Does a purpose-built nailing gantry beat a stock humanoid at fastening sheathing, and at what house volume does it pay?
research/buildbot-custom
Does a purpose-built nailing gantry beat a stock humanoid at fastening sheathing? A modelled two-axis bridge nails a 62-nail sheet in 61 s (6.1x hand nailing); a $100k machine needs about 150 houses a year to break even on walls (50 with roof cassettes), where the H1 saves money in 0.4% of draws (price and times swept, not sourced).
What it found
Designs the machine the earlier buildbot projects point to: a two-axis bridge with a nailer over a flat panel table. On the 62-nail schedule of a 4x8 sheet from buildbot-fasten, 1 m/s axes, 3 m/s2 and 0.5 s a nail give 61 s a sheet, 6.1× faster than hand nailing at the taskbook's minutes; over the sweep 20 to 115 s. The nail windows need a total tip scatter under 1.24 mm, so finding the studs, not the axes, takes the budget. Nailing is 3.3 crew-hours on the walls and 9.0 with roof cassettes ($125 and $341 a house at a $37.86 loaded hour), so a $100k machine breaks even at about 150 houses a year (walls) or 55 (with roof); over a sweep the median is 326 and 120. The H1 saves money in 0.4% of draws on the same wall hours, the G1 cannot take them, and neither reaches the roof. Machine price, load and scan times, axis performance and operator share are swept, not sourced; no machine is built or quoted and the hand minutes are the taskbook author's estimate.
A figure from its own output
Paper, code and results
More in Robots on the building site
- Can a stock Unitree G1 or H1 humanoid push a nailer into a wall and place nails inside a stud face?Can a stock Unitree G1 or H1 push a nailer into a wall and place nails inside a stud face? One 4x8 sheet is 62 nails, 34 with a +-3.9 mm window; the tip needs 1.24 mm scatter for a 95% sheet. The G1 pushes up to 142 N and reaches 18 of 62 nails from one stance, the H1 up to 276 N and 36 of 62 (static, peak torque, no policy).
- Can a G1 humanoid nail a 4x8 sheet lying flat on a table?Can a G1 humanoid nail a 4x8 sheet lying flat on a table? Not better than on a wall: pushing down at 50 N with its feet planted it holds only 30 of 693 reach cells, reaching 48 of the sheet's 62 nails in 16 stances (13 stances for all 62 on a wall), and matching hand nailing (372 s) only if each re-stance takes under 23-24 s (push, standoff and re-stance time are swept, not sourced; static IK, no walking).
- With the simulated carry and nailing results in place of a swept slowdown, does a stock humanoid pay on a house site?With simulated carry and nailing results in place of a swept slowdown, does a stock humanoid pay on a house site? The H1's carry slowdown is 0.41 to 1.70 against a one-stud labourer (pays in 72% of cells) and 1.63 to 6.82 against a bundle carrier (never pays), against a break-even of 1.07; the dollars are tens per house.
- How long can a Unitree G1 hold a building load before its motors overheat?Can a stock humanoid hold a stud for a minute? G1 static joint torques re-read against a swept continuous-torque fraction: the waist is at 0.54 of peak with empty hands at the carry pose and 0.91 with a 3.8 kg stud, so below a 0.6 continuous ratio nothing is sustainable (0.6 kg at 0.6) and a stud reaches the winding limit in 0.14 thermal time constants at 1/3. A 150-pose search finds a posture (hands 0.15 m ahead, 0.7 m high) at 0.33 of peak, so the 0.91 is not the floor.
- Does a searched posture let a Unitree G1 hold a stud longer than the default pose?Does a searched arm and waist posture let a G1 hold a stud longer? In MuJoCo the stud's peak joint fraction falls from 0.91 (default) to 0.76 with the centre of pressure on the sole point and to 0.49 (best of three noisy seeds) 3 cm inside the toe edge, enough to make a stud sustainable at a continuous rating of 0.5 of peak; a 2x6 goes 0.93 to 0.68 and sheets stay above peak.
- Does a hand-written flexing-sheet plant agree with MuJoCo, and do its sim-to-sim transfer failures hold there?Does the hand-written flexing-sheet plant of buildbot-skills agree with MuJoCo? Pass/fail labels agree on 99.8% of 2600 episodes across 13 policies, so the sim-to-sim transfer failure (0-11% vs 81% where tuned) is not an artefact of my equations.
All 27 Robots on the building site projects →
Other research topics
Proper scoring & elicitation28Effort & peer prediction9Verification games & slashing19Detection & tolerances14Markets as routers13Prediction-market design9Paying contributors8Decentralised training systems23Digital twins & sim-to-real164LLM chip design loops40Robot manufacturing loops53Ambient AI assistant46Self-replicating manufacturing34Simulation-first design64Engineered clothing45Task-specific AI accelerators40Algorithmic chip-design loops42Open-source purpose-built robots44The software-engineering loop43Adversarial human imitation41Open-source residential construction41Autonomous multi-angle capture34Sensing, imaging and 3D31Energy supply and dense energy storage39Robot societies among people31Agentic co-design of low-cost robots27Laundry-hanging robots25Machine learning on qualitative data23Robots on the building site27Civilizations & group psychology1History & human sciences1
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.