Reproduce “Expressive Whole-Body Control for Humanoid Robots” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xuxin Cheng, Yandong Ji, Junming Chen, Ruihan Yang, Ge Yang, Xiaolong Wang
Researchers at UC San Diego developed Expressive Whole-Body Control (ExBody), a framework that enables humanoid robots to generate diverse and expressive whole-body motions by selectively imitating human motion capture data. The method combines robust lower-body locomotion with upper-body expressiveness, successfully transferring various human gestures and dances to a real Unitree H1 robot and demonstrating superior stability and expressivity compared to traditional imitation techniques.
Key insights
- A relaxed imitation strategy, prioritizing upper-body expressiveness while maintaining robust lower-body locomotion, is essential for successfully transferring diverse human motions to real humanoid robots with hardware constraints.
- Utilizing large-scale human motion data for diverse state initialization (Guiding State Initialization - RSI) and deriving root movement goals directly from this data is crucial for efficient and robust deep reinforcement learning of complex humanoid behaviors.
- Direct full-body imitation from human motion capture data or naive application of adversarial motion prior (AMP) methods often leads to unstable or infeasible control on physical humanoid robots, highlighting the need for specific regularization and adaptive control strategies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.