Abstract: our recent work proposes a robot system which contains three main components: 1. Data processing module for extracting human hand motion trajectory and object motions. 2. Retargeting module based on MANO hand model for mapping from human motions to robot motions. 3. Residual action learning with RL for learning how to refine imperfect motion priors. With the proposed system, robot can quickly learn how to complete complex dexterous manipulation task with one single human video demo.
Date:
Thursday, 19 June, 2025 - 13:00
Speaker:
Yucheng Xu
Affiliation:
University of Edinburgh
Location:
Informatics Forum. G.03