pass quality is flat up to 0.77 (the cheapest pipeline at 0.40, 0.50, 0.60 and 0.75 first-pass clean costs $7.10, $7.14, $7.24 and $7.42) and then jumps to $12.17 for 1.000. A line laser beats three point reads at per-nail resolution above about 0.23 m/s; per-nail reads by point cost 15 cents and 31 s more than three reads.
- What it means for the programme. In dollars, the paper's question about framing is now a question about missed nails: if a missed nail found at inspection (or not found) costs more than about $5.50 beyond its re-drive, double the joint; otherwise run single 2x with whichever remedies the line allows, since they are nearly free. The next input to get is that cost and the detection rate of a missed nail, not more sensing.
- Inputs not sourced: the slide-and-shim time, the sweep speed and the push time (all swept); every missed nail assumed found.
2026-11-13 · Milestone 13 planned: what the single-2x remedies cost, and whether doubling pays on measured studs
Written before any milestone-13 pipeline was priced. Inputs are milestone 12's committed results.json (clean sheets and missed nails for the 12 single-2x cells on measured studs and the 12 with the common part removed) and milestone 7's cost model and 4,000 draws (seed 11).
- Why. Milestone 7 found doubled joints never pay in dollars unless a missed nail costs about $6.51 more than its re-drive, but it priced parabolic low-warp studs (0.79 missed nails a sheet with three reads). On measured studs the same pipeline misses 1.09 nails a sheet, and the remedies of milestones 9, 10 and 12 (per-nail reads, a mid push, a slide and shim) each buy clean sheets with seconds milestone 7 never charged. Milestone 12 left the per-nail read's time open. The question a panel-line integrator asks: which single-2x pipeline is cheapest per clean sheet on real studs, and is the extra joint stud now worth it?
- Model. Milestone 7's G1 on a drive-and-lift base (the cheapest robot class), 100, 300 and 1,000 houses a year; every missed nail found and re-driven (t_fix). Seconds added per sheet: point reads at milestone 7's t_read U(1, 3) s each (three: 3 per stud; dense: 17 per stud; local: 17 on the joint stud and 7 on each of two field studs; plus 2 edge reads); or, for dense and local, a line-laser sweep along the three new studs, 3 × 2.353 m / v with v log-uniform 0.05–0.5 m/s (unsourced; ISO 10218-1's 250 mm/s reduced speed sits inside it) plus the 2 edge reads; the mid push U(10, 40) s (milestone 9); slide and shim U(5, 30) s plus 2 more edge reads (unsourced, swept). Doubled joints: single read, ledger, the extra stud and a framer's time (milestone 7), clean sheets and misses from a new 20,000-sheet run on the measured studs.
- H57 (doubling still does not pay). At 100 houses a year, the cheapest single-2x pipeline on measured studs (chosen per draw over all 12 cells and both read hardwares) is cheaper than doubled joints in at least 90% of draws. Refuted otherwise.
- H58 (the remedies buy first-pass quality, not dollars). In the median draw at 100 houses a year, the full stack (local reads by line laser, mid push, slide and shim; 0.771 clean) costs more per clean sheet than the cheapest single-2x pipeline. Held if it costs more; refuted otherwise.
- H59 (measured studs lower the break-even). The median extra cost a missed nail must carry, beyond its re-drive, for doubled joints to tie the per-draw cheapest single-2x pipeline is below $5.00 (milestone 7: $6.51 on parabolic studs). Refuted otherwise.
- Also reported, not scored: median dollars and seconds of every pipeline; the cost of first-pass quality (the cheapest pipeline reaching each clean-sheet level); the line-laser speed at which per-nail reads stop paying for themselves; the common-removed studs as a bracket; 300 and 1,000 houses a year, where the machine's capacity can bind.
- Limits known in advance. The slide-and-shim time and the sweep speed are unsourced and swept; missed nails are all assumed found, so an escaped miss (a shear-wall nailing defect) has no price here, which is exactly what H59's break-even asks the reader to supply; one robot-built wall's six studs.
2026-11-13 · Milestone 12 result: per-nail reads and the mid push both survive measured studs; together with slide and shim they reach 0.77, not 0.90
experiments/m12_remedies/ (run.py, results.txt, results.json, figure.py, figure.png); tack.simulate gains a local read policy and a mid push on any stud shape; measured.model_profiles holds milestone 11's profile variants; ceiling.clean takes profiles; tests/test_remedies.py (5 tests). 39 s to run. Milestone 11's three cells reproduce to four decimals (0.4253, 0.4707, 0.5792).
- A bug in the first run, fixed before any result was used. Milestone 11's profiles are measured about each stud's least-squares line, not the chord through its ends; the first run took the push from the least-squares line, so the push moved a stud by a third of its bow in the wrong place and made things worse (three reads 0.425 to 0.428; the end-chord RMS of five studs went up). The registered model says "offset from its chord" and the push is physically pinned at the plates, so the push now uses the end chord (
test_push_is_measured_from_the_end_chord). The results below are from the fixed code only; the parabola and straight cells were unaffected (their profiles are zero at the ends). - H53 held: most of the measured-stud loss is the estimate. Reading the stud at every nail height and aiming at each read gives 0.530 on measured studs (ledger, no push), above the 0.502 bar; the waviness cost falls from 15.4 points (three reads) and 12.5 (dense parabola) to 4.7. On the parabola and straight studs per-nail reads gain nothing (0.577 and 0.813 against 0.579 and 0.811): their only use is waviness.
- H54 held: the push still pays. A mid push (three reads, ledger) lifts measured studs from 0.425 to 0.552 (+12.6 points; bar +10), against +16.3 on their own parabolas (0.579 to 0.742). Pushed onto the end chord, four of the six studs straighten (end-chord RMS 0.83–3.49 mm to 0.45–1.20) and the most leaning end stud gets worse (0.94 to 1.56), because its shape is not a bow. Forces 10–72 N at a mean Stud-grade E, inside milestone 9's range.
- H55 refuted: the remedies stack, but to 0.77. Per-nail reads, the mid push and a 0.5 mm slide plus shim give 0.771 on measured studs (0.796 with the common part removed), against 0.860 on their own parabolas and 0.92 on straight studs. 95% then needs a tool held to 0.61 mm (1σ; 0.63 with the common part removed) against 0.84 mm on straight studs; three reads with the push need 0.27 mm.
- H56 held. Doubled joints give 1.0000 in all six estimate-placement cells on measured studs.
- What it means for the programme. The single-2x route is now: read at every nail (a line laser along the stud, not three spots), push the joint stud to its end chord, slide and shim the sheet, and hold the tool to about 0.6 mm. Each step is needed and none is enough alone. Doubled joints still need none of it. The paper's single-2x recommendation narrows accordingly.
- Inputs not sourced: the slide and shim residuals (0.5 mm, milestone 10's sweep), the push error (= the 0.5 mm read), linear superposition of the push on the scanned shape, and the length-squared scaling from 2,220 mm.
2026-11-13 · Milestone 12 planned: do the single-2x remedies survive measured studs?
Written before any milestone-12 pipeline was run. Inputs are milestone 11's committed profiles.json and its 110 mm running mean.
- Why. Milestone 11 found the six measured studs are not parabolas, and the best single-2x pipeline drops from 0.579 (each stud's own parabola) to 0.425 on their measured shapes. Milestones 9 and 10 proposed two remedies, a mid-height push onto the line and a slide plus shim before tacking (0.92–0.94 on straight studs), both scored only on parabolic or straight studs. Milestone 11 also left open whether the loss is the robot's estimate (a parabola fitted to three or seventeen reads cannot follow the waviness) or the geometry (a straight sheet edge against a wavy face narrows the window however well it is known). These are different fixes: more sensing, or more framing.
- Model. Milestone 11's pipeline (single 2x, smooth-rock Drake seats, tool 0.9 mm, reads 0.5 mm, edge aim, 20,000 sheets, seed 7) on four stud sets: measured (110 mm mean), measured with the cross-member common part removed, each stud's own parabola, and straight. Three switches: (1) the estimate, three (parabola through three reads), dense (parabola through 17) or a new local policy that reads the stud at every nail height (17 on a joint stud, 7 on a field stud) and aims each nail at its own read, no fitted shape; (2) a mid-height push, generalised to any measured shape by superposition: the stud moves by (its mid-height offset from its chord + N(0, s_read)) times the pinned point-load shape, and the robot believes its own estimate moved by its estimated offset times the same shape; (3) placement, milestone 9's ledger seat or milestone 10's slide plus shim (0.5 mm residuals). Doubled joints on the measured studs as a check.
- H53 (the loss is mostly the estimate). On measured studs (110 mm mean), ledger placement, no push, local reads close at least half of the gap from three-read measured (0.425) to three-read own-parabola (0.579), i.e. reach 0.502 or more. Refuted otherwise.
- H54 (the push still pays). On measured studs, three reads, ledger placement, the mid push lifts clean sheets by at least 10 points over no push (milestone 9 found +15 on parabolic low-warp studs). Refuted otherwise.
- H55 (the remedies stack). The best single-2x cell on measured studs (any estimate, with or without push, slide plus shim 0.5 mm) reaches 0.90 clean sheets. Refuted otherwise.
- H56. Doubled joints give at least 0.99 on measured studs in every estimate and placement. Refuted otherwise.
- Limits known in advance. Six studs from one robot-built wall; the push is linear superposition on the scanned (already installed) shape, which assumes the plates hold the ends pinned as milestone 9 did; part of the waviness may be scanner artefact (milestone 11, post hoc), which the common-removed set brackets; the local policy's read count (24 per new stud pair against 3) has a time cost milestone 5 prices, not scored here.
2026-11-13 · Milestone 11 result: measured studs bow by the published amount, but not as parabolas; the parabola flattered single 2x by 11–15 points
experiments/m11_measured/ (extract.py, profiles.json, run.py, results.txt, results.json, figure.png); src/buildbot_sheathing/measured.py; tack.simulate gains a profiles argument; 5 s to extract, 16 s to run.
- H49 refuted, and the reason sets the resolution. The faces read 35.6–36.0 mm wide (bar 38.1 ± 2) and centrelines jump 1.2–2.6 mm between some neighbouring slices. The scan takes a profile every 2 mm along the wall, so an edge between two profiles is known to about ±1 mm. A synthetic 38.1 mm face with eased corners on the same grid reads 36.0 mm wide and recovers bows of 0–6 mm to 0.25 mm RMS, leaving 0.27 mm RMS of quantisation after a perfect parabola. Bow amplitudes are well resolved; millimetre-scale shape is resolved only after averaging along the stud.
- H50 held. Scaled to 8 ft, the six studs bow 0.021–0.283 in, median 0.106 in, on the published low mean of 0.11 in; with the five inclined members, median 0.074 in, mean 0.105 in. The programme's bow magnitudes, taken from search snippets, are confirmed by a measurement.
