Expressive Whole-Body Control for Humanoid Robots

• DOI: 10.48550/arxiv.2402.16796
• Publication Date: 2024-02-26
• Authors: Xuxin Cheng; Yandong Ji; Junming Chen; Ruihan Yang; Ge Yang; Xiaolong Wang

Can we enable humanoid robots to generate rich, diverse, and expressive motions in the real world? We propose learn a whole-body control policy on human-sized robot mimic human as realistic possible. To train such policy, leverage large-scale motion capture data from graphics community Reinforcement Learning framework. However, directly performing imitation learning with dataset would not work robot, given large gap degrees of freedom physical capabilities. Our method Expressive Whole-Body Control (Exbody) tackles this problem by encouraging upper body imitate reference motion, while relaxing constraint its two legs only requiring them follow velocity robustly. With training simulation Sim2Real transfer, our can walk different styles, shake hands humans, even dance world. conduct extensive studies comparisons diverse both world show effectiveness approach.