Open-source purpose-built robots

A stdlib model compares an XY gantry with an equal-link planar arm over a square work area, on cost per good pick per hour. At 2 mm tolerance the gantry wins up to 0.35 m and above about 1.0 m; the arm wins between 0.36 and 0.99 m. Tolerance below 1 mm removes the arm’s band, and the band depends heavily on the assumed cost coefficients (halving the gantry’s removes it). Stylised: every cost, mass, speed and accuracy figure is assumed, and only one task is modelled.

A stdlib planar two-link arm model searches the split of a fixed 0.50 m reach together with PD gains, against an equal-link body with tuned gains, at equal search budget. The tray is reachable only for splits of 0.40 to 0.60. Co-design gained 1.9% in held-out move time (0.283 against 0.289 s), within three-seed noise, while tuning the controller cut hand-set gains' 0.460 s by about 37%. Stylised, preliminary and mostly negative: one body parameter and an assumed torque limit.

A stdlib deck-and-pipette twin tracks tips, well volumes, liquid labels and straight-line travel over labware, and checks protocols before they run. On 522 injected faults across three reference protocols it flags 366 (355 with the intended code) where a syntax-only baseline flags 100; the misses are mostly equivalent mutants. Collisions appear only where a move crosses taller labware (a tube rack in one protocol; forced travel heights up to 70 mm collide, 90 mm and above do not). With an assumed pipetting-error curve, a 12-well two-fold serial dilution has a final-well CV of 0.187 at 10 µL transfers, 0.100 at 20 and 0.035 at 100 µL. Stylised and circular by construction: the faults match the checks, and every geometry, timing and error value is assumed.

The 522 seeded faults of osrobot-lab (three protocols, 11 fault kinds) were translated into Opentrons Python API scripts and run through the real simulator (opentrons==8.0.0, OT-2, p300). It raises on 270 of them against the twin's 366, and every one it catches the twin also catches. It catches scripting errors (unknown labware or well, missing or doubled tip pick-up, 350 µL or 5000 µL aspirate) and accepts tip reuse, a 5 µL dispense, a tip left on and a half-empty source, even with starting liquids declared. Those are mostly policy differences, not simulator misses. Stylised: the fault set was written for the twin.

A stdlib model searches link lengths, joint count and mount position of a planar arm, with its inverse-kinematics controller tuned in the loop, against a fixed three-link body at an equal cost budget (300 evaluations, 10 seeds, held-out targets). On a shelf task that needs the tool pointing down, the search reaches 0.95 success against 0.26 for the fixed body with a tuned controller and 0.50 with the mount also tuned; on a long position-only reach the fixed body already succeeds and the search adds nothing. Random search beat a (1+4) evolution strategy here, weakly. Stylised: kinematics and static torque only (no MuJoCo dynamics), and every cost, mass and limit is assumed.

A stdlib mass-closure sizing model closes a quadrotor, a fixed wing and a fixed wing with a vertical-lift system for the same mission, choosing rotor diameter or wing area for least mass. For 3 km out, 20 min on station and a 0.5 kg payload the plain fixed wing is lightest (1.27 kg), against 2.14 kg for the quadrotor and 2.08 kg for the VTOL wing. The quadrotor does not close at 45 min on station; the VTOL wing is lighter than it from 30 min or beyond about 10 km. Stylised: every technology constant is assumed, no flight stack was run, and a fixed wing cannot hover or launch unaided.

A stdlib extension of the osrobot-uav mass-closure sizing adds a steady wind: headwind out, tailwind back, airspeed re-optimised, and station-keeping at airspeed equal to the wind. For 3 km out, 20 min on station and 0.5 kg the plain fixed wing stays lightest at every wind from 0 to 18 m/s (1.27 kg at 0, 1.30 at 10, 1.50 at 18); the quadrotor goes 2.14 → 2.33 → 2.86 → 7.0 kg at 0, 10, 15 and 18 m/s and the VTOL wing 2.08 → 2.54. The multirotor’s one win in still air (1 min on station) lasts to about 8 m/s. Designs sized for still air keep 17.2 (fixed wing) and 18.5 min (multirotor) of 20 at 9 m/s and 8.8 and 13.8 at 15 m/s; building for 15 m/s costs 34% (multirotor) and 9% (fixed wing) in mass. Crediting the multirotor with translational lift makes its mass fall to 1.71 kg at 10 m/s, an optimistic artefact of a steady known wind. Stylised: assumed coefficients, no gusts, crosswind or launch limits, nothing flown.

A stdlib extension of osrobot-uavwind draws one along-track wind per mission from a normal distribution (exact discretisation, no sampling noise) and sizes the airframe for a quantile of it. For a 3 km, 20 min, 0.5 kg mission with wind ~ N(8, 3) m/s, designing to the 95th-percentile wind (12.9 m/s) costs the fixed wing +3.7% mass (1.333 against 1.286 kg) and the multirotor +11.7% (2.518 against 2.254 kg). The standard 20% reserve, if it may be spent, already lifts a mean-wind design to 0.95 (fixed wing) and 0.99 (multirotor) success, and a larger reserve and a quantile design cost within a few percent of each other. At a 1 min stay the 95th-percentile design flips the lightest airframe from multirotor to fixed wing. Stylised: one steady wind per mission, assumed constants, no turbulence, no launch limit, and no flight.

