A bowed sheet edge is a rocker, not a seesaw

Modelled as a curve, a 1/16 in bowed sheet edge rolls back to level on a ledger. The 0.149° rock we had charged came from modelling the bow as three feet.

The buildbot-sheathing programme asks whether a robot can hang a 4x8 sheet of wall sheathing and land all 62 nails where the code and the panel maker want them. One of its placement aids is a ledger board: a level strip screwed to the wall that the sheet rests on, so the board squares the sheet instead of the robot. Yesterday's sim-to-sim check (MuJoCo against Drake) agreed on almost everything. The exception was the sheet's own bottom edge. The panel standard (PS 2) lets that edge bow up to 1/16 in, and on the ledger the bowed sheet tipped 0.107–0.123° in Drake and 0.139–0.145° in MuJoCo. Single 2x joint studs leave a hang budget of about 0.03°, so that disagreement was larger than the whole budget.

We had modelled the bow as three feet: the two corners, and one at mid-width 1/16 in lower. Resting on that middle foot, the sheet is a seesaw. Its centre of mass sits 1.2 m above a single point, so it has to fall onto a corner, and the rock comes out at 2 s / W = 0.149°. A real bowed edge is a curve, though, not a point. A parabolic 1/16 in bow across 48 in has a radius of 117 m. A curved base whose radius is larger than the height of the centre of mass above it is a rocker. Tip it and the centre of mass rises, so gravity rolls it back. Here the radius is about 100 times that height.

So we wrote down four hypotheses and rebuilt the edge as 25 to 401 small feet along the curve, in both engines.

  • A smooth bow rolls back to level. Once gravity restores the sheet, only friction against the wall can hold a tilt, and that has a closed-form limit: θ_s = μ F_n r̄ / (m g (R − h0)). All 48 cells, in both engines, seated inside it. That means 0.007° at most under a 5 N press against the wall, and 0.058° at 50 N. Over 2,000 random seats, the 99th percentile was 0.06° in both engines, against up to 0.149° before.
  • It matters, but not where it would change the conclusion. Put into the end-to-end pipeline model, the best single-2x pipeline goes from 56.0% to 62.5% clean sheets. Every single-2x pipeline gains 1.8 to 6.9 points, and the two engines now agree within 0.11 points. Doubled joint studs move by under half a point; they were already at 94–100%. Single 2x joints are still nowhere near 95%, because what caps them is bowed studs (an earlier milestone), not the bowed sheet.
  • A kink keeps the rock, but not reliably. We predicted that a sharp point at mid-width (a tent-shaped edge) would bring back the full rock. That prediction failed. Balanced on a point, the sheet's tipping moment at 0.149° is 0.68 N·m, while a 5 N press against the wall holds 3.1 N·m of friction. The sheet stops wherever its approach leaves it, anywhere from 0.01° to 0.15° depending on the engine. That is the whole reason the two engines disagreed yesterday: neither was wrong, and the three-foot model has no stable answer.
  • Grit is worse than the bow. One millimetre of grit on the ledger, or a burr on the edge, tips the sheet onto its far corner: a median of 0.05–0.06° and up to 0.09°. The PS 2 bow allowance does less damage than that.

Our pre-registered convergence test (single seats change by less than 0.005° between 25 and 97 feet) also failed. A sheet resting anywhere inside the friction band is at rest, so exactly where it stops is sensitive to everything. We added a check afterwards, marked as post hoc in the code, to see whether the band converges. In Drake it does: every seat lands inside the band at every foot count from 25 to 401. In MuJoCo, at a 5 N press, up to 0.037° gets past the band at every foot count, so that excess comes from MuJoCo's soft contact, not from the number of feet.

For anyone building a ledger-based sheathing cell, two rules follow: back the press off to a few newtons once the sheet is down, and keep the ledger clean. Neither has been checked on hardware. PS 2 bounds how far an edge may bow, not what shape the bow takes, and we found no measured edge profiles. The parabola and the tent are the two ends of what we tried. Code, numbers, the figure and the pre-registered plan: research/buildbot-sheathing.

All writing · More in Research findings · More on buildbot-sheathing · All tools and research