Reproduce “Partial Motion Imitation for Learning Cart Pushing with Legged Manipulators” on OpenResearch

About this paper

Mili Das, Morgan Byrd, Donghoon Baek, Sehoon Ha

A partial imitation learning framework from Georgia Institute of Technology enables legged manipulators to stably push carts along various paths, outperforming traditional reinforcement learning and full-body imitation methods in simulation. The approach, which selectively applies an adversarial motion prior to lower-body kinematics, achieves a 94.4% survival rate and superior tracking accuracy during sim-to-sim transfer and under environmental stress.

Key insights

• Selectively applying imitation learning to only locomotion-relevant states with a partial AMP effectively transfers robust locomotion 'style' without overly constraining the manipulation arm's adaptability.
• Decoupling locomotion stability from manipulation flexibility results in more robust and generalizable loco-manipulation policies compared to full-body imitation or purely reward-based learning.
• Combining a reinforcement-learned low-level policy with a classical high-level path follower enables accurate and stable long-horizon path tracking for mobile manipulation tasks.

Summary by alphaXiv.