A stdlib extension of osrobot-uavgust sizes the airframe for a launch limit wind and launches only if the forecast of the wind is at most that limit, with a normal forecast error (exact discretisation, no sampling noise). With a perfect forecast every failure becomes a scrub (fixed wing, wind ~ N(5, 4) m/s: 48% scrubbed at the median limit, 4.6% at the 95th percentile, 1.274 to 1.313 kg). A 2 m/s forecast error brings the failure rate to 6.8% at the median limit but only 1.1% at the 95th-percentile limit. Stylised: one steady wind per mission, assumed constants, normal forecast error, no cost put on a scrub.

A stdlib extension of osrobot-uavnogo lets the crew look again at the forecast in up to 8 slots a day, with wind following a Markov chain of persistence ρ (exact probabilities, no sampling noise). Fixed wing, wind ~ |N(5, 4)| m/s, limit at the median wind, perfect forecast: one slot completes 52.0% of days, and eight slots complete 99.7% with independent wind but 97.0%, 87.2% and 71.5% at ρ = 0.5, 0.8 and 0.95. With a 2 m/s forecast error each retry is another draw of the error, so eight slots raise the failure share from 6.8% to 13.6% of days (23.7% at ρ = 0.95). Stylised: assumed constants, one steady wind per slot, forecast errors independent between slots, no abort, no value on a mission. Nothing was flown.

A stdlib extension of osrobot-uavnogo puts a money value on a completed observation, the loss of an aircraft on a failure and a scrubbed day, then searches the design wind limit L and the launch threshold T together (exact probabilities, no sampling noise). With a cheap airframe ($1,500/kg, assumed) the best design is for nearly every wind (L at 14 m/s or more in 17 of 18 cases, 11 on the grid edge) and the rule barely matters. With an expensive one ($30,000/kg, assumed) the optimum is interior, and with a 2 m/s forecast error the best T sits 1 to 6 m/s below L in all 9 cases, worth up to $284 per attempt over tying T to L, with failures at or below 0.1%. Money figures are assumptions, many optima sit on the grid edge, and a failure is assumed to lose the whole airframe. Stylised, no aircraft designed or flown.

A stdlib extension of osrobot-specialise equips a shop of 2 to 8 pick-and-place trays three ways: a robot per job, one shared gantry spanning all trays, or one shared arm, on capital cost over 200 random job mixes per row. Sharing wins only while the total load fits one machine (4 jobs at scatter 0.4 m: shared cheapest in every mix at demand 0.2 of a gantry per job, a robot per job from 0.3) and only while changeover is short (a robot per job in 99% of mixes at 1,800 s). Trays spread beyond about 0.6 m erode the saving. The shared arm can hold 1 to 2 mm in only 13% of default mixes, but is cheapest in about half of mixes with loose (5 and 10 mm) tolerances. Stylised: every cost, speed, demand and changeover figure is assumed.

A stdlib extension of osrobot-reconfig adds independent unit failures (availability A) and buys the fewest spares that meet a service target, either per job or for every job at once (200 random job mixes per row, seed 7). While the shared robot's load fits one machine, spares change nothing in the per-job comparison: the shared fleet costs 0.68 of a robot per job at 4 jobs, A = 0.95, target 0.99, with or without spares. Asking every job to be up at once favours sharing (0.34 to 0.51 at 4 jobs, A 0.90 to 0.99; 0.41 at 8 jobs, A = 0.95), because a shared robot fails every job together while four independent jobs each need a spare. At demand 0.3 per job spares close the gap load opened (1.34 bare, 1.01 with spares). Limits: “served in full” is all or nothing, failures are independent with unlimited repair crews, and one assumed availability covers every machine. Stylised; no price or MTBF is sourced.

A stdlib planar point-mass twin of a feeding arm and a swaying mouth, with 200 paired trials per row. With a 0.05 s forward lead the push on a blocked tip is 20 N·s/m × speed, so peak lip force tracks approach speed (1.0 N at 0.05 m/s, 4.0 N at 0.20). A camera lowers the share of trials that touch the lips (0.40 against 0.72 at 30 mm sway) without lowering the peak, and its advantage fades with latency (touch share 0.07 at no delay, 0.71 at 0.4 s). A 1 N force cap trims the worst peak but not the share of comfortable feeds. Stylised: all constants are assumed, the peak-force result is largely an identity of the chosen controller, and no human model or Assistive Gym was used.

A stdlib planar model gives 2 to 6 joint feeding arms an operational-space tip mass (servo, link and spoon masses plus each servo's rotor inertia reflected through a 345:1 gearbox) and feeds it to osrobot-assist's lip-contact trials (200 paired trials per row, 30 mm sway, camera arm). The tip mass falls only from 0.63 to 0.50 kg from 2 to 6 joints; the worst peak lip force is identical (2.0 N at 0.10 m/s, 4.0 N at 0.20) and the safe share at 0.20 m/s is 0.60 to 0.65, within sampling error. Servo cost alone triples, from $52–98 to $156–295 (two retailers' listings, 2026-10-19). Most of the mass is the geared-up rotor. Stylised: planar, one pose, position-only task, assumed constants, so the real reasons for more joints (spoon orientation, 3-D motion) are not tested.

