ExBody

A 2024 UC San Diego method where a humanoid's upper body imitates human motion while the legs just track velocity steadily.

ExBody was released in February 2024 by Xiaolong Wang's group at UC San Diego, published at RSS 2024, and is one of the earlier works to apply large-scale human motion-capture data to whole-body control on a real humanoid robot. The difficulty is that humans and robots differ greatly in degrees of freedom and strength, so having the whole robot imitate a person frame by frame tends to make it fall over. ExBody's solution is to split the work: the upper body imitates the reference motion's joint angles and keypoints frame by frame, while the legs aren't required to match frame by frame at all, only to robustly follow the overall velocity and heading given by the reference motion. The data comes from about 780 clips in the CMU motion-capture database, retargeted onto the 19-degree-of-freedom Unitree H1, trained with massively parallel reinforcement learning in Isaac Gym, and then transferred to the real robot. Later work such as ExBody2 often uses it as a comparison baseline.

ExampleOn the real robot, the H1 can walk while performing motions in different styles at the same time — such as a zombie walk, high-fiving and shaking hands with a person, or even dancing together with someone.

Also called

Expressive Whole-Body Control for Humanoid Robots

Related

ExBody2 · Motion Tracking · Whole-Body Control · Motion Retargeting · Unitree H1 · Massively Parallel Reinforcement Learning

Sources

Expressive Whole-Body Control for Humanoid Robots (arXiv 2402.16796)
ExBody 项目页 (Chinese)

As of

2024-07