Morphometric Imitation: From Morphology and Contact Aware Hand Retargeting to Sim-to-Real Visuomotor Policy
The paper reports an 89.3% zero-shot real-world success rate for robot-hand policies trained from retargeted human hand-object interactions.
The method converts reconstructed human hand motions into robot demonstrations through contact-preserving retargeting, residual RL, and policy distillation.
Across three robot hands and ten interactions, its retargeting step beat five baselines on contact F1 by at least 8 points for every hand.
The authors also report downstream dynamic retargeting gains of up to 35 points, with object pose and contact cues both helping in ablations.
Source: HF Daily Papers' note
The method converts reconstructed human hand motions into robot demonstrations through contact-preserving retargeting, residual RL, and policy distillation.
Across three robot hands and ten interactions, its retargeting step beat five baselines on contact F1 by at least 8 points for every hand.
The authors also report downstream dynamic retargeting gains of up to 35 points, with object pose and contact cues both helping in ablations.
Source: HF Daily Papers' note
score 5