The osrobot-assist spoon-feeding trial rebuilt in MuJoCo (a sphere on two slide joints, two box lips that follow the mouth's sway, MuJoCo soft contact), with the same controller and 200 paired seeds per row. With the arm blocked, the steady contact force equals the controller identity KP×LEAD×speed (2.00 N at 0.10 m/s) but the impact peak is 1.84× higher at every speed (3.68 N), and falls with more contact damping. With a swaying mouth the share fed at most 2 N drops from the stylised 1.00 to 0.45 at 0.10 m/s (30 mm sway, camera arm), with peaks of 11 to 23 N because the kinematic lips sweep into the tip, a contact the stylised model leaves out. Feeding time agrees within 4 % in all but one row. Stylised and preliminary: rigid lips that cannot yield, assumed contact damping, no human model.

A stdlib sweep of osrobot-assist's feeding-arm twin over lip stiffness, loop stiffness and forward lead (200 paired trials per row, 30 mm sway, camera arm at 0.20 m/s). Mean peak lip force rises from 0.32 N at 100 N/m to 1.15 N at 800 and 2.02 N at 6400; the worst trial reaches 6.7 N, above the controller's 4.0 N stall push, so that push is a ceiling only for soft and medium lips. At 0.05 m/s every stiffness stays under 2 N. Stylised: assumed constants, a point mass and a passive user.

A stdlib extension of osrobot-leannoise's planar feeding-arm twin in which the user's misjudgement of the spoon position shrinks, linearly over the last 40 mm, by a swept fraction (200 paired trials per row, 30 mm sway, 0.20 m/s, 100 mm/s lean). A blind arm's share of feeds with peak lip force at most 2 N is 0.83, 0.60, 0.47 and 0.41 for a constant misjudgement of 10, 20, 40 and 80 mm, and 0.99, 0.95, 0.84 and 0.70 when the error falls to zero at the lips; feeds that never reach the mouth (0.50 fed at 40 mm) disappear. With the error shrinking, the blind arm is within 0.02 of the camera arm up to 20 mm; the camera keeps its lead at 40 and 80 mm. Stylised: an assumed linear shrink law, assumed constants, a point mass and no real person.

A stdlib extension of osrobot-lean's planar feeding-arm twin in which the user leans toward a perceived spoon position that is delayed and offset by a per-feed misjudgement (200 paired trials per row, 30 mm sway, 0.20 m/s, 100 mm/s lean). A blind arm's share of feeds with peak lip force at most 2 N falls from 1.00 with a perfect leaner to 0.83, 0.60 and 0.47 at 10, 20 and 40 mm of misjudgement, and only 0.50 of feeds reach the mouth at 40 mm; the camera arm falls from 0.97 to 0.69 and always feeds. A perception delay of up to 0.2 s changes little; 1.0 s drops the blind arm to 0.46. Stylised: constant offset, assumed constants, no safety case.

A stdlib extension of osrobot-assist's planar feeding-arm twin in which the user shifts the mouth toward the spoon tip at a swept lean speed (200 paired trials per row). At 30 mm sway and 0.20 m/s the blind arm's share of feeds with peak lip force at most 2 N rises from 0.21 (passive user) to 0.69 at 50 mm/s and 1.00 at 100 mm/s, against 0.71 to 0.97 for the camera arm; at 50 mm sway even 100 mm/s gives only 0.62 for the blind arm. Leaning removes contacts but not the peak force of the ones that remain (4.0 N at 0.20 m/s). Stylised: every constant, including the lean speed, is assumed; no person was modelled or tested.

A stdlib extension of osrobot-pipeline that adds a hard finish deadline for 16 equal fields and searches up to four scouts, up to 16 treaters and up to 32 inline robots (no new simulation; every price an assumed round number). One scout cannot finish 16 fields faster than 5.3, 15.9 and 32.5 h (0.18, 0.72, 1.62 ha fields) however many treaters it feeds, against 0.7, 2.8 and 6.2 h for 32 inline robots. Cheapest pipeline over cheapest inline: 1.06 to 1.12 at 0.72 ha, rising as the deadline tightens, and 1.49 at 0.18 ha; at 1.62 ha the pipeline is cheaper (0.87 to 0.93) at every deadline it can meet but cannot meet 8 h, where only inline robots can. Stylised: equal fields, hard cut-off, no late-yield curve.

A stdlib scheduling follow-up to osrobot-pairprice using its fitted per-field minutes (no new simulation; every price an assumed round number). Set-up and battery swaps make the scout nearly as slow as the treater, so one scout keeps only 0.99 to 1.33 treaters busy and the cheapest pipeline is always one scout with one treater. Overlapping the stages cuts the held-pair penalty to 1.04 times the inline fleet at 0.72 ha fields (32 fields) and 0.86 at 1.62 ha, from 1.34 and 1.16; at 0.18 ha it stays 1.48 to 1.81 times. Extra treaters finish sooner but raise cost per hectare (13.92 to 24.56 USD for 1 to 5). Stylised: deterministic, equal fields, no weather or failures.