- H51 refuted. On the four studs bowed over 1 mm, the parabola explains 21–79% of the shape about the chord (bar 80%) and leaves 0.59–0.94 mm RMS (bar 0.5). A 110 mm running mean brings it to 0.39–0.76 mm.
- Post hoc, not registered: some of that is the scanner. The residuals correlate across all eleven members (median pairwise r 0.47) and their mean carries 48% of the residual variance, against 9% if independent; wood waviness should not line up across members at the same heights, a pass registered slightly off would. Removing the common part leaves 0.31–0.60 mm RMS on the studs, still at or above the quantisation floor.
- H52 refuted. Single 2x, three reads: measured studs give 0.425 clean sheets (raw slices 0.311, common part removed 0.471, twist off 0.461) against 0.579 for each stud's own parabola and 0.625 for the model's low-warp draw. The parabola is 11–15 points optimistic, because a three-read (or 17-read) parabola cannot follow a millimetre of waviness inside a ±3 mm window. Doubled joints give 1.0000 in every variant.
- Also found. Layout about a 402 mm grid: sd 2.0 mm at the bottom, 5.1 mm at the top; the two end studs lean 14.6 and 6.5 mm and run 20.5 and 10.9 mm out of plane, the inner four studs within 1.7 mm of lean and 2–3.5 mm out of plane. The nail-laminated panel's benchmark poses reproduce the paper's Table 1 exactly (mean 1.282 mm); the adaptive file does not (1.956 against 0.401 mm), so it was not used. The paper's bibliography had two author first names wrong (Wes/Saeed for Wesley/Salma) and no DOI; both fixed from the NSF public-access record.
- What it means for the programme. The single-2x numbers in sections 5.2–5.9 are upper bounds by about 11–15 points; the doubled-joint recommendation is unchanged and better supported. The slide-and-shim route of milestone 10 (0.92–0.94 on straight studs) needs straight studs in the strong sense, a pushed stud keeps its waviness. The next measured input worth getting is a population: framer-built walls scanned along the studs, finer than 2 mm across them.
- Inputs not sourced: the length-squared scaling from 2,220 mm to the 92 5/8 in stud; treating the cross-member common part as scan artefact (it may include real shape the robots' fitting imposed).
2026-11-13 · Milestone 11 planned: the first measured studs, from a scanned robot-built 2x4 wall
Written before any stud profile was extracted. What had been looked at: the data set's README, the paper's description of the wall module, and one 100 mm slice at mid-height (to see which members the scan shows and that their faces are about 40 mm wide).
- Why. Every stud in milestones 2–10 is a parabola whose amplitude is drawn from two published means seen through search snippets; milestone 2 left "measured stud profiles in place of means and assumed shapes" open, and the paper's section 2 names the Adel et al. (2024) data set as the most promising source. It turns out to hold a laser-profiler scan (0.15 mm profile, 0.019 mm depth resolution, 2 mm between profiles) of a full-scale 2x4 wall module two industrial robots built: 2.05 m by 2.32 m, two plates, six vertical studs at about 16 in, five inclined members, scanned from one face. The scanned face is the 1.5 in face sheathing is nailed to. So the shape of six installed studs, as a sheathing robot would meet them, can be measured instead of assumed.
- Model. Fit the wall's plane to the plate and stud faces. For each vertical stud, slice the face every 10 mm of height; from each slice take the face's two side edges (centreline and width), its mean height off the plane (in and out of the wall: crook plus warp of the wall) and its slope across the face (twist). The flatwise bow is the centreline's deviation from the chord through its ends; lean and layout are the chord's ends against the nominal 16 in grid fitted to the six studs. The inclined members and plates give more in-plane straightness profiles, about their own chords. Then the milestone-9 single-2x ledger pipeline (
tack.simulate, three reads, s_x 1 mm, edge aim, smooth-rock seats, 20,000 sheets, seed 7) with each sheet's four studs drawn from the six measured shapes (random left-right and top-bottom flips, deviation scaled from the measured length to the 92 5/8 in stud by length squared), against the same pipeline with each measured stud replaced by its own fitted parabola. - H49 (the scan resolves a stud). On every vertical stud the median face width is 38.1 ± 2 mm and the centreline has no step over 0.3 mm between neighbouring slices beyond what its own slice-to-slice scatter explains (a step at a pass boundary would be stitching). Refuted otherwise.
- H50 (magnitude). Scaled to 8 ft, every vertical stud's flatwise bow is under the Stud-grade limit (3/4 in) and their median is at or below the high published mean (0.21 in). Refuted otherwise.
- H51 (shape). On every vertical stud whose bow exceeds 1 mm, the chord-plus-parabola fit explains at least 80% of the variance of the deviation from the chord and leaves an RMS residual of at most 0.5 mm. Refuted otherwise.
- H52 (the parabola is a safe stand-in). The measured-shape pipeline's clean sheets lie within 5 points of the fitted-parabola pipeline's, single 2x and doubled joints, and doubled joints give at least 0.99 on the measured studs. Refuted otherwise.
- Also reported, not scored: measured twist and in-and-out-of-plane profiles, stud spacing and lean, the inclined members' and plates' straightness, and the data set's per-board pose deviations from the nail-laminated panel experiment (an industrial robot placing 2x4s with and without its feedback loop) as a measured placement scatter beside the 0.9 mm tool scatter.
- Limits known in advance. One wall, six studs, lumber of unstated grade and species, built by robots with a feedback loop that adjusts element lengths (not by a framer), scanned from one side; the scan is stitched from passes on the robot's calibration. A sample this small can refute a shape assumption, not estimate a population.
2026-11-13 · Milestone 10 result: the tool carries half the single-2x ceiling, the pose the rest; a slide and a shim lift it to 0.92–0.94, not 95%
experiments/m10_ceiling/ (run.py, results.txt, results.json, figure.png); src/buildbot_sheathing/ceiling.py; about a minute on one CPU.
- H45 held. Tool scatter alone gives 0.9838 clean in Monte Carlo against 0.9833 from the product of the 62 nails' normal probabilities.
- H46 refuted; milestone 9's reading was half right. On straight studs the Shapley shares of −ln(clean) are tool 44.6% / 46.9% (three-foot / smooth rock), stud read 7.0 / 8.4, edge read 5.0 / 6.1, slide 17.5 / 21.7, ledger 4.1 / 5.9, panel 21.9 / 11.0: pose 43.5% / 38.6%. The more useful number: no group misses alone. Pose alone is 0.995 / 1.000 clean, tool and reads alone 0.957, tool and pose 0.805 / 0.874. The window is ±3.14 mm and a sheet off by d shrinks one side's half by d/2, so a tool tail that would land and an offset that would be harmless combine into a miss.
- H47 held. Sliding the sheet along the ledger to centre its gap on the read studs (0.5 mm residual) adds 6.1 and 5.8 points on straight studs (0.737→0.798, 0.811→0.869). A slide that leaves 1 mm is no better than the 1 mm slide already assumed.
- H48 refuted. Slide and shim (0.5 mm each) with a mid push give 0.893 at low warp (bar 0.95); 0.916 at 0.25 mm, 0.922 with dense reads. On straight studs the same remedies give 0.918 (0.938 at 0.25 mm). Once the shim plumbs the sheet, the three-foot and smooth rock models give the same result.
- What 95% takes. With slide and shim at 0.5 mm, the tool must hold 0.84 mm (straight studs) or 0.79 mm (pushed low-warp studs) 1σ, against the 0.9 mm
buildbot-fasten used; without them, 0.41–0.66 mm. So single 2x reaches 95% only with a pose fix and a tool about 10% tighter than assumed. Doubled joints, at 0.9994–1.000 with no fix, stay the recommendation. - What it means for the programme. The paper's section 5.2 should say the single-2x cap is a joint budget shared by the tool and the pose, not the pose alone. The tool's scatter is now the single input the single-2x case depends on most, and it comes from one source; a measured scatter for a nailer on an arm is the next number worth getting.
- Inputs not sourced: the slide and shim residuals (swept 0.25–1 mm); the shim as a mechanism.
2026-11-13 · Milestone 10 planned: what is the straight-stud ceiling made of, and can placement lift single 2x to 95%?
Written before any milestone-10 experiment was run.
- Why. Milestone 9 pushed the bowed joint studs straight and found single-2x joints still stop at 0.74 (three-foot rock) and 0.81 (smooth rock) clean sheets, the same as on perfectly straight studs. It read the gap as "the sheet's own pose on the ledger", but never ran the ablation that would show it. On straight studs six error sources remain: the tool's scatter (0.9 mm,
buildbot-fasten), the stud read (0.5 mm a read, three reads), the edge read (0.5 mm), the slide's translation (1 mm), the ledger's slope (laser rating plus 0.5 mm at each corner) and the panel's own geometry (PS 2 squareness and the bottom-edge rock). With edge aim each joint nail's half-window is 3.14 mm less half the sheet's local offset from where the robot meant it, so the pose and the tool compete for the same 3 mm. - Model. Milestone 5's ledger pipeline (
tack.simulate, single 2x, three reads, edge aim, 20,000 sheets, seed 7) on straight studs, with each source switchable. Shapley shares of −ln(clean) over the six sources (all 64 subsets), for both rock models. Two placement remedies, each applied after the sheet is seated and before it is tacked: slide, the robot reads both sheet edges against its stud estimates and slides the sheet along the ledger to centre the gap (translation left: the edge read over √2 plus a slide residual, swept 0.25–1 mm); shim, the robot also reads the edge's slope and lifts the low corner on a shim to plumb (tilt left: the slope of two edge reads 2.39 m apart plus the shim's own step, swept). Then both on the pushed studs of milestone 9 (mid push, pinned) at the two published warps. - H45 (closed form). Tool scatter alone (perfect pose, perfect reads, straight studs): Monte Carlo clean within 0.003 of the product of the 62 nails' normal probabilities. Refuted otherwise.
- H46 (milestone 9's reading). On straight studs, the three pose sources (translation, ledger slope, panel geometry) carry at least half of the Shapley share of −ln(clean), in both rock models. Refuted otherwise.
- H47 (slide). The slide (0.5 mm residual) raises the straight-stud ceiling by at least 5 points in both rock models. Refuted otherwise.
- H48 (95%). Slide and shim (0.5 mm residuals) with a mid push (pinned) give single-2x joints at least 0.95 clean at the low warp under the smooth rock. Refuted otherwise.
- Also reported, not scored: the largest tool scatter that meets 95% on each pipeline (bisection), the high warp, and dense stud reads.
