ExBody2

A humanoid whole-body motion-tracking method from UC San Diego and others that lets the Unitree G1 walk, squat, and dance.

ExBody2 was released in December 2024 by Xiaolong Wang's group at UC San Diego together with UC Berkeley and MIT, an upgrade over the original ExBody. ExBody only had the upper body imitate a reference motion; ExBody2 switches to full-body tracking, mainly through three changes: first, automatic data filtering — a base policy is run over the CMU motion-capture data, and motions the robot can't achieve are filtered out by tracking error, balancing feasibility against diversity; second, teacher-student distillation, where a teacher policy uses privileged information (state unavailable on a real robot) in simulation, and is then distilled into a student policy that uses only onboard observations; and third, tracking keypoints in a local coordinate frame, decoupled from overall velocity tracking. The authors also found that training a general policy first and then fine-tuning it on a specific category of motion gives higher accuracy.

ExampleOn the Unitree G1, ExBody2 can dance the cha-cha following a reference motion; a specialized policy fine-tuned on dance data has noticeably lower joint-tracking error than either the general policy or one trained from scratch.

Also called

Advanced Expressive Humanoid Whole-Body Control

Related

ExBody · Motion Tracking · Whole-Body Control · Teacher-Student Distillation · Privileged Information · OmniH2O

Sources

ExBody2: Advanced Expressive Humanoid Whole-Body Control (arXiv 2412.13196)
ExBody2 项目页 (Chinese)

As of

2025-03