A stdlib follow-up to osrobot-scoutpair that prices its matched-kill pair (scout at 1 m/s, 3 m swath, 0.343 killed) against one inline robot, with every price an assumed round number. Simulated time is linear in area (R² ≥ 0.9999 over 0.045 to 1.62 ha), so cost per field is a fixed set-up part plus a per-hectare part. Charging each machine only for the minutes it works, the pair costs about 2.5 times the inline robot on a 0.045 ha field, breaks even near 0.5 ha (0.8 ha at a 4,000 USD scout, 1.3 at 6,000) and is about 0.7 times as much per hectare at 5 ha (extrapolated). If the two machines are held for the whole job, as one dedicated pair would be, it never wins at the reference prices (1.08 times at 20 ha). The verdict flips with utilisation (2,000 occupied hours a year gives 2.09 times at 0.72 ha) and with labour and battery life. Stylised: assumed prices and constants, a straight-line extrapolation beyond 1.62 ha, no registration error or wind; nothing was built or flown.

A stdlib follow-up to osrobot-fieldtour that makes the scout a second platform with its own swath and detector, then compares at matched kill because both platforms lose detections as they speed up. With a 3 m scout swath at 1 m/s the pair kills 0.343 in 26.6 min against about 42 min for one inline robot interpolated to the same kill; with the scout's swath cut to the ground robot's 1 m the same kill takes 47.3 min and the pair loses by about 13%. The saving is the swath ratio, not the second vehicle. A mud-aware tour changes the total by 0.1 min or less at 0 to 80% mud. Assumed constants, no registration error and no pipelining or capital cost: stylised, simulation only.

A stdlib follow-up to osrobot-field that swaps nearest neighbour for 2-opt as the second stage of scout-then-treat. 2-opt cuts the straight-line tour over about 232 detections from 539 m to 468 m (random order: 5,403 m), but a field still takes 56.1 min against 51.7 min treating inline (nearest neighbour: 57.0), and scouting is slower at every weed density from 0.05 to 1.2 per m². Stylised: assumed constants, a mud-blind planner, 30 fields per row; no robot was built.

A stdlib simulation of a lane-following weeding robot on a 60×30 m field with clustered weeds, a motion-blurred noisy detector and muddy squares. Scouting then touring the targets kills the same share of weeds as treating inline (0.399 against 0.401 at 1 m/s) and is slower in all 23 paired rows, by 2.8 to 7.8 min. Lane overlap repays a weak sensor at a steep price: 0 to 0.75 overlap raises the kill share from 0.40 to 0.86 and the time from 52 to 178 min; no grid configuration reaches a 0.9 kill share at a 0.5 hit rate. Stylised: every constant is assumed, the speed results move with the blur assumption, and the planner comparison is a statement about this cost model.

An open research programme on which hydraulic circuit a small shop can build from catalogue parts to get a 1.5 to 3 t excavator closest to stock energy use, and at what parts cost. Milestone 1 is a stdlib model of pump-shaft energy for three circuits (fixed-displacement open-centre, load-sensing, ideal independent metering; analytic single-actuator check, 6 tests) and a 2,916-case grid on one four-phase cycle: load-sensing uses a median 47% of open-centre shaft energy (22 to 81%), saving more as flow demand falls (31%, 47%, 63% of open-centre at 0.5, 0.75 and 1.0 of full flow), and metering uses a median 80% of load-sensing, 88% with equal actuator pressures and 74% at 1.5 times the reference spread. The ordering holds in every case but only because each rung removes a loss term, so it is a code check rather than a finding. The cycle, pressures, flows and margins are illustrative values written from memory and swept, not sourced; there are no dynamics, leakage, line losses, energy recovery or parts costs; the DOI registry was unreachable so references were checked by web search, not cite_check. Milestone 2a repeats the comparison on 20,000 random cycles inside a machine envelope (pump 45 L/min, 20 to 180 bar, taken from an unopened search summary of one mini-excavator listing): load-sensing uses a median 44% of open-centre energy, metering 72% of load-sensing (87% at equal actuator pressures), and H3 fails per case in about a third of equal-pressure cycles. The cycles are random, not measured. Preliminary; arithmetic only, nothing run on a machine.

Milestone 1 of track 23's marine-colregs step, filed under the marine-hull programme. Kinematic own boat (4 m/s, 6 deg/s turn limit) against a target on constant course (it never gives way). 115 scripted head-on, crossing (both sides, speed ratios 0.5 to 1.5), overtaking and overtaken encounters with lateral offsets of -100 to 100 m. Roles are classified from Rules 13 to 15; planners: rule-based 35 degree starboard turn, velocity obstacle limited to starboard for give-way, and sampling MPC. Scores: closest point of approach (safe = at least 100 m), rule compliance (starboard turn, pass astern in a crossing, stand-on holds course until the action range) and detour. Check: straight-course CPA matches the analytic offset. Sweeps: target-velocity error 0, 0.3, 0.6 m/s (3 seeds) and stand-on action range 200 to 600 m. Preliminary: no vessel dynamics beyond a turn-rate limit, the safe distance and speeds are chosen not sourced, the target never manoeuvres, no AIS replay, and the MPC cost weights are untuned (no CMA-ES arm). Simulation only; research, not a certified system.