2026-11-13 · Milestone 9 result: a push of 28–232 N straightens the joint stud, but single-2x joints still stop at the straight-stud ceiling
experiments/m9_straighten/ (run.py, results.txt, results.json, figure.png); src/buildbot_sheathing/straighten.py; 48 s on one CPU (MuJoCo 4.0.1 for the humanoid capacity).
- H41 held. The cubic-Hermite FE beam reproduces the closed forms exactly: c = 48.000 and 192.000, residual 0.07407 b and 0.29630 b at u = 1/6.
- H42 held. A mid-height push on a Stud-grade 2x4 bending flatwise between pinned plates takes a median of 28 N and a 99th percentile of 126 N at the lower published warp (53 and 232 N at the higher), and 376 N for a grade-limit stud at DF-L's 95th-percentile E. Fixed ends cost four times that. Nail by nail equals the mid push for the median stud (the mid push alone brings every point within 0.5 mm when the bow is under about 6.8 mm), but on the studs it does not, holding each nail while pushing the next makes the stud a short beam: p99 355 N at low warp and 1,258 N at high, up to 2,821 N at the grade limit, and 791 N (p99) if the tolerance is tightened to 0.25 mm. Push first, then nail.
- A G1 can just do it. Milestone 3b's static whole-body model gives a sideways push of 81–88 N at 1.2–1.3 m (right ankle roll binds, at either foot friction), an H1 120–123 N (shoulder yaw). That covers 93% of low-warp studs and 69% of high-warp ones for the G1, 99% and 86% for the H1. A G1 that cannot move a stud can read it and fall back to the extra stud.
- H43 refuted, and the reason is the useful part. Straightening lifts the best single-2x pipeline from 0.560 to 0.710 clean sheets at low warp and 0.307 to 0.677 at high (0.625→0.782 and 0.344→0.750 under milestone 6b's smooth-bow rock), 6 to 37 points, but not to 95%: on perfectly straight studs the same pipeline gives 0.737 (0.811 smooth-bow). Bow was costing single-2x joints up to 37 points; the sheet's own pose on the ledger costs the remaining 19–26. Doubled joints (0.9994 single read, 1.000 three reads) stay the only route to 95%.
- The push policy barely matters, the end restraint does. Nail-by-nail and a three-point simultaneous push land within 0.4 points of the mid push. Fixed ends lose up to 5 points at high warp, because they leave 4 times as much shape (0.296 b at u = 1/6).
- H44 held, on a low bar. Single 2x plus straightening beats doubled joints in 100% of draws at 100 and 300 houses a year; milestone 7 had already found doubled joints never pay. The informative number is the break-even: the median draw keeps favouring the push until about 925 s of pushing a sheet. Against not pushing, straightening pays in 92% of draws at high warp and only 7.5–10% at low: at low warp its rework saving (0.79 → 0.47 nails a sheet) is worth about what 10–40 s of a supervised robot costs.
- What it means for the programme. Three ways to deal with a bowed joint stud are now priced: double it ($5.42 a sheet, 95%+ clean), push it (28–232 N, free in dollars, 0.68–0.78 clean), or live with it (0.31–0.63). The recommendation stands as doubled joints where 95% first-pass is wanted; the push is the cheap fallback, and it removes bow as the cap, which turns the single-2x question into a placement question. The paper's section 5.2 can now say what the cap is made of.
- Inputs not sourced: the push tolerance (0.5 mm, swept 0.25–1), the push time (10–40 s a sheet), the end restraint (both ends bracketed), and whether a pushed stud stays put between sheets.
2026-11-13 · Milestone 9 planned: push the bowed joint stud onto its line instead of doubling it
Written before any milestone-9 experiment was run.
- Why. Milestone 2 found that bowed studs, not sensing, cap single-2x joints: the sheet edge is straight and the stud's face is not, so no placement leaves every joint nail a window. Doubling the joint stud fixes it (0.9997 clean) but milestone 7 found the stud and a framer's minute or two cost $5.42 a sheet, more than the rework it saves. Carpenters do a third thing: they push a bowed stud over to the line and nail it there. Once the first sheet's joint nails hold it, the stud stays straight for the second sheet. Nobody has asked what force that takes, what it leaves, or whether a robot doing it beats the extra stud. The 2026-11-12 "Next steps" asked how far sheathing nails pull a bowed stud straight; on their own they pull nothing (a nail goes in where the stud is), so the question becomes a push.
- Model. The stud as an Euler-Bernoulli beam bending flatwise (I = 3.5 x 1.5^3 / 12 in^4, span 92 5/8 in) between plates that hold it pinned or fixed (end nails are near pinned; fixed is the bound), its stress-free shape milestone 2's parabolic bow. E from the reference design values for Stud grade (Douglas Fir-Larch 1.4, Hem-Fir 1.2, Spruce-Pine-Fir 1.2, Southern Pine 1.3 million psi) with the coefficient of variation 0.25 that the NDS uses for visually graded lumber, lognormal. Two policies: mid, one push at mid-height until a read there is on the line through the stud's ends; nails, the robot pushes the stud onto that line at each joint nail height, bottom up, before driving the nail (the carpenter's way), each nail then holding its point. A finite-element beam (cubic Hermite elements) gives the forces and the shape left; clean sheets come from milestone 5's ledger pipeline (
tack.simulate) with the joint studs replaced by their straightened shape. - H41 (closed form). The FE mid push needs 48 E I b / L^3 (pinned) and 192 E I b / L^3 (fixed) within 1%, and leaves at most b u (1 - 2u)^2 peaking at 0.074 b at u = 1/6 (pinned) and 4 b u (1 - 2u)^2, 0.296 b (fixed), within 1% of b. Refuted otherwise.
- H42 (force). Pinned, at the lower published warp, the 99th percentile mid-push force is under 200 N; at the grade limit (3/4 in over 8 ft) with Douglas Fir-Larch's 95th-percentile E, under 400 N. The nail-by-nail push's largest force is no more than twice the mid push's in the median stud. Refuted otherwise.
- H43 (clean sheets). On milestone 5's best single-2x ledger pipeline (s_x 1 mm, three reads), nail-by-nail straightening gives at least 95% first-pass clean sheets at both published warps (low and high), within 3 points of doubled joints; the mid push alone, pinned, also reaches 95% at the low warp. Refuted on any of these.
- H44 (dollars). On the G1 on a lift base at 100 houses a year (milestone 7's draws), single 2x with nail-by-nail straightening costs less per clean sheet than doubled joints in at least 90% of draws when straightening adds 10 to 40 s a sheet (swept, unsourced). Refuted otherwise.
- Also reported, not scored: the force against what a G1 and an H1 can push sideways at the wall (milestone 3b's whole-body model) if Menagerie loads; the fixed-end and mid-push policies on the cost; what a grade-limit stud leaves.
2026-11-13 · Milestone 6b result: a smooth bow rolls back to level; the 0.149° rock was the three-foot model, and single-2x pipelines gain up to 6.9 points
experiments/m6b_bowed_edge/ (run.py, results.txt, results.json, figure.png; envelope.py and envelope.txt post hoc); MuJoCo 4.0.1 and Drake 1.51.1 SAP; 278 s and about 3 min on 4 CPUs.
- H37 held, in all 48 cells and both engines. A parabolic bow (1/32 or 1/16 in) seats inside the stick band θ_s = μ F_n r̄ / (m g (R − h0)) + 0.005°: at most 0.007° (Drake) and 0.006° (MuJoCo) at a 5 N press, 0.058° and 0.054° at 50 N. With R = 117 m and h0 = 1.22 m the edge is a stable rocker; gravity rolls the sheet back and wall friction keeps what it can.
- Over 2,000 random seats (holder error ±1°, μ 0.3–0.5, press 5–50 N, the same draws as milestone 6): 1σ 0.022° in both engines, 99th percentile |θ| 0.059° (Drake) and 0.061° (MuJoCo), against a three-foot rock of up to 0.149° on half the sheets. |θ| rises with the press (correlation +0.62, +0.59): the lighter the press once seated, the straighter the sheet.
- H40 held. With that rock in milestone 5's ledger pipelines, the best single-2x pipeline (1 mm, low warp, three reads) goes from 0.560 to 0.625 clean (+6.5 points in both engines; 0.626 with no bow rock at all), every single-2x pipeline gains 1.8–6.9 points, and doubled joints move by 0.45 points at most. The engines now agree within 0.11 points on every pipeline. Single 2x joints are still far from 95%: the bow rock was not what caps them, the bowed studs are (milestone 2).
- H38 refuted. A kink (tent, p = 1) at a 5 N press leaves 0.013–0.038° in Drake and 0.037–0.147° in MuJoCo, not ≥ 0.8 × 0.149°. The three-foot model does the same at 5 N (0.026–0.107°, 0.038–0.150°). Balanced on a point, the sheet's toppling moment m g h0 θ is 0.68 N·m at 0.149°, less than the 3.1 N·m wall friction holds at 5 N and μ 0.4, so it stops wherever the approach leaves it. That is why milestone 6's engines disagreed: the outcome on a kink is set by the approach, not by statics, and neither engine is wrong.
- H39 refuted as registered, and the post-hoc check says why. Single seats on the parabola move by up to 0.017° between 25 and 97 feet: a sheet anywhere inside the stick band is at rest, so where it stops is sensitive to everything. The envelope converges in Drake: 100% of 24 seats inside the band at every n from 25 to 401 (max 0.005–0.006° at 5 N, 0.056–0.061° at 50 N). MuJoCo's envelope at 5 N exceeds the band at every n (max 0.017–0.037°, 79–88% inside) and does not shrink with n: the soft contact, not the discretisation. At 50 N both engines are inside the band at every n.
- Shape and grit (not scored). p = 4 (flat-bottomed) leaves 0.008° at most; p = 1.5 sits between. A 1 mm bump under a straight edge (a burr, or grit on the ledger) tips the sheet onto the far corner: median 0.050° (Drake) and 0.061° (MuJoCo), up to 0.090°, which MuJoCo matches to the slope 1 mm / (W/2 + |y|). Grit costs more than the whole PS 2 bow.
- What it means for the programme. The model-dependent term milestone 6 left open is closed for a smooth bow: charge θ_s (or the simulated seats), not 2 s / W. Two placement rules come out of it: once the sheet is on the ledger, back the press off to a few newtons, and keep the ledger clean (1 mm of grit undoes it). The paper's conclusions do not change: single 2x joints stay well under 95% and doubled joints stay at 94–100%. A kink in a real edge, a chipped corner or grit is the case where the rock stays; how often real sheet edges and ledgers have one is not known (no measured edge profiles found).
- Inputs not sourced: the edge shapes (PS 2 bounds the bow, not its shape; parabola and tent are the two ends), the 1 mm bump, and the press range (milestone 3's).
