InterMimicGen scales humanoid skills via motion-imitation flywheel

A new paper on Hugging Face, posted October 5 by eleven authors led by Yucheng Zhang and Sirui Xu, pitches humanoid loco-manipulation as a joint data-and-policy problem rather than a data problem alone. The framework, InterMimicGen, consolidates captured human-object interactions, retargets them to humanoids with dexterous hands, and then feeds a physics-based tracker that generates and filters its own training variants.

The motivating premise is stated plainly in the abstract: "Captured human-object interactions provide rich supervision for humanoid loco-manipulation, but they are sparse, heterogeneous, and not directly executable by robots." The authors' response is to introduce "a self-evolving motion-imitation framework in which robot motion data and a tracking policy improve each other."

The loop itself is three-part. Human-object clips are first retargeted onto the robot while preserving whole-body coordination and hand-object contact. A single generalist tracker then learns to execute those references in simulation. Finally, task-preserving edits are applied to the interactions, the tracker is fine-tuned on variants, and only runs that complete the task are retained to seed the next round.

The abstract does not publish per-task success rates or name specific humanoid hardware, and the framing is qualitative throughout: broader coverage, more diversity, continually expanding executable motions. It lands into a run of robot-learning work AI Weekly has been tracking this week, alongside efforts like RACE to stretch VLA policies further per action chunk.