Bimanual Imitation

Policy replay for a two-arm insertion task

Bimanual policy rollout · MuJoCo simulation

Overview

Coordinating two arms for insertion requires that their poses agree over time. This project brings task data, policy inference, and simulation replay together in a quad-insertion environment.

From data to rollout

Demonstrations are converted to a LeRobot-style dataset for ACT inference. The rollout shows the resulting sequence in MuJoCo, where the relative motion of the two end effectors can be inspected frame by frame.