2026-11-13 · Milestone 6b planned: is the bowed-edge rock a property of the sheet or of three feet?
Written before any milestone-6b experiment was run.
- Why. Milestones 3, 5 and 6 model a bottom edge bowed outward by s as three feet (the two corners and one at mid-width, s lower). On its middle foot the sheet is an inverted pendulum, so it must fall onto a corner: milestone 5 charges a rock of 2 s / W (0.149° at PS 2's 1/16 in), and milestone 6 found the two engines disagree on it by 0.02–0.03°, the size of the whole single-2x hang budget. A real bow is a smooth curve, not a point. A smooth convex edge of radius R under a body whose centre of mass sits h0 = H/2 + s above it is a rocker: rolling by θ raises the centre of mass by (R − h0) θ²/2, so the sheet is stable about level whenever R > h0. A parabolic 1/16 in bow over 48 in has R = W²/(8 s) = 117 m against h0 = 1.22 m. If so, gravity rolls the sheet back to level and only wall friction can keep a tilt: at most θ_s = μ F_n r̄ / (m g (R − h0)), with r̄ the mean distance of the wall contacts from the mid-width contact (1.56 m on corner contacts). That is 0.005–0.009° at a 5 N press and 0.054–0.090° at 50 N (μ 0.3–0.5), against the 0.149° charged. A kink (a point, a chip, grit on the ledger) has R → 0 and keeps the full rock.
- Model. The bottom edge as n feet (2 mm spheres) on the profile z = −s (1 − |2y/W|^p): p = 2 a parabola (the smooth bow), p = 1 a tent (a kink at mid-width), 1.5 and 4 in between; n = 25, 49, 97. Same sheet, holder, release and press as
ledger.seat; both engines (MuJoCo 4.0.1 and Drake 1.51.1 SAP). - H37 (a smooth bow rocks back). p = 2, s of 1/32 and 1/16 in, holder error ±0.25 and ±1°, wall friction 0.3 and 0.5, press 5, 20 and 50 N: the seated |θ| is at most θ_s + 0.005° in every cell, in both engines. Refuted on any cell.
- H38 (a kink keeps the rock). p = 1, s = 1/16 in, press 5 N: |θ| at least 0.8 × 2 s / W in every cell (the three-foot result). Refuted otherwise.
- H39 (the discretisation does not decide it). p = 2: the seated tilt moves by under 0.005° between n = 25 and n = 97 in every cell of a 12-cell subset.
- H40 (clean sheets). Charging milestone 5's ledger pipelines with the smooth-bow residual (drawn from the simulated seats, press U(5, 50) N) instead of 2 s / W raises the best single-2x pipeline (s_x 1 mm, low warp, three reads; 0.56) by at least 2 points, and moves every two-2x pipeline by under 0.5 points. Refuted otherwise.
- Also reported, not scored: p = 1.5 and 4; grit (a 1 mm sphere on the ledger at a random point under a straight edge); and whether a light press after seating (5 N) is a placement rule worth stating.
2026-11-12 · Milestone 8: the paper draft, and what writing it changed
paper/draft.md, paper/references.bib, paper/check_numbers.py (70 checks, all ok; run by tests/test_paper.py).
- The draft. One article for Automation in Construction in its structure: highlights, abstract, an error-model section, methods and results for milestones 1–7 ordered by the proposal's hypotheses, discussion, limitations, a generative-AI declaration and references. Authorship is left for the owner.
- Scoring the proposal's hypotheses for the first time. H-B, H-C and H-E held. H-D held in its first half (the base beats the walking H1 in every draw) and is refuted as registered in its second: the gantry is cheaper than the base in the median draw only from about 501 houses a year, not 30–300, though the ranking does flip inside the swept range. H-A held at a 0.5 mm stud read (no pose tolerance without edge aim), but at 0.3 mm aiming at the design window allows 0.38 mm and 0.0276°, so it depends on a stud reader's accuracy no commercial source we found states; and on an as-built wall the joint framing, not the aim, decides.
- H-B's own refutation test (a learned hang policy) was never run; the paper says it is supported by the static analysis, not tested.
- Before submission: the Elsevier
elsarticle template, the citations marked "not confirmed" (Construction Robotics 2024's authors, Lachmayer et al. 2022's title, the Adel et al. 2024 DOI, the SDPWS edition), APA E30 and SDPWS read in full, and authorship.
2026-11-12 · Milestone 6 result: on Drake the ledger still seats the sheet and clean sheets move by 0.06 points at most; the bowed-edge rock and a soft held wrist depend on the contact model
experiments/m6_drake/ (results.txt, results.json, figure.png); Drake 1.51.1 SAP against MuJoCo 4.0.1; 404 s on 4 CPUs.
- H35 (H-E) held, by a wide margin. Over 2,000 seats on the same random draws, Drake leaves a seated residual of 0.00142° (1σ; 0.0047° at most) against MuJoCo's 0.00001°. Both are far under the 0.03° a sheet may hang on single 2x joints, so the 24 ledger pipelines of milestone 5 keep their first-pass clean sheets within 0.06 points of each other (bar 10).
- H32 held. The press at which wall friction holds a 1° error is 0.965–0.998 of the closed form on corner contacts in Drake, against 0.872–0.987 in MuJoCo; the worst Drake/MuJoCo gap is 12.7% (μ 0.5). In MuJoCo wall friction holds the sheet from a little less press than rigid Coulomb friction predicts (the cautious side for a press limit); Drake is closer to the closed form. The 6 of 200 released-seat cells where the two disagree all sit at the threshold.
- H33 held. At 50 N or less and wall friction 0.3–0.5, Drake leaves 0.0025° at most (bar 0.005°).
- H34 refuted. 70% of held-wrist cells are within max(0.005°, 10%) of the closed form, bar 90%. Where the closed form keeps an error the simulators agree within 0.07°; where it predicts none (soft wrists, 300–3,000 N·m/rad) Drake keeps 0.001–0.016° and MuJoCo nothing. Small against a hang budget of 0.03°, but not zero, and a released holder (the policy milestone 5 uses) avoids it.
- H36 held. The push height that tips a sliding sheet is 0.995–1.007 of W / (2 μ_l) in Drake.
- Not scored, and the one that matters next: the bowed bottom edge. A sheet resting on its bowed middle rocks 0.107–0.123° in Drake and 0.139–0.145° in MuJoCo, both under the 2 s / W = 0.149° milestone 3 charged. On its middle foot the sheet is balanced, so how far it rocks is set by friction and the approach, not by statics: the full rock milestones 3 and 5 used is the conservative side, and the two engines disagree by 0.02–0.03°, the size of the single-2x budget. A flexible sheet (
buildbot-mjsheet) or a mock-up would settle it. - Drake's own settings (stiffness 10^5–10^7 N/m, dissipation 0.5–50 s/m, 0.5–2 ms, SAP, TAMSI, lagged) move the μ 0.4 threshold over 187.5–199.7 N and a released seat by 0.003° at most, except the softest contact (0.023°).
- What it means for the programme. The ledger results the cycle and cost milestones rest on are not a MuJoCo artefact. Milestone 8 (the paper draft) can report H-E as held, with the bowed edge as the open, model-dependent term.
2026-11-12 · Milestone 6 planned: do the MuJoCo results hold on a second contact model?
Written before any milestone-6 experiment was run.
- Why. Every result the cycle and cost milestones lean on that came out of a simulator came out of one: MuJoCo's soft, convex contact with an elliptic cone,
impratio 10 and 20 no-slip iterations (milestones 3 and 3b). Soft contact can creep under static friction, and the ledger's whole job is a friction question (does the sheet rotate onto the ledger, or does wall friction hold it at the holder's error?). H-E of the proposal asks that clean-sheet rates agree within 10 points on a second contact model. - Second model. Drake (Toyota Research Institute;
pip install drake, 1.51.1) MultibodyPlant, discrete at 1 ms, point contact with the SAP solver (Castro et al. 2022, convex and near-rigid, with a regularised Coulomb cone) and, as a third check, Drake's TAMSI solver (Castro et al. 2020, Stribeck-regularised friction). Same sheet, same contact points (spheres on the back face at the corners or the stud lines, feet on the bottom edge), same holder law, applied as an external wrench each step. Contact stiffness and dissipation are Drake's per-geometry point_contact_stiffness and hunt_crossley_dissipation, swept by a decade either side. - H32 (press threshold). The least press that leaves over half of a 1° holder error, corner contacts, wall friction 0.2–0.7: Drake SAP within 25% of the closed form m g W / (2 mu r_bar) at every friction, and within 15% of MuJoCo's. Refuted on any friction.
- H33 (a light press seats). Released at a press of 50 N or less with wall friction 0.3–0.5 and a holder error of 0.25–2°, Drake SAP leaves at most 0.005° (milestone 3's H13 on the second model). Refuted on any cell.
- H34 (held wrist). A wrist that keeps its stiffness leaves the closed-form tilt within max(0.005°, 10%) in at least 90% of milestone 3's section C cells in Drake SAP.
- H35 (H-E, clean sheets). Draw 2,000 seats per simulator with the holder error uniform in ±1°, friction uniform 0.3–0.5 and press uniform 5–50 N; add each simulator's seated residual to milestone 5's ledger tilt draw; the first-pass clean rate of every milestone-5 ledger pipeline (framing, s_x, warp, sensing policy; 24 cells, 20,000 sheets) agrees between MuJoCo and Drake within 10 points. Refuted on any cell.
- H36 (slide and tip, milestone 3b). The push height that tips a sliding sheet in Drake is within 10% of the closed form W / (2 mu_l) in each of milestone 3b's ten pushed-at-the-press cases.
- Also reported, not scored: the bowed bottom edge (section E), creep under a stuck press, the SAP against TAMSI difference, and the run time per seat.
2026-11-12 · Milestone 7 result: a G1 on a lift base is the cheapest robot below about 500 houses a year, the walking H1 never wins, and the extra joint stud does not pay in dollars
experiments/m7_cost/ (results.txt, results.json, figure.png); 4,000 draws, seed 11, common random numbers; under a second.
- H28 held. The G1 on a drive-and-lift base costs less per clean sheet than a walking H1 in every draw, at every volume and with either sensing policy forced (H1 about 2.5 times dearer: a higher price and 85–265 s of walking nails against 54–121 s).
- H29 refuted. In the median draw the gantry is dearer than the base at 10 houses a year (+$43.67 a sheet) and cheaper only from about 501 houses a year, outside the predicted 30–300 (at 300 it is still +$0.51). Per draw, 62% cross inside 10–1,000, at a median of 54 houses; the gantry's unsourced $40k–600k price spreads the crossing over two decades.