An open research programme on which body (monohull, catamaran, SWATH, planing hull, hydrofoil, glider, AUV) carries a survey payload furthest per kWh, and where that ranking flips under uncertain drag inputs. Milestone 1 is a stdlib calculator (ITTC 1957 friction line checked analytically, form factor, wetted-area constant, hotel load; 9 tests) and a 6,561-case grid for a 200 kg craft: a catamaran of two slender demi-hulls out-ranges a monohull in 27 cases (0.4%) and in none of the 2,187 equal-length cases; the monohull's best speed is about 1 m/s, set by a 30 W hotel load, with range 560 to 1,188 km as an upper bound. The wave-resistance term is a placeholder and no published resistance curve is reproduced yet, so this is a friction-limited comparison; hull-form constants, efficiencies and water properties are swept or unsourced, and stability and deck area are not modelled. Two DOIs written from memory were wrong and were corrected by the citation check. Preliminary; simulation only. Milestone 2 replaces the placeholder wave term with Michell thin-ship theory on Wigley hulls: the catamaran wins 3.3% of 6,561 cases and none at equal length, and wave resistance changes best range by about 4% because the optimum speed (about 1 m/s) is at Froude number 0.16. Milestone 3a checks it against one published Wigley tank point (Ju 1983, via Wang et al. 2017): Michell plus the ITTC 1957 line with no form factor is within +1 to +5% of the measured total-resistance coefficient 4.16e-3 (water temperature unknown, 10 to 20 C), while the programme's central form factor 0.25 is 18 to 25% high; the form factor that fits is about 0, below the swept 0.10 to 0.40 box. One speed, second-hand source, geometry assumed. Milestone 4 compares a quasi-steady buoyancy glider with a same-mass propeller AUV over 19,683 swept cases (no glider coefficient is sourced: the papers could not be fetched): at equal hotel load the glider has the longer best range in 5.9% of cases, at 5 times the AUV hotel load in 89%, so the advantage comes from hotel load rather than hydrodynamics; glider speed tops out at 0.13 to 0.70 m/s, and the proposal's 0.4 to 0.9 m/s crossover is not supported. Not checked against a fielded glider's range. Milestone 4b re-ran it with Slocum aerodynamics read from the MIT-licensed gliderflight package (lift slope 7.5/rad, K 10.5/rad²): the glider out-ranges the AUV in 0 of 2,430 cases at equal hotel load, and the break-even AUV hotel load is a median 4.9 times the glider's (10th to 90th percentile 2.4 to 10.7); hotel loads and pump efficiency remain swept, not sourced. Milestone 4c reads a REMUS-100-class AUV's total power from vendor endurance (1.5 kWh, 12 h at 1.5 m/s to 8 h at 3 kn: 125 to 188 W), subtracts the model's 13 to 21 W of propulsion, and gets a 104 to 175 W non-propulsion load; against it the glider has the longer best range in all 810 x 18 cases. The vendor figures are approximate and mutually inconsistent, and energy per metre ignores that the glider is about seven times slower. Milestone 5 rungs: a lifting-line hydrofoil (all foil coefficients swept) and a Savitsky (1964) planing hull both fail to out-range the 4 m monohull (0 of 2,916 and 0 of 48 cases; range ratios up to 0.25 and 0.14), so they are transit-speed options, not endurance options; no tank data checked. Milestone 5b adds a SWATH rung (two fully submerged parabolic torpedoes on Wigley-form struts, Michell wave drag for torpedoes and struts, 4,374 pairs against the 4 m monohull): the SWATH out-ranges the monohull in none, at its best speed (range ratio median 0.69, best 0.83) or at a fixed 3 m/s (median 0.46), because strut wave drag and wetted area outweigh the torpedoes' small wave drag at 200 kg; torpedo wave drag falls with depth at about 2 k₀ as theory says, but no SWATH tank curve was reachable, and stability and seakeeping are not modelled. Milestone 5d charges panel mass to the battery (areal mass 2, 5 and 10 kg/m², swept, not sourced) and lets the array area vary: where the array cannot carry the boat the full deck is still the best layout in every finite-range case, but the range multiplier falls from a median 2.01 (no mass) to 1.95, 1.83 and 1.64; neutrality at 1 m/s is unchanged at 25.1%, by construction. Milestone 5g runs three speed policies against ten years of daily NASA POWER insolation at five sites: slowing when the pack is under 30% never raises distance (hotel load dominates at 0.5 m/s) but cuts pack-empty days, e.g. 25.4% to 4.2% (slow) and 1.5% (hold) for a 4 m boat at 25 N; above 42 N the deficit is seasonal and policy barely helps. Threshold and slow speed are not swept. Milestone 3b checks the placeholder wave term against five bare-hull tank speeds of a published 2.8 m catamaran (Zenodo 18459868, CC BY 4.0; geometry and water type inferred from the tables): the central model (k = 0.25) is 35 to 43% low at Fn 0.33 to 0.58 and within 10% at Fn 0.72 and 0.79; no single wave scale keeps all five speeds within 20% (best worst-case error 29 to 32%), so the placeholder is not safe for catamarans above Fn 0.3. The programme's survey speeds (Fn about 0.16) are below the data's range, and H3 (Holtrop-Mennen) is untouched.

