InterMimicGen: Scaling Humanoid Loco-Manipulation through Self-Evolving Motion Imitation
InterMimicGen turns sparse human-object motion captures into an expanding set of executable humanoid robot motions.
The paper describes a self-evolving imitation framework where retargeted human interaction data and a physics-based tracking policy improve each other. It consolidates heterogeneous motion-capture datasets, preserves whole-body and hand-object coordination, and trains one generalist tracker in simulation. Each augmentation round makes small task-preserving motion changes, keeps only successful simulated executions, and uses them for the next round. The authors report broader executable coverage over iterations, contact-preserving retargeting, and transfer to real robots. Source: HF Daily Papers' note.
The paper describes a self-evolving imitation framework where retargeted human interaction data and a physics-based tracking policy improve each other. It consolidates heterogeneous motion-capture datasets, preserves whole-body and hand-object coordination, and trains one generalist tracker in simulation. Each augmentation round makes small task-preserving motion changes, keeps only successful simulated executions, and uses them for the next round. The authors report broader executable coverage over iterations, contact-preserving retargeting, and transfer to real robots. Source: HF Daily Papers' note.
score 5