- H30 refuted, for every class. Doubled joints are cheaper in 0% of draws at 100 houses a year: the stud and a framer's minute or two cost $5.42 a sheet (median), while the rework it saves (0.79 missed nails with three reads on single 2x) is under a dollar of robot and supervisor time. Doubling pays only if a missed nail costs $6.51 [5.21, 7.88] more than its re-drive, about ten loaded person-minutes. Milestone 5's 18–65 s are real but cheap; the milestone 5 conclusion that the joint framing moves the cost is reversed when the cost is dollars rather than seconds, unless missed nails are expensive (found at inspection, or on a shear wall where a missed nail is a structural defect) or the doubled joint is required anyway.
- H31 held. Hand crew $5.21 a sheet (median; $1.93–14.46 over the taskbook's sweep of its minutes); at 100 houses a year the base is $7.03 and the H1 $17.23; the gantry is $3.93 at 300 and $33.49 at 30. Paired, the base is cheaper than the crew in 37% of draws at 100 houses a year, 74% at 300 and 94% at 1,000.
- Where the base's dollars go (two 2x, single read, 100 houses a year): ownership 44%, the stud 44%, people 11%. Its cost is moved most by its life (ρ −0.58) and its base price (+0.58), both unsourced; supervision is fourth (+0.26).
- What it means for the programme. The robot design question has a plain answer on cost (put the arm on a lift base, not on legs) and an open one on volume (a robot pays against a crew only when it sheathes about 100 or more houses a year, which a site robot moving between jobs would struggle to do). The joint-framing recommendation of milestones 2 and 5 holds for clean sheets and seconds, not for dollars, unless a missed nail is costly. The paper should report both. Milestone 6 (a second contact model) is the remaining check before the draft.
- Inputs not sourced: the base and gantry prices, every life, interest and upkeep, every stage time but the nailing; the hand crew's minutes are its author's estimate and its own misses are not charged; site robots are dedicated to wall sheathing with no transport or set-up cost.
2026-11-12 · Milestone 7 planned: dollars per clean sheet, by robot class and house volume
Written before any milestone-7 experiment was run.
- Why. Milestone 5 put every stage in seconds and found doubled joint framing buys back 18–65 s a sheet of rework. Seconds are not the decision: a builder compares dollars a sheet against a two-person crew, and the robot classes differ in price, speed and who still has to carry the sheet. H-D of the proposal (a base with a lift beats a walking humanoid; a gantry beats both above a volume) has not been tested.
- Classes. Hand crew (taskbook S1 carry 0.8, S2 hold and square 1.5, S4 nail 62 x 0.1 person-minutes a sheet, each scaled by the taskbook's one-third-to-three-times sweep; clean rate not modelled, taken as 1). Walking H1 (nailing 84.7–264.7 s, the walking-G1 range of
buildbot-mobilebase at 5–20 s a step; a person carries the sheet to the ledger, since milestone 3b found neither humanoid can carry it). G1 on a drive-and-lift base (nailing 54–121 s, buildbot-mobilebase; a person carries the sheet; the base can also hang and tack without a ledger, milestone 5's lift-tack). Gantry over a framing table in a plant (nailing 20–115 s, scan 10–60 s and load 30–180 s, buildbot-custom; the table fence plays the ledger). - Pipelines. Each class picks, per draw, the cheapest of framing {single 2x, two 2x} by sensing {single, three reads} (and for the base, ledger or lift-tack), at the lower published warp; missed nails and first-pass clean rates are milestone 5's (
m5_cycle/results.json); every missed nail is re-driven, so every sheet ends clean and the cost of a clean sheet is the cost of a sheet with its rework. - Dollars. Ownership: capital recovery of the price over its life at an interest rate plus upkeep, spread over the sheets of V houses a year (32 wall sheets a house,
buildbot-taskbook), capped at the capacity of 2,000 machine hours a year (a fleet beyond that). People: loaded wage (BLS $29.12 times a burden of 1.0–1.6, buildbot-cost) times carry seconds plus supervision (0.1–1.0 person-hour a robot-hour on site, buildbot-cost; operator share 0–0.5 on the gantry, buildbot-custom). The extra joint stud: a 2x4 92 5/8 in precut stud at $3.41–5.56 (retail listings seen through search on 2026-11-12) plus 60–120 s of a framer's time. Prices: G1 $13,500–21,600 (Unitree's own page and 2026 retailer listings seen through search), H1 $90,000–150,000 (press reports, buildbot-cost), base and lift $10,000–40,000 (not sourced), gantry $40,000–600,000 (not sourced, buildbot-custom's sweep); site robots live 2–7 years, the gantry 7–20; interest 4–10%; upkeep 5–20% (site) and 3–12% (gantry) of the price a year. Every other stage time is milestone 5's sweep. 4,000 draws, seed 11, common random numbers across classes. - H28 (H-D, first half: the base beats the walker). At every volume tried (10, 30, 100, 300, 1,000 houses a year) and with either sensing policy forced, the G1 on a base costs less per clean sheet than a walking H1 in at least 95% of draws. Refuted otherwise.
- H29 (H-D, second half: the gantry wins above a volume). In the median draw the gantry is dearer than the base at 10 houses a year and cheaper at 1,000, crossing between 30 and 300. Refuted if it never crosses inside 10–1,000 or crosses outside 30–300.
- H30 (doubled joints pay in dollars). For each robot class, at 100 houses a year, the best two-2x pipeline is cheaper than the best single-2x pipeline in at least 90% of draws once the stud and its framing time are charged. Refuted for a class below 90%.
- H31 (no site robot beats the crew). At 100 houses a year both site robots cost more per clean sheet than the hand crew in the median draw, and the gantry costs less than the crew in the median draw at 300 houses a year and more at 30. Refuted on any of the three.
- Also reported, not scored: dollars a sheet by class and volume (median and 5–95%), the share of draws in which each class is cheapest, which inputs move the base's cost most (rank correlation), and the higher warp as a check.
2026-11-12 · Milestone 5 result: doubled joints and one read a stud give a clean sheet in 76–216 s; single 2x joints cost 18–65 s more in rework, and fire-at-once recoil is not free there
experiments/m5_cycle/ (results.txt, results.json, figure.png); 20,000 sheets a cell, seed 7; under a minute.
- H22 held. On single 2x joints at the lower warp, with a ledger seat (PS 2 panels, 0.5 mm a corner, 1 mm left after the slide), first-pass clean sheets are 0.417 (single read), 0.560 (three), 0.583 (dense); 2.16, 0.79 and 0.70 nails missed a sheet. With 3 mm left after the slide: 0.25–0.31.
- H23 held. On two 2x the same pipeline gives 0.9997 with a single read (1.0000 with three); at the higher warp a single read falls to 0.938 and three reads hold 1.0000.
- H24 held. By seconds a sheet including rework, a single read is fastest on two 2x in all 108 time cells, and three reads are fastest on single 2x in all 108 (dense reads never: 53 reads a sheet cost more than the rework they save).
- H25 held. Before the extra joint stud is charged, the best two-2x pipeline beats the best single-2x one by 17.8–65.2 s a sheet; the place, ledger and nailing times cancel, so the break-even depends only on the read and rework times: 29.6–41.6 s at 30 s of rework, 17.8 s at the cheapest corner (15 s rework, 1 s reads). At the higher warp 33.8–129.2 s.
- H26 refuted. Firing at once after a 4 N s recoil, modelled conservatively as a 0.37 mm Gaussian added to the 0.9 mm tool scatter, costs up to 7.6 points of clean sheets on single 2x joints (dense, 0.583 to 0.507), though at most 0.12 points on two 2x.
buildbot-nailseq found no effect on single-2x joints with no panel gap; with the gap, bowed studs and a seated panel the single-2x window is tight enough to feel it. The residual there is a deterministic transient of at most 0.37 mm, so a Gaussian of that size overstates it. - H27 refuted. In the two-2x single-read pipeline nailing is the largest stage in 90 of 108 cells; placing the sheet (60 s, plus 10–20 s of ledger) is larger in the other 18, all with fast 54 s nailing. Shares: place 10–57%, sensing 2–17%, nailing 37–86%.
- Seconds a sheet. Two 2x with a single read: 76–216 s (median 146). The best single-2x pipeline: 94–281 s (median 183). For scale, nailing alone takes 61 s on the gantry of
buildbot-custom and 372 s by hand at the taskbook's estimated minutes. - Unplanned: on doubled joints the ledger is optional. A lift that hangs the sheet to 0.25° and 5 mm (1σ) and tacks it gives 0.978 clean with a single read and 0.996 with three; at the middle time cell that is 97 s against 106 s with the ledger, since no ledger is set. On single 2x the same hang gives 0.08–0.09.
- What it means for the programme. The design choice that moves the cost of a clean sheet is the joint framing, not the robot's sensing or its recoil handling: doubled joints buy back 18–65 s a sheet of rework and let the robot read each stud once. Milestone 7 should price that extra stud (lumber and framing minutes) against these seconds, and milestone 6 should check the clean-sheet rates on a second contact model.
- Inputs not sourced: every stage time except the nailing (itself
buildbot-mobilebase's model with swept speeds) is swept: place and slide 15–60 s, ledger 0–20 s, a read 1–3 s, rework 15–60 s a missed nail. Every missed nail is assumed found and re-driven at the same cost; a closed joint window (a bowed stud) is charged the same per nail, where a real fix may need a block or a sister stud.
2026-11-12 · Milestone 5 planned: the whole cycle, in seconds per sheet and clean sheets
Written before any milestone-5 experiment was run.
- Why. Milestones 1 to 4 each scored one stage (gap and pose, as-built studs, ledger and humanoid, tack-then-measure) on the clean-sheet rate, and
buildbot-mobilebase, buildbot-nailseq and buildbot-custom timed the nailing alone. Nobody has put the stages in one cycle and asked what a sheet costs in seconds once placement, sensing, recoil and the rework of missed nails are all charged, and which design choice moves that number. - Pipeline. Steady state along a row: a lift (or a person) stands the sheet on a laser-set ledger and the robot slides it to its stud line and presses lightly, so it seats (milestone 3); two tacks freeze it; the robot reads each new stud (three a sheet, the left joint stud is shared with the last sheet) and both sheet edges, then nails the 62 nails with edge aim (milestones 1 and 4). Clean sheets come from milestone 4's model (
tack.simulate) with milestone 3's tilt draw (laser rating, 0.5 mm a corner of ledger error, PS 2 panels uniform within the limits) and a translation left after the slide, s_x 1 and 3 mm. A second placement, lift-tack, has no ledger: the lift hangs the sheet to σθ 0.25° and σx 5 mm and tacks it. - Choices compared. Joint framing (single 2x; two 2x fastened, 3.0 in); sensing (single, three, dense reads a stud); recoil (none; fire without waiting at 4 N s, which leaves at most 0.37 mm at the next nail in
buildbot-nailseq, added in quadrature to the 0.9 mm tool scatter as a conservative Gaussian; wait, which costs 2.4 s a sheet on a fast arm and 0 on a slow one). Stud warp: the lower published means (0.11 in bow, 0.08 in twist), the higher (0.21, 0.16) as a check. - Seconds. Nailing with moves and re-stancing from
buildbot-mobilebase (54 s fast, 121 s slow, both with a 2 s settle); place and slide 15, 30, 60 s; ledger set-up 0, 10, 20 s a sheet; a read (stud or edge) 1 or 3 s including the move; rework 15, 30, 60 s per missed nail (find it, drive one beside it); the extra joint stud of doubled framing 0, 60, 120 s a sheet, with its break-even reported. None of these times is sourced except the nailing, which is itself a model with swept speeds; they are swept and the ranking is what is scored. - H22 (single 2x has no 95% pipeline). On single-2x joints at the lower warp with a ledger seat, no sensing policy reaches 95% first-pass clean sheets (milestone 2: perfect sensing 0.87 with a perfect hang). Refuted if any reaches 0.95.