Milestone 1 of track 23's marine-path step, filed under the marine-hull programme. A stdlib 3-DOF (surge, sway, yaw) 200 kg boat with a rudder and surge PI, in a constant earth-frame current and a six-sinusoid wave disturbance (4 to 9 s periods, force scale swept, not sourced), follows a straight line under plain LOS and integral LOS (Borhaug, Pavlov and Pettersen 2008). Check: the simulated plain-LOS offset equals the analytic delta*tan(asin(c/U)) to 1% (0.3 m/s, 3.774 m). Sweep: cross current 0 to 0.5 m/s, wave scale 0 to 1, look-ahead 6, 12, 24 m, integral gain 0.5, 1, 2, 3 seeds. Integral LOS has lower rms cross-track error in 136 of 144 cases with current at least 0.2 m/s; the 8 exceptions are all look-ahead 6 m with gain 2, which oscillates. Energy per ground km differs by a median 1.4% (max 37%, in the oscillating cases, which also include the transient). Damping, rudder and wave-force coefficients are not sourced; the model has no Coriolis cross-coupling beyond u*r and v*r, no MPC arm, no JONSWAP spectrum and no field data. Preliminary; simulation only.

A stdlib extension of osrobot-deadline (no new simulation; fleet prices still assumed round numbers) that charges each late field 2% of its revenue per leaf stage of delay past the window, the pooled corn and soybean figure of Knezevic, Evans & Mainz (2003, Crop Management 2(1)). Leaf-stage length (2, 4, 8 days), an 8 h working day and revenue (1,000 and 2,000 USD/ha) are swept, not sourced. For 16 equal fields the fleet minimising fleet cost plus lost yield is searched over up to four scouts, 16 treaters and 32 inline robots. Pipeline over inline cost is 1.06 to 1.20 at 0.72 ha and 1.49 to 1.51 at 0.18 ha in every cell; at 1.62 ha it ranges 0.89 to 1.18, with inline cheaper when the window closes by 8 h (12.27 against 13.07 USD/ha at 2,000 USD/ha revenue and a 4-day stage) and the pipeline cheaper from 12 h. Late loss reaches 8.56 USD/ha in the worst cell against about 12 of fleet cost. Limits: the 2% rate is for corn and soybean up to canopy closure and is applied linearly; stage length and revenue unsourced; equal fields, one window, no weather or failures; search capped at four scouts.

osrobot-dof priced servos alone and said extra joints triple the cost. Pricing the whole arm from the published SO-ARM101 parts list (board, cable, power supply, clamps, screwdrivers; the model reproduces the table's $121.94 follower and $229.88 two-arm totals exactly) gives a 6-joint to 2-joint ratio of 1.84 at the table's $13.89 servo ($66.38 against $121.94), 2.15 at a $26.00 retail listing and 2.44 at $49.22; it reaches 2.5 only at a $57.90 servo. Printed links add $1.50 to $4.00 across a swept $15 to $40 per kg filament price. Prices are a snapshot, the 7.4 V and 12 V servo classes are mixed, and camera, mount and safety stop are not priced, so every total is a lower bound. Only the filament price is not from a source (swept).

osrobot-field ran on round numbers. This follow-up replaces the swath (2 m), time per weed (0.73 s) and kill rate (99.2%) with figures from laser-weeding literature and trade sources (read from search summaries, to be re-checked), and sweeps what has no source. Treating inline still beats scout-then-treat with equal kill: at 0.3 weeds/m² and 1 km/h, 71.7 min against 79.3 and 180 Wh against 216; at 3 weeds/m² 110.8 against 133.4 min. Scout never wins in any row. The sourced constants cut the time from 151 to 72 min for the same field. Speed trades kill for area rate (0.65 to 0.38 kill and 0.08 to 0.53 ha/h from 0.5 to 4 km/h); detector hit rate sets kill almost one for one. Stylised and preliminary: the detector hit rate (0.7) and blur law, false positives, mud, energy per metre and beam-steering speed are assumed, kill is limited to a single pass, the laser stops for each weed, and no weed-density source was found (swept 0.3 to 10 per m²).

