Learning Object-Centric Motion Priors from Human for Robotic Dexterous Manipulation

This research paper introduces a method for robotic dexterous manipulation that learns object-centric motion priors from human-object interaction datasets. Instead of solely imitating human demonstrations, the approach holistically predicts future hand-object states to serve as a general-purpose reward term for reinforcement learning. This reduces reliance on task-specific reward engineering and enhances generalization across tasks. The method was tested in simulation and the real world on three manipulation tasks, outperforming existing state-of-the-art methods in success rate and generalizability on novel objects. Additionally, the study validates cross-embodiment compatibility by deploying the learned skills on different robot hands.