- H23 (doubled joints let the robot read less). On two 2x at the lower warp with a ledger seat, a single read a stud gives at least 95% clean sheets. Refuted below 0.95.
- H24 (sensing choice follows framing). Ranking by seconds per sheet including rework, a single read is fastest on doubled joints, and three or dense reads are fastest on single 2x once rework costs 30 s or more, each in at least 90% of the swept time cells. Refuted otherwise.
- H25 (the extra stud pays). At rework 30 s, the doubled-joint pipeline's best seconds per sheet beat single 2x's best by at least 20 s before the extra stud is charged, so framing it may take 20 s or more and still win, in every place, read and nailing setting. Refuted if the break-even is under 20 s in any of them.
- H26 (recoil stays out of the budget). Firing without waiting at 4 N s costs under 2 points of clean sheets in every pipeline.
- H27 (the nailing dominates). In the doubled-joint single-read pipeline, nailing (with moves and settles) is the largest share of the sheet's seconds in every swept cell. Refuted if place, sensing or rework is larger in any.
- Also reported, not scored: the time per sheet set beside the 61 s gantry of
buildbot-custom and the taskbook's 372 s of hand nailing (both nailing only); the share of the cycle by stage; the lift-tack placement; the higher warp.
2026-11-12 · Milestone 3 (second part) result: a stock G1 or H1 can press and let the ledger seat the sheet, but neither can carry it there
experiments/m3b_humanoid/ (results.txt, results.json, figure.png); deterministic, about 6 min.
- H18 refuted in part. The G1 holds at most 2.8–8.7 kg in both hands at 0.6–1.3 m and presses 66–110 N on the ledger, as predicted. The H1 does not carry it either: 12.4–18.8 kg against the sheet's 21.8 kg (elbow torque). It presses 128–281 N. A lift or a person must stand the sheet on the ledger for either robot.
- H19 held. The push height that tips a sliding sheet is W/(2 μ_l) when the press is at the push height, 2.03 m at μ_l 0.3 to 0.87 m at 0.7; MuJoCo's threshold is 0.984–1.007 of the closed form in 14 cases, including a press fixed at 0.6 or 1.2 m (the general form).
- H20 held. At foot friction 0.6 a G1 slides the sheet at μ_l 0.3 with up to 96 N of press, and at μ_l 0.5 with at most 41 N (none above 1.0 m); the H1 manages 105–206 N and 34–141 N. On a foot friction of 0.3 the G1 slips at about 50 N.
- Unplanned: not through the palms. Sliding by palm friction needs F_n ≥ μ_l m g / (μ_h − μ_w), at least 107 N and up to 1.5 kN, and is impossible when the palm grips no better than the wall. The robot has to push the sheet's edge.
- H21 refuted. At Unitree's published gains (G1 arms kp 60 N·m/rad) the sheet's stiffness about the landing corner is 182–1,946 N·m/rad on a G1 against part one's 5,641 at 1°, 2.9 to 31 times under; H1 palm contacts 876–2,740. Only an H1 rigid clamp is too stiff (10,423–613,376; it keeps 0.46–0.99° of a 1° error), because its 4-joint arms cannot turn the forearm end about the wall normal. Releasing a G1's wrist roll changes the stiffness by 0.1–19%.
- The catch. The same softness lets the sheet's own gravity moment turn a stock hold 10–14° (G1) if nothing feeds the weight forward: the ledger carrying the weight is what makes the soft hold work, and a controller has to drop the feedforward at ledger contact.
- What it means for the programme. A stock humanoid is not the hanger; it is the presser and nailer once a lift or a person sets the sheet on a ledger. Milestone 5's cycle should carry that division of labour.
- Modelling choices, not sourced: static whole-body model, forces split 50/50 between hands, PD gains taken as the only compliance, feet rigidly planted, ledger and palm friction swept.
2026-11-12 · Milestone 3 planned (second part): can a stock G1 or H1 work a sheet the ledger carries?
Written before any part-two experiment was run. Part one found a ledger squares the sheet if the
holder goes limp in roll and presses under about 100 N. This part asks whether a Menagerie G1 or
H1 can be that holder, with the ledger (or a lift that stood the sheet on it) carrying the weight.
- Scene. The robot stands square to the wall in front of the sheet, hands at height z above the floor and reach x, feet flat (buildbot-payload's static whole-body model, buildbot-fasten's lean for a push). Three jobs: (a) hold the 21.8 kg sheet with no ledger; (b) press it into the wall with F_n once it sits on the ledger; (c) press and slide it sideways along the ledger to its stud line, a lateral force P = mu_l m g + mu_w F_n. Feet must stay inside torque limits, keep both soles loaded and not slip (rubber on concrete, mu 0.3 and 0.6, swept as in buildbot-fasten).
- Gains. Unitree's G1 arm SDK example drives arms and waist at kp = 60 N m/rad (kd 1.5), and unitree_rl_gym's G1 deploy config the legs at 100 (hips), 150 (knee), 40 (ankles); the H1 training config uses 150 shoulder, 100 elbow, 300 torso, 150/200/40 legs. The sheet's rotational stiffness about the wall normal follows from those gains through both arms' Jacobians with the feet planted, for a rigid grasp and for point contacts at the palms, rotating about the landing corner. kp is also swept.
- H18 (the ledger is what makes a G1 usable). A G1 cannot hold the 21.8 kg sheet at any hand height from 0.6 to 1.3 m within its torque limits; with the ledger carrying it, it can press at least 50 N at some height in that band. The H1 can do both.
- H19 (a high push tips the sheet). Sliding a seated sheet by pushing at height h above its bottom edge, with the press applied at the same height and wall contacts at the corners, lifts the trailing corner once h > W / (2 mu_l) (1.52 m at mu_l 0.4): the wall friction's moment cancels. Held if the MuJoCo threshold is within 10% of the closed form for mu_l 0.3 to 0.7.
- H20 (a G1 can slide only on a slippery ledger). At F_n = 50 N and mu_w 0.4, a G1 can apply the slide force at mu_l = 0.3 (84 N) at some height, and not at mu_l = 0.5 (127 N); the H1 can at both.
- H21 (stock gains are too stiff to let the ledger seat it). At the sourced gains the sheet's rotational stiffness about the landing corner exceeds part one's limit (m g W/2 - mu F_n r_bar) / theta0 at theta0 = 1 deg for both robots and both grasps, so a stock position hold keeps part of its error; releasing the wrist roll alone does not bring it under, because the hands translate about the corner. Also reported: the kp at which it would.
2026-11-12 · Milestone 3 (first part) result: let go and press lightly, and a ledger squares the sheet; then the panel's squareness is what single 2x joints cannot absorb
experiments/m3_ledger/ (results.txt, results.json, figure.png); MuJoCo seating runs are deterministic, Monte Carlo 20,000 sheets a cell, seed 5; about 95 s.
- H13 held. Released with a press of 50 N or less at μ 0.3–0.5, the sheet ends within 0.00003° of the ledger from θ0 up to 2°. In fact every cell up to 100 N, μ 0.2–0.7, seats; descent speed, release time, step size and contact layout change it by under 0.0001°.
- H14 held. The press at which wall friction holds the error is 380, 253, 189, 143, 115 N at μ 0.2 to 0.7 (corner contacts), 0.87–0.99 of μ F_n r̄ = m g W/2; with contacts along four stud lines 0.90–1.00. The transition is sharp: below it the sheet seats, above it the sheet keeps 99.6–100% of its error.
- H15 refuted, narrowly. The held-wrist closed form max(0, θ0 − (m g W/2 − μ F_n r̄)/k) matches 29 of 30 cells within 0.005° or 10%; at 100 N, 0.5° and 10,000 N·m/rad MuJoCo leaves 0.089° against 0.119°, where wrist and friction share the moment and friction is not fully mobilised. The practical reading stands: a wrist stiffer than about 130 N·m / θ0 (7,500 N·m/rad at 1°) keeps part of its own error.
- H16 held. With a laser-set ledger (0.5 mm a corner) and panels uniform up to the PS 2 limits, the tilt σ is 0.085°, 6 times inside the doubled-joint target: 1.0000 clean on two 2x at the lower warp with σx = 3 mm, in every ledger and panel cell up to 2 mm a corner. Single 2x straight, σx 0: 0.844.
- H17 held. With a perfect ledger and seat, panels at the PS 2 limits give 0.857 on single 2x; 95% needs panels within 0.244 of the limits (0.020° out of square, 1/64 in of edge bow). Even perfect panels give only 0.953, the laser's 0.020° rating included.
- Unplanned: a bowed bottom edge rocks. An outward bow of 1/16 in (PS 2's limit) puts the sheet on its middle and one corner, 0.139–0.145° in MuJoCo against 0.149° for a rigid edge; which corner follows the side the robot held it tilted to.
- What it means for the programme. H-B now has a concrete form: a humanoid does not need to hang a sheet plumb if it can set it on a ledger, go limp in roll and press under about 100 N. Whether a G1 or H1 can do that while carrying 22 kg is the second part of milestone 3. On single 2x joints the floor is the panel, not the robot: the shear-wall case for doubled joint studs (milestone 2) gets stronger.
- Modelling choices, not sourced: rigid sheet on point contacts; lumped impedance holder; uniform distributions within PS 2 limits; ledger height error swept; wall friction swept around a secondary-source range.
2026-11-12 · Milestone 3 planned (first part): does a ledger board square the sheet, and what does it leave?