A stdlib linter for Opentrons Python API scripts follows tip state, used tip wells, declared well volumes and the liquid in the tip without running the script. On the 522 seeded faults of osrobot-lab it flags 362, against 270 raised by the real simulator and 366 flagged by the twin, adding reused tips, tips left on, sub-minimum volumes and carry-over; it cannot see the travel-height collisions. Agreement with the twin is partly circular (same author, same fault set), and every limit is an assumed value. Stylised.

osrobot-mjfeed assumed an 800 N/m lip and a 2 N comfort limit. This project re-runs the same MuJoCo feeding trial with the ISO/TS 15066 face effective spring constant (75 N/mm, 75,000 N/m) and the standard's face force limits (65 and 130 N; secondary sources disagree on which is quasi-static, so both are used). A 94× stiffer lip raises the blocked-arm impact peak about 2× (3.68 to 7.56 N at 0.10 m/s, 7.35 to 15.21 N at 0.20 m/s) and leaves the steady push unchanged. For the camera arm against a 30 mm sway (200 seeds), the share fed with peak contact within 2 N falls from 0.45 to 0.32 at 0.10 m/s, but within 65 N and 130 N it is 1.00 in every row; the worst peak is 53.3 N. So the assumed 2 N, not the physics, drove the earlier unsafe reading. Contact damping has no sourced value and is swept (blocked-arm peak 12.7 to 49.3 N from 100 to 1000 N·s/m at 0.20 m/s). Limits: the standard's tables were read only from secondary excerpts, lips are kinematic and cannot yield, pressure limits were not applied, tip mass, gains and sway remain assumed, no human model. Simulation only, not a safety case.

osrobot-mjmodulus showed the half-space Hertz model is out of range at published 4 to 45 kPa face moduli. Here the tissue is a layer of 5 to 20 mm bonded to rigid bone, using two published bottom-effect polynomials (Dimitriadis and Garcia, as tabulated by Dal Fabbro et al., ν = 0.5). The force that indents the layer by 30 % of its thickness is 0.05 to 9.4 N across tip radii 4 to 15 mm; for a 15 mm lip and the 4 mm tip it is 0.2 to 2.2 N. The simulated trial forces (12.6 to 53.3 N) exceed it by 1.3 to 1000 times, and at them the model's depth is usually beyond the layer thickness. The two corrections agree within about 40 %. Limits: homogeneous linear bonded layer, polynomial fits extrapolated from micro-scale indentation, the 30 % limit is my choice, thicknesses are lip values from one summary, no finite-element check. Not a safety case.

osrobot-mjpress found the pass or fail of the feeding arm's tip pressure flipped on an unsourced face modulus swept from 0.05 to 5 MPa. Published in-vivo values are 4 to 45 kPa (a one-subject lip and cheek device, Hertz fits to skin indentation, an aspiration study), all below that sweep. At those moduli a Hertz sphere of the 4 mm tip at the worst trial force (53.3 N) has a contact radius 3.5 to 7.8 times the tip radius and an indentation depth of 48 to 241 mm, so the half-space model is out of range and its 1.8 to 8.8 N/cm² pressures are a bound, not a result. Hertz holds (a/R ≤ 0.5) only for a tip radius of about 73 to 245 mm. The forces come from a stiff kinematic-lip simulation that is not self-consistent with a soft face; the modulus values come from small studies read as summaries; pressure limits remain unconfirmed. Not a safety case.

osrobot-mjface found every camera-arm trial under the 65 N face force limit (worst 53.3 N) but did not apply pressure limits. This project divides the same 200-seed peak forces by a swept contact area. The worst trial needs 0.48 cm² to stay within 110 N/cm², 0.97 cm² within 55 and 2.67 cm² within 20; a Hertz contact of the 4 mm tip gives 0.41, 1.93 or 8.9 cm² for a face modulus of 5, 0.5 or 0.05 MPa, so the verdict depends on a modulus that is not sourced. Lightly damped contact (b = 4 to 100) needs a third of the area of b = 1000. Pressure limits (20, 55, 110 N/cm²) are from conflicting secondary excerpts and not confirmed; damping, modulus, tip mass, gains and sway are assumed; lips are kinematic. Simulation only, not a safety case.

osrobot-mjlayer found the force that indents a lip 30 % of its thickness is 0.05 to 9 N, far below the 12 to 53 N a kinematic spring produced. Here the osrobot-mjfeed arm (MuJoCo dynamics, same controller and 30 mm sway) meets a lip whose force is that layer law, with assumed viscous damping and a bone stop, across 4 to 45 kPa, 5 and 15 mm and 0.02 to 0.20 m/s, 100 seeds per cell. Peaks are mostly 1 to 3 N, but the indentation stays within 30 % of the layer in only 36 to 100 % of trials: 100 % at 0.02 m/s for a 10 kPa, 15 mm lip (feed time 8.7 s), 52 % at 0.10 m/s, 36 to 53 % for a thin soft lip at any speed. A blocked arm pushes 1 N by its own controller; a 0.2 to 0.5 N cap brings the depth under the limit in the 10 kPa case and a 0.2 N cap in the softest, and the simulated depth matches the closed-form law. Caps barely trim fast impacts. More lip damping raises the share within the limit (52 to 85 %) and the peak force about 8 times. Limits: homogeneous linear bonded layer, MuJoCo used as an integrator with hand-coded contact, unsourced damping, bone stop and controller gains, no pressure limit, no human model. Not a safety case.

