Labour is a large part of what a new house costs. Humanoid robots like Unitree's G1 and H1 are now on sale to anyone. So it is fair to ask whether one could do useful work on a residential building site, perhaps beside a crew. Before anyone trains a robot to frame a wall or hang drywall, there is a plainer question. Can it even hold the materials?
This repository opened a long-term research track on that question, robots on the residential building site (track 20). Its first project, buildbot-payload, gives a first answer in simulation.
What was measured
The project loads the open MuJoCo Menagerie models of the G1 and H1. For each hand position, from 0.2 to 2.0 m high and 0.2 to 0.6 m in front of the feet, it finds the heaviest load the robot can hold in both hands while standing still. The feet stay flat. The robot leans back so that it and the load balance over its feet. Every joint must stay inside the torque limit the model lists. A test checks the solution against MuJoCo's own forward dynamics: with the computed torques applied, nothing accelerates.
What it found
- The G1 (33 kg) holds about 4.5 kg. It manages 1.5 to 4.7 kg with its hands 0.4 to 0.5 m out, and up to 12 kg held close. Unitree rates the G1's arm at 2 kg (3 kg for the EDU) and notes that the figure depends on posture, so the simulation and the vendor broadly agree. What gives out is not the arm but the waist pitch motor (50 N·m), in 16 of the 28 reachable positions.
- The H1 (51 kg) holds about 14 kg, 12 to 15 kg over most of its reach, and its 18 N·m elbows are the limit.
- On real tasks. The G1 can carry an 8 ft 2x4 or hold a framing nailer, but not a 2x6. The H1 can carry studs and take one end of a 4x8 sheet of OSB or drywall while a person takes the other, and it can set a lightweight concrete block. Neither can carry a whole sheet alone, set a normal-weight block or lift a 60 lb bag, and neither reaches an 8 ft ceiling. The task masses come from maker and standards data: USG for drywall, the NDS for lumber, a masonry survey for blocks, and a nailer's listing.
So a stock humanoid today is a helper for light handling and a second pair of hands on a shared sheet. The person's half of that sheet, 10 to 12 kg, is well under the NIOSH lifting guideline. Heavy pieces still go to people, lifts and cranes.
How far to trust it
These are static limits at each model's peak torque. A robot holding a load for minutes runs nearer its continuous rating and will manage less. Walking, gusts on a sheet, grip and the robot's hands (the G1's have no fingers, and the H1 has none at all) are not modelled. The Menagerie models are the community's approximations of the robots, and nothing was tested on hardware. The pose solver also picks a posture near the robot's home pose rather than its strongest one, so some limits, particularly the waist, may be pessimistic.
What comes next
The track's plan, in research/PROJECTS.md, goes from here to dynamic carrying over site terrain, fastening, a person and a robot holding one sheet together under contact-force limits, and a published library of trained skills. It then turns to a purpose-built machine for whichever task holds the most labour cost, and finally to what all of it would take off a house's price. This first result already points the purpose-built branch somewhere specific: stronger waist and elbow motors.
The research was done by an AI coding agent (Claude) in one session, with the code, tests and data in the repository.