Written before any milestone-3 experiment was run. Milestone 4 set the target: a hang within
about 0.5° and 10 mm (1σ) on doubled joints, 0.03° and 0.4 mm on single 2x. Translation is easy
(the robot sets x from its stud reads); rotation is the hard part. This first part of milestone 3
asks whether a temporary ledger board under the sheet, the guide H-B names, can set the rotation
for a robot that cannot, and what the ledger, the panel and the robot's wrist leave behind. The
humanoid whole-body hang (Menagerie G1/H1) is left for the second part.
- Scene (
src/buildbot_sheathing/ledger.py, MuJoCo). A rigid 4x8 sheet of 7/16 in OSB (1.219 x 2.438 x 0.011 m, 21.8 kg, buildbot-payload's figure) pressed against a wall plane by a holder that pushes at the sheet's centre with force F_n, and lowered onto a ledger board. The holder is an in-plane impedance on (y, z, θ) aimed at its believed-plumb pose, which is off by θ0. Once a corner touches, the holder ramps its vertical support and translation gains to zero over 0.5 s; its rotational stiffness k_θ about θ0 is either released too ("release") or kept ("hold", swept 0 to 10⁴ N·m/rad). Swept: wall friction μ 0.2 to 0.7 (wood on wood, 0.25 to 0.5 in a secondary table; the Wood Handbook's own figure was not found), F_n 0 to 600 N, θ0 0.25 to 2°, descent speed. The tilt left is read once the sheet has settled. - Panel and ledger. PS 2-18 §5.2.1.3 (opened this session): panels square within 1/64 in per foot of the longest edge measured along the diagonals (1/8 in on an 8 ft panel, a corner angle of 0.083°), and edges within 1/16 in of the corner-to-corner line. A bottom edge bowed out by s rests on its middle and one corner, tilted 2s/W (up to 0.15°); bowed in, it rests on both corners. The ledger is set to a laser line; pro cross-line lasers are rated 1/8 in at 30 ft (0.020°; Fluke PLS 180, Stabila LAX 320 and others, maker listings seen in search), and the board is fixed to it with a height error at each sheet corner (σ swept 0 to 2 mm, not sourced).
- H13 (a released sheet seats). Released, with a light press (F_n ≤ 50 N) at μ 0.3 to 0.5, the sheet's bottom edge ends parallel to the ledger within 0.005° from any θ0 up to 2°.
- H14 (a hard press freezes the hang). Wall friction holds the tilt once μ F_n r̄ exceeds the gravity moment about the landing corner, m g W/2 = 130 N·m. With MuJoCo's four corner contacts on the wall, r̄ = (W + H + √(W² + H²))/4 = 1.596 m, so F_n* = 81.7/μ N (204 N at μ = 0.4). Held if the simulated threshold (the least press that leaves over half of θ0) is within 25%.
- H15 (a stiff wrist fights the ledger). Held, the residual tilt is max(0, θ0 − (m g W/2 − μ F_n r̄)/k_θ): a wrist stiffer than about 7,500 N·m/rad at θ0 = 1° with no press keeps part of its own error. Held if MuJoCo matches within 0.005° or 10%.
- H16 (the ledger meets the doubled-joint target, not the single). A released seat on a laser-set ledger (σ 0.5 mm a corner), with squareness and bottom-edge bow drawn uniformly up to the PS 2 limits, gives ≥ 95% clean sheets on two-2x joints at the lower published warp with σx = 3 mm, and < 95% on straight single-2x joints even with σx = 0 (milestone 4's model).
- H17 (the panel, not the ledger, sets the single-2x floor). With a perfect ledger and a perfect seat, panels drawn uniformly up to the PS 2 limits give under 90% clean sheets on straight single-2x joints. Also reported: the fraction of the PS 2 limits a panel would have to meet for 95%.
2026-11-12 · Milestone 4 result: on doubled joints a robot may hang the sheet 18x coarser; on single 2x joints re-hanging cannot reach 95%
experiments/m4_tack_measure/ (results.txt, results.json, figure.png); 20,000 sheets a cell, seed 4; about 100 s.
- H10 refuted. On straight single-2x studs with a hang of σx = 1 mm, σθ = 0.1°, tack-measure-reject peaks at 0.87 clean sheets (re-hang below a 2.0 mm predicted half-window, 2.1 hangs a sheet). A stricter threshold re-hangs more and does worse (0.81 at 2.5 mm, 3.6 hangs), because the fifth hang is nailed whatever it predicts. At the lower published warp the best is 0.66. Even a perfect hang gives only 0.96 (straight) and 0.81 (low warp): the narrow joint has no room left for the hang once sensing and tool scatter are in.
- H11 held, by a wide margin. On two-2x joints at the lower published warp, tacking and aiming from the measured edge allows σθ = 0.54° (11 mm at the top nail) and σx = 10.4 mm for 95% clean sheets, against 0.029° and 0.37 mm on straight single-2x joints: 18 and 28 times coarser. A 3x allows 0.38° and 7.2 mm. Studs that bow cost almost nothing on wide joints (0.556° straight against 0.537° low warp).
- H12 held. A held sheet that sways under the nail after the edge was measured may sway 0.19 mm (1σ per nail) on straight single-2x joints and nothing at the lower warp, against 4.1 mm on two-2x joints and 3.0 mm on a 3x.
- Edge aim is worth having only when the hang is coarse. With three reads per stud, aiming at the design window beats aiming from the measured edge on single-2x joints (0.972 against 0.958 with a perfect hang; 0.44 mm against 0.37 mm of translation allowed), because the 0.5 mm edge reading adds noise the hang does not. On wide joints edge aim nearly doubles the hang allowed (0.29° to 0.54°).
- Re-hanging is cheap insurance on wide joints. A 3x with a hang of 3 mm and 0.3° goes from 0.967 to 0.9996 clean with 1.1 hangs a sheet.
- What it means for the programme. Milestone 3 now has its target: a hang within about 0.5° and 10 mm (1σ) on doubled joints, or 0.03° and 0.4 mm on single 2x. The first is a placement a guided humanoid or a lift might reach; the second is not plausible for any site robot co-holding a sheet. Tacking matters less than the framing: on wide joints a robot can hold the sheet and nail with millimetres of sway.
- Modelling choices, not sourced: hang and sway errors Gaussian and swept; sway is an independent translation per nail with no rotation; a re-hang is a fresh independent hang (a robot that corrects from its measurement would do better; table A's budget is that residual); the measured edge line has the true slope. Studs, sensing and tool are milestone 2's.
2026-11-12 · Milestone 4 planned: tack, then measure; how coarse may a robot hang the sheet?
Written before any milestone-4 experiment was run. Milestone 3 (placement in MuJoCo) needs a
target to score a G1, H1, lift or ledger against; this milestone writes that target, so it runs
first.
- Why. Milestone 1 gave the sheet pose a robot may leave as about 0.4 mm and 0.03° (1σ) on single-2x joints, even aiming from the measured edge, and milestone 2 found the joint framing, not sensing, sets the window once studs bow. Two things were left out. (1) A robot that holds the sheet while it nails sees it move between the edge measurement and each nail (sway, recoil). Two tack nails in the field studs freeze the pose, and the edge is measured after. (2) A robot that measures the edge can see a bad hang and re-hang before it drives the joints (pull the tacks, or nudge a held sheet: the pose-based correction of Adel et al. 2024).
- Model (
src/buildbot_sheathing/tack.py, on milestone 2's studs: bow, twist, three reads at 0.5 mm, edge line 0.5 mm, tool 0.9 mm, 1/8 in gap centred on each joint stud's estimated range). The sheet is hung off by a translation x and a rotation θ about its centre, both Gaussian, both swept. Hold: each nail sees the hung pose plus an independent sway s (swept); the edge is measured once. Tack: two field tacks freeze the pose (any shift while tacking folds into x), then the edge is measured. Reject: after measuring, the robot predicts each joint nail's window from its measured edge and stud estimate and re-hangs if the narrowest half-window is under a threshold (swept 1 to 2 mm), up to five hangs. Joint framing: single 2x, 3x, two 2x. - H10 (re-hanging rescues single-2x joints). On straight single-2x studs, with a hang of σx = 1 mm and σθ = 0.1°, tack-measure-reject reaches ≥ 95% clean sheets with ≤ 2 hangs a sheet on average, at some threshold. Refuted otherwise.
- H11 (wide joints relax the hang). On two-2x joints at the lower published warp (three reads), tack-then-measure with no re-hang allows σθ ≥ 0.2° (with x = 0) and σx ≥ 4 mm (θ = 0) for 95% clean sheets. Refuted if either is smaller.
- H12 (holding costs little on wide joints, much on narrow). With a perfect hang, the per-nail sway allowed for 95% clean sheets is under 0.5 mm on straight single-2x joints and at least 2 mm on two-2x joints at the lower published warp.
- Also reported, not scored: a grid of σx by σθ on each framing; the hang a 95% rate allows written as millimetres at the top of the sheet; the mean hangs a sheet under each threshold.
2026-11-12 · Milestone 2 result: a bowed 2x closes the joint whatever the robot sees; a 3x or doubled joint stud with three reads gives 100% clean sheets
experiments/m2_asbuilt_studs/ (results.txt, results.json, figure.png); 20,000 sheets a cell, seed 3; about 40 s.
- H6 held. At the Stud-grade bow limit (3/4 in over 8 ft, 17.7 mm on the stud) no sheet is feasible with the gap centred on the stud's range: the arithmetic limit is 4h = 12.5 mm (0.53 in over 8 ft). Single and two-read policies pass 25% of sheets, but only by putting all the bow on one side, which closes the neighbour's.
- H7 held. At the lower published warp (bow 0.11, twist 0.08 in mean, half-normal) a single mid-height read gives 0.55 clean sheets (1.7 nails missed a sheet).
- H8 refuted. Three reads and a fitted parabola give 0.80, a read at every joint nail 0.83, and perfect stud and edge knowledge 0.87. The ceiling is the window the bow leaves, against 0.9 mm of tool scatter, not the sensing: no policy reaches 95% at any read scatter (table E).
- H9 held. Twist costs 1.4 points under three reads (0.818 bow only, 0.803 both); twist alone 1 point. The band shrinks on the side the nail runs toward: 0.44 mm at 1°, 1.35 mm at 3°.
- Unplanned, and the useful result: widen the joint, don't sense harder. A 3x joint stud (2.5 in face) or two 2x fastened together (3.0 in), which SDPWS 4.3.7.1 names for adjoining panel edges of some shear walls, widens the window by 1/2 or 3/4 in a side; the largest closable bow becomes 37.9 or 50.6 mm, beyond the grade limit, and three or dense reads give 1.0000 clean sheets at both warp levels and both distribution shapes. Even a single read gives 0.85–1.00. Section H was added after C–F showed no sensing policy could reach 95% on single 2x joints.
- Culling is the expensive alternative. On single 2x joints a framer would reject studs bowed over 1/16 in (65% of them at the low warp) to bring dense reads to 0.96 and three reads to 0.94.
