Abstract

This paper enables real-world humanoid robots to maintain stability whileperforming expressive motions like humans do. We propose ExBody2, a generalizedwhole-body tracking framework that can take any reference motion inputs andcontrol the humanoid to mimic the motion. The model is trained in simulationwith Reinforcement Learning and then transferred to the real world. Itdecouples keypoint tracking with velocity control, and effectively leverages aprivileged teacher policy to distill precise mimic skills into the targetstudent policy, which enables high-fidelity replication of dynamic movementssuch as running, crouching, dancing, and other challenging motions. We presenta comprehensive qualitative and quantitative analysis of crucial design factorsin the paper. We conduct our experiments on two humanoid platforms anddemonstrate the superiority of our approach against state-of-the-arts,providing practical guidelines to pursue the extreme of whole-body control forhumanoid robots.