InterMimicGen: Scaling Humanoid Loco-Manipulation via Self-Evolving Motion Imitation
Automatically prepared brief
InterMimicGen is a self-evolving framework that enhances humanoid robot motion imitation by iteratively improving motion data and tracking policies.
The research aims to address the challenge of scaling humanoid loco-manipulation using sparse and heterogeneous human-object interaction data that is not directly executable by robots.
The authors propose InterMimicGen, a framework where robot motion data and a tracking policy evolve together. They consolidate human-object interaction datasets, retarget them for humanoid robots, and train a physics-based tracker to execute these motions in simulation.
The evaluation involves iterative refinement of motion data and tracking policies, demonstrating improvements in motion quality and task completion across different robot configurations and real-world transfers.
About this brief
This article summarizes a public preprint abstract. Its findings have not been independently reproduced for this publication. Consult the paper for methods, evaluation scope and limitations. Automated briefs may contain errors; the original source takes precedence.