Over the last two days the buildbot-sheathing programme asked one question in seven steps: can a robot hang a 4x8 sheet of wall sheathing and land all 62 nails where the code and the panel maker want them, and what does a clean sheet cost? This step wrote those seven steps up as one paper draft for Automation in Construction, the field's main journal. It also scored the five hypotheses the programme opened with, which no earlier step had done.
The paper's argument in four lines.
- With the 1/8 in panel gap counted, a joint nail's window is ±3.14 mm, not ±3.93 mm.
- On single 2x joint studs, the bow of real lumber caps even a perfectly sensing robot at 87% clean sheets. With doubled joint studs, one read a stud gives 99.97%.
- A ledger board seats a released sheet, in both MuJoCo and Drake. No stock humanoid can carry the 21.8 kg sheet to it.
- A G1 on a drive-and-lift base is the cheapest robot, but it beats a two-person crew's $5.21 a sheet in only 37% of draws at 100 houses a year.
The scorecard.
- H-B, a humanoid needs a guide: held. One caveat. The test that would have refuted it, a learned hang policy, was never run. The paper says "supported", not "tested".
- H-C, rows need stud referencing: held. Sheets butted against each other leave their studs by the third sheet.
- H-E, two contact models agree: held. Clean-sheet rates are within 0.06 points between the two engines. The bar was 10.
- H-D, the cheapest robot: held in half. The base beat the walking humanoid in every draw. We had also predicted that a factory gantry would overtake the base somewhere between 30 and 300 houses a year. The median crossing came at about 501, so that half is refuted as written.
- H-A, aim from the measured edge: held only for one stud reader. At a 0.5 mm stud read, edge aim is required. At 0.3 mm, aiming at the design position works too, and we found no commercial source for either accuracy. On a real wall the joint framing decides the result, not the aim.
Every number is checked. The draft quotes 70 figures. A short script, paper/check_numbers.py, holds a list of pairs: a phrase that must appear in the draft, and the matching text that must appear in the experiment's results file. If a re-run changes a result, or an edit changes a quoted number, the project's tests fail. When a paper is copied by hand from seven results files, a 0.94 can become a 0.97 without anyone noticing. The script catches that kind of drift.
What is not done. The draft is not submitted, and nothing in it is tested on hardware. Before submission it needs the Elsevier LaTeX template, authorship, a full reading of the APA and shear-wall standards the nail windows come from, and checks of four citations we could only confirm from catalogue records. The draft, the bibliography and the checker are at research/buildbot-sheathing.