A robot learned to seat a prefab window in a 2-millimeter fit with 12 minutes of human help

A research team ran a robot through the most punishing step of prefab window installation: seating a modular unit into a two-millimeter clearance on every side, with friction and pose errors randomized around the workflow. After three hours of online training inside MuJoCo, a robotics simulator, the policy hit 100% autonomous seating. The human installer only had to grab the controls for twelve to fifteen of those minutes, the preprint reports.

The mechanism is the contribution. Rather than dense demonstrations, the team recorded a small number of teleoperated runs and then let the policy ask for help only at the moment a contact was failing. A Q-chunking policy extended each action into a short recovery sequence, and a non-updating warm-start phase stabilized the handover from offline to online learning. Sparse takeovers and temporal abstraction together turned a few minutes of installer intervention into a tolerance-critical recovery policy.

The result is single-team simulation evidence, not a jobsite deployment. No hardware trial, no third-party replication, and no construction partner is named. The two-millimeter figure is a property of the simulator's contact physics. Whether sparse human takeover still works against dust, real friction, and an unpredictable crew is the next test the paper cannot run.