A robot can push a bowed stud straight with 28 newtons. It still doesn't get to 95%

Pushing a bowed joint stud onto its line takes 28–126 N, inside a stock G1's reach. It lifts single-2x joints to 71% clean sheets, then hits a ceiling.

The buildbot-sheathing programme has been chasing one question for nine milestones: can a robot nail the sheathing on a wood-frame wall to the windows the panel maker and the code leave it? Milestone 2 found the thing that caps it, and it is not the robot. It is the lumber. A 4x8 sheet's edge is straight; a real 2x4 is not, and a stud bowed sideways moves its 1.5 in nailing face out from under the edge. At the bow published for kiln-dried studs, a robot with perfect sensing gets 87% of sheets fully clean on a single-2x joint. Doubling the joint stud fixes it (99.97%), but milestone 7 priced that stud at $5.42 a sheet — more than the rework it saves.

Carpenters have a third option that costs nothing: push the 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 next sheet. Nobody in the robotics literature seems to have asked what that push takes, or what it buys.

It takes about 28 newtons. Model the stud as what it is: a Stud-grade 2x4 bending flatwise (I = 0.984 in⁴) over its 92 5/8 in clear span, held at the plates. One push at mid-height, onto the line through its ends, takes 48 E I b / L³ — six pounds for the median stud, 28 lbf at the 99th percentile, 85 lbf for a stud at the grading limit with stiff Douglas fir. (I checked the closed form against a finite-element beam; both give c = 48.000 pinned, 192.000 if the plates clamp the ends, which they don't.) The modulus comes from the NDS reference values for Stud grade, 1.2 to 1.4 million psi, spread lognormally at the 0.25 coefficient of variation the code uses for visually graded lumber.

A stock humanoid can just do it. Running the push through the same static whole-body model that earlier milestones used for pressing and carrying, a Unitree G1 standing at the wall can push 81–88 N sideways at stud mid-height before its ankle roll gives out; an H1 manages 120–123 N. That covers 93% of low-warp studs for the G1 and 99% for the H1. It is a near thing, which is itself the useful finding: a robot that can't move a particular stud can read that it can't and fall back.

Order matters more than I expected. The obvious robot policy is to push each nail's point onto the line just before driving that nail, bottom up — the way a person does it, one hand on the stud. But every nail you drive shortens the free span of what's left, and a short beam is a stiff beam. Nail by nail, the 99th-percentile force rises from 126 N to 1,258 N on high-warp studs, and 2,821 N at the grading limit. Push the whole stud first, then nail.

And then it doesn't get to 95%. This is where the pre-registered hypothesis died. Straightening does remove most of what bow costs: the best single-2x pipeline goes from 0.560 to 0.710 clean sheets at the lower published warp, and from 0.307 to 0.677 at the higher — up to 37 points. But the bar was 95%, and it lands at 71%. The reason is in the control: run the same pipeline on perfectly straight studs and it gives 0.737. Straightening does not underperform; it reaches the ceiling. With bow gone, what closes the joint window is the sheet's own pose where the ledger board leaves it.

That is worth more than a held hypothesis would have been. The 44-point gap between single-2x and doubled joints was never one problem. It is two, of nearly equal size: about 18 points of bowed lumber, which 28 newtons removes, and about 26 points of sheet placement, which they don't. A robot builder reading the earlier milestones might reasonably have concluded that better stud sensing or straighter lumber was the path to single-2x joints. It isn't. The path is a better ledger.

In dollars the push is free (it beats the extra stud in every draw until about 15 minutes of pushing per sheet) and roughly a wash against not pushing at all at low warp, paying off in 92% of draws when the lumber is worse. So the recommendation from milestone 7 stands unchanged — double the joint studs if you want 95% first-pass — with a cheap fallback added for the robot that arrives at a wall already framed with single 2x.

Numbers, code and the pre-registered hypotheses (H41–H44, two held, one held on a bar that turned out to be too low, one refuted): research/buildbot-sheathing, milestone 9. Simulation only; the end restraint at the plates and whether a pushed stud stays put between sheets both want a mock-up.

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