InterMimicGen: Self-Evolving Motion Imitation Scales Humanoid Loco-Manipulation

Yucheng Zhang · hf · 2026-10-06

InterMimicGen: Scaling Humanoid Loco-Manipulation via Self-Evolving Imitation

Captured human-object interactions are rich supervision for humanoid loco-manipulation, but sparse, heterogeneous, and not directly executable. InterMimicGen is a self-evolving motion-imitation framework where robot motion data and a tracking policy improve each other:

Experiments show contact-preserving retargeting across robot configurations, broad tracking with a single generalist policy, growing executable motions over augmentation rounds, and real-robot transfer.

Original post →

More from Embodied

Embodied channel →