A stdlib model times a 96-well plate fill on an XY gantry and a two-link arm with 1 or 8 channels. At 10 µL per well one channel takes 69.7 s (gantry) or 66.7 s (arm); eight channels take 9.5 and 9.0 s. Channel count cuts the time by 86–88%; robot type moves it by 4–14%, because 76–79% of the time is fixed aspirate and dispense actions, so well order saves only about 2%. Which robot is faster flips with the assumed speeds; the 8-channel gantry is cheapest per plate-hour by a small margin. With assumed joint noise the arm misses the 1.5 mm well radius often enough to hit all 96 wells with probability 0.005. Stylised: every speed, price and accuracy figure is assumed.

A stdlib re-run of osrobot-deadlineloss (16 equal fields, 2% yield loss per leaf stage late, Knezevic, Evans & Mainz 2003) with two list prices read on 2026-10-23 in place of round numbers: a Holybro X500 V2 PX4 kit (609 USD, no battery, 18 min hover) and an AgileX Scout Mini (4,500 USD). Pipeline over inline cost moves from 1.49 to 1.54 at 0.18 ha, 1.11 to 1.14 at 0.72 ha and 1.07 to 1.15 at 1.62 ha (8 h window, 2,000 USD/ha). At 1.62 ha and a 12 h window the fleets tie when the ground platform costs 4,500 USD. A 12 min loaded flight time raises the ratio by 0.12 and can flip the verdict. Not sourced and swept: battery, camera, weeding tool, inline-robot premium (0 or 1,500 USD) and loaded endurance; no weeding tool is priced for the treater, so both ground fleets are too cheap. Two retail prices; simulation only.

Checks the osrobot-uav multirotor power model (imported unchanged) against two aircraft with a published mass, pack energy and endurance: DJI Mavic 3E (1.050 kg, 77 Wh, 45 min) and Parrot Anafi UKR (1.450 kg, 78.5 Wh, 38 min at 21.6 km/h), taken from Wikipedia's specification tables. Rotor diameter is not in the sources, so it is swept from 0.15 to 0.30 m: the model's endurance over the published one runs from 0.68 to 2.21 (Mavic) and 0.47 to 1.47 (Anafi). The model matches the published endurance at 0.181 m (Mavic) and 0.248 m (Anafi) at 90 % usable energy. At 0.22 m the errors have opposite signs (1.31× and 0.85×). Two aircraft of one class only; the fixed-wing and VTOL models are not tested; figures are from a secondary source and manufacturer maxima. Update: DJI's listing gives the Mavic 3 propellers as 9.4 in (0.239 m); at that diameter the model gives 1.30 to 1.63× the published endurance (80 to 100 % usable), so the earlier matching diameter of 0.18 m was a fit to a wrong model, not a rotor size. The Anafi diameter is still unsourced.

osrobot-fieldreal left two next steps: a second pass over missed weeds and treating on the move. This project adds both to its sourced-constant weeding model (2 m swath, 0.73 s per weed, 99.2% kill from laser-weeding literature; 60×30 m field, 20 seeds per row). At 1 weed/m², three passes at 4 km/h kill 0.780 of the weeds in 75.1 min, against 0.607 in 82.8 min for one pass at 1 km/h; two passes at 2 km/h reach 0.778 in 90.5 min. Because the assumed detector loses hits as speed rises, the second pass recovers more than speed costs. Treating on the move rather than stopping cuts time by 6% at 0.3 weeds/m² and 39% at 3 weeds/m² (1 km/h) with identical kill, and is bounded by laser throughput at 10 weeds/m². If 30% of weeds are never seen by the detector, kill falls (three passes at 2 km/h: 0.894 to 0.627) but the pass-count ranking is unchanged. Stylised and preliminary: detector hit rate (0.7, swept 0.5 to 0.9) and blur law, false positives, mud, energy per metre and beam steering are assumed; passes are independent except for the swept hidden share; no weed-density source was found; simulation only.

osrobot-weedpass assumed a detector hit rate of 0.7 falling as exp(−v/2 m/s). This project fits p_max·exp(−v/v_blur) through two published detection rates (88.94% at 0.2 m/s, Frontiers in Plant Science 2022; 86.2% at 0.425 m/s, arXiv 2512.10319), giving 0.914 and 7.19 m/s, and re-runs the multi-pass sweeps (60×30 m field, 20 seeds per row, treating on the move). The detector is far better than assumed (0.846 against 0.530 at 2 km/h), so one pass at 2 km/h kills 0.842, and two passes at 4 km/h kill 0.952 in 39.8 min against 0.875 in 68.4 min for one pass at 1 km/h: fast multi-pass still wins, and a third pass adds only 0.020. With a share of weeds never seen, three passes cap at 0.996, 0.946, 0.897, 0.799 and 0.700 for 0, 5, 10, 20 and 30% hidden. Limits: two points from two different robots read from search snippets (the papers were not opened), a flat two-point fit extrapolated to 4 km/h, independent detections apart from the swept hidden share, and assumed mud, energy and false-positive figures.