- Lean adds to bow. Plate layout error of 1.6 mm at each stud end takes three-read from 0.80 to 0.72.
- What it means for the programme. The robot design question (H-B, H-D) should be asked on doubled or 3x joint framing, where the sheet's pose and the tool scatter set the budget again, or the robot must straighten studs. A panel line that frames its own walls can choose that cheaply; a site robot sheathing someone else's framing cannot.
- Inputs not sourced or seen only in snippets: warp limits and means come from search snippets of USDA and Wood and Fiber reports (southern and ponderosa pine), not opened; the distribution shape is assumed (two shown); the 10D penetration is from memory; plate layout error is swept. The model takes the free-piece warp as installed; nailing may pull bowed studs partly straight.
2026-11-12 · Milestone 2 planned: real studs bow, twist and lean, and the sheet edge is straight
Written before any milestone-2 experiment was run.
- Why. Milestone 1 kept every stud straight and plumb. Stud-grade 2x4s are not: the grading limits for an 8 ft Stud-grade 2x4 are crook 1/4 in, bow 3/4 in and twist 3/8 in (USDA Forest Service reports quoting the 1978 NHPMA rules and Shelly, Arganbright and Birnbach 1979, Wood and Fiber 11(1):50–56, seen through search, not opened; one report gives bow 1/2 in), and kiln-dried southern pine studs averaged crook, bow and twist of 0.06, 0.11, 0.08 in in one Koch FPJ note and 0.13, 0.21, 0.16 in in another (Koch 1974, serrated sticks), also seen through search. In a wall the 3.5 in depth is normal to the sheathing, so bow (flatwise) moves the nailing face along the wall, twist rotates it, and crook moves it out of plane (not modelled here). A sheet edge is straight, so a bowed joint stud moves under it; no sensing can widen a window the bow has closed.
- Model. Each stud's face line is a parabola of bow b (sign random, zero at the plates) plus a linear twist φ(z) = φ_T (z/H − 1/2) with φ_T = t / 3.5 in, scaled from 8 ft to the 92 5/8 in precut stud (bow by length², twist by length). Magnitudes are half-normal or exponential with the sourced mean, truncated at the grade limit. A joint nail must have its shank on the stud and at least P = 10 nail diameters (1.31 in, the NDS full-value penetration, from memory, not checked this session; 6D swept) of wood along its path, through the twisted section, within the 2 1/16 in an 8d nail has past 7/16 in sheathing. Sheets are straight and plumb, edges set gap/2 either side of a reference on each joint stud (panels cut to fit; milestone 1's width question is left aside). The robot reads each stud with 0.5 mm scatter per read and aims from the measured edge (0.5 mm) and its estimate of the stud at each nail height, with 0.9 mm tool scatter.
- Sensing policies. single: one read at mid-height, stud assumed straight (studfind). ends: reads near the top and bottom, a straight line. three: bottom, middle, top, a parabola. dense: a read at each of the 17 joint heights, least-squares parabola. Every fitted policy centres the gap on the middle of its estimated range (minimax).
- H6 (a grade-limit bow closes the joint). A Stud-grade stud at the 3/4 in bow limit leaves no window on one side of the joint under any straight-sheet placement. Arithmetic before the run: minimax centring needs b ≤ 4h = 12.6 mm, and the scaled limit is about 17.7 mm. Held if the Monte Carlo finds zero feasible joints at the limit.
- H7 (a single read fails on typical studs). At the low published mean bow (0.11 in) and twist (0.08 in), half-normal, a single mid-height read gives under 80% clean sheets. Refuted at ≥ 80%.
- H8 (three reads suffice). Reading each stud at three heights and fitting a parabola gives ≥ 95% clean sheets at the same warp. Refuted below 95%.
- H9 (twist is second order). At that warp under the three-read policy, removing twist raises clean sheets by under 3 points. Refuted at ≥ 3.
- Also reported, not scored: the window against twist angle; clean sheets at the higher published means (0.21 in bow, 0.16 in twist) and with exponential tails; the share of studs a framer would have to cull (bow over a threshold) for a single read to reach 95%; stud read scatter allowed for 95% by policy; plate layout lean swept 0 to 1.6 mm.
2026-11-11 · Milestone 1 result: the 1/8 in gap cuts the stud budget from 0.90 to 0.41 mm, sheets must hang plumb to ~0.03°, butted rows leave the studs by sheet 3–5
experiments/m1_gap_placement/ (results.txt, results.json, figure.png); 20,000 sheets a cell, seed 1; rows 4,000 trials, seed 2; about 12 s.
- H1 held. With APA's 1/8 in gap centred on the joint stud, the joint window shrinks from ±3.93 to ±3.14 mm, and the single-read stud scatter allowed for 95% clean sheets falls from 0.90 mm (no gap;
buildbot-studfind gave 0.89) to 0.41 mm; the tip scatter allowed with perfect studs falls from 1.23 to 0.98 mm. At studfind's own 0.89 mm budget, clean sheets drop from 0.95 to 0.75 when the robot plans for the gap, and to 0.56 (1.1 nails missed a sheet) when it aims for no gap and the sheet is hung with one. - H2 held, trivially at the pre-registered setting. With stud 0.5 mm and tip 0.9 mm a robot aiming at the design window is already at 0.93 clean sheets with a perfectly hung sheet, so it has no pose tolerance at all. Unplanned, at stud 0.3 mm: 0.028° of rotation (1σ; 0.58 mm at the top nail) and 0.38 mm of translation.
- H3 held at the pre-registered setting, refuted at a tighter one. Aiming from the measured edge (edge line good to 0.5 mm) allows 0.44 mm of translation and 0.031° of rotation where nominal aim allows none (ratio unbounded). At stud 0.3 mm (unplanned) the ratio is 0.61/0.38 = 1.6, under the 2 the hypothesis asked for. With the edge line good to only 1 mm, edge aim allows nothing either.
- Unplanned: edge aim halves the weight of the stud error. The aim point sits midway between the measured edge (+3/8 in) and the estimated far stud face, so a stud error moves it by half. With no pose error, clean sheets rise from 0.93 to 0.96 at stud 0.5 mm.
- H4 partly refuted, held in outcome. Full 48 in panels butted with a 1/8 in gauge drift 3.2 mm a sheet; the third sheet's joint window exists in 46% of trials, not 0% as predicted (6.35 mm of drift against a 6.27 mm window, so gap and width noise open it), but no third sheet is clean (sheet 2: 0.04 nominal, 0.22 edge aim). Panels trimmed 1/16 in keep a joint window in at least half the trials for 4 sheets (sheet 5: 47%): held. Even panels trimmed by the full 1/8 in lose clean sheets along the row (0.95 to 0.66 by sheet 10) because the gauge's 0.5 mm scatter accumulates as a random walk.
- H5 refuted. Setting each sheet from its own studs keeps every joint window open and every sheet at 0.98 clean (1/16 in trim, 0.5 mm width scatter, 0.1% of joints overlapping), but at the pre-registered 1 mm width scatter 5.6% of joints would need a negative gap: a 1/16 in trim leaves a 1.6 mm mean gap, too little for 1 mm of width scatter. A robot setting sheets from the studs must either trim the panel or accept an edge off-centre by up to a millimetre or two there.
- What it means for the programme. The joint window, not the stud reading alone, is the binding constraint once the gap is in, and a robot must see the sheet edge and set each sheet from its own studs. The rotation budget (~0.03°) is two orders of magnitude below the 2–6° tilt
buildbot-cohold saw during a co-held lift, which makes H-B (placement is the humanoid's bottleneck) the next thing to test, in MuJoCo (milestone 3). - Inputs not sourced, swept or fixed by choice: tip scatter (0.9 mm, a
buildbot-fasten model result), stud scatter, edge-line error, gauge scatter (0.5 mm) and panel width scatter (0.5 and 1 mm). The 1/8 in gap, 3/8 in edge distance and the "sized for spacing" trim come from APA pages seen through search, not opened.
2026-11-11 · Milestone 1 planned: the panel gap and the sheet's own pose in the joint-nail budget
Written before any milestone-1 experiment was run (the model and its tests, src/ and tests/, were written first; the tests only check it against buildbot-studfind and closed forms).
- Why.
buildbot-fasten and buildbot-studfind put each sheet edge exactly on its joint stud's centre line. On a real wall the panels are hung with APA's recommended 1/8 in gap, centred on the joint stud, so each edge sits 1/16 in past the centre and the 3/8 in edge-distance rule moves with it. The sheet is also hung with a translation and a rotation error, and its width differs from the 48 in stud module. None of that has been in the budget. - Fixed inputs. Robot tip scatter 0.9 mm per nail (
buildbot-fasten); stud estimate scatter a shared error per stud, single read (buildbot-studfind's policy); 8d common nail, 1.5 in face, 3/8 in edge distance, 1/8 in gap (APA builder tips, via search). - H1 (the gap alone tightens the stud budget). Perfect placement, nominal aim: the single-read stud scatter allowed for 95% clean sheets falls below 0.70 mm (from
buildbot-studfind's 0.89 mm with no gap). Refuted if ≥ 0.70 mm. - H2 (a nominal-aim robot needs a near-perfectly plumb sheet). With stud scatter 0.5 mm and tip scatter 0.9 mm, the sheet rotation (1σ) allowed for 95% clean sheets is under 0.05°. Refuted if ≥ 0.05°.
- H3 (seeing the edge buys placement tolerance). Aiming each joint nail from the measured sheet edge (edge-line error 0.5 mm) at least doubles the sheet translation (1σ) allowed for 95%, against aiming at the design window. Refuted if the ratio is under 2.
- H4 (sheet-referenced rows drift off the studs within a wall). Butting each sheet against the last with a 1/8 in gap: with full 48 in panels the third sheet's joint window is empty (arithmetic: 3.2 mm of drift a sheet against a 6.3 mm window); with panels trimmed by the typical 1/16 in ("sized for spacing"), no more than 5 sheets in a row keep every joint window non-empty. The Monte Carlo adds gap and width noise; refuted if a 1/16 in row stays feasible (≥ 50% of trials) past sheet 5.
- H5 (stud-referenced placement fixes the drift). Setting each sheet from its own studs keeps every joint feasible, and with 1/16 in trimmed panels and a width scatter of 1 mm (1σ) the realised gap is positive at ≥ 99% of joints. Refuted below 99%.
- Also reported, not scored: clean-sheet rate along a 10-sheet row for each placement and aim; nails missed per sheet; the allowed stud scatter as a function of gap.
Produced with AI assistance (Claude). Results are from the code in this repository and are simulated or stylised unless the abstract says otherwise; nothing here implies affiliation with or endorsement by any lab, researcher or company cited. MIT licensed.