Bimanual Lifting
Human-like bimanual object lifting and carrying for a Unitree G1 via residual AMP–PPO over a frozen locomotion policy.
Whole-body lifting is a coordination problem: a humanoid must bend its legs, lean its torso, close both arms around an object, and stand back up — without the arms fighting the balance controller. This project trains a Unitree G1 to squat, grasp, and carry objects bimanually in MuJoCo, structured as a residual policy over a frozen locomotion actor: the pretrained locomotion policy keeps the robot balanced and walking, while a residual network adds the manipulation behavior on top.
Style without imitation traps
An AMP discriminator trained on human motion capture acts as a style regularizer — but a capped one. The imitation reward is bounded so it can never outpay task completion: the policy is rewarded for lifting like a human, but only after it actually lifts. This ordering matters; uncapped style rewards routinely produce policies that pose beautifully and never pick anything up.
Training machinery
- Success-adaptive curriculum — a virtual “object assist” force helps early training and is annealed away as the success rate climbs.
- Asymmetric actor–critic — the critic sees privileged contact forces; the actor sees noisy object-state observations only.
- Domain randomization over object mass, friction, center of mass, and external base pushes.
- Reward-hacking whack-a-mole — successive exploits (lunging, waist-folding instead of squatting) were closed with gated posture penalties that activate only when the task is being cheated, not during legitimate motion.
Results (simulation)
Converged policies consistently hold objects at goal height, centered within 3 cm, with human-like posture — torso pitch stays inside the mocap prior’s 2–13° range. This project is simulation-only; no sim-to-real transfer was attempted.