ExBody2 is a revolutionary system that allows humanoid robots to perform expressive motions with stability, similar to humans. The framework utilizes reinforcement learning in simulation to train the model before implementing it in the real world. This approach separates keypoint tracking from velocity control and uses a teacher policy to enhance mimic skills in the robot. The paper provides a detailed analysis of the design factors and showcases the superior performance of ExBody2 compared to existing methods. The team behind ExBody2 includes Mazeyu Ji, Xuanbin Peng, Fangchen Liu, Jialong Li, Ge Yang, Xuxin Cheng, and Xiaolong Wang.
https://exbody2.github.io/