Reproduce “From Experts to a Generalist: Toward General 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
Yuxuan Wang, Ming Yang, Ziluo Ding, Yu Zhang, Weishuai Zeng, Xinrun Xu, Haobin Jiang, Zongqing Lu
A framework called BumbleBee (BB) enables robust, agile, and generalizable whole-body control for humanoid robots by clustering diverse human motions, training specialized expert policies with motion-specific sim-to-real adaptation, and distilling knowledge into a Transformer-based generalist. This approach achieves superior performance against state-of-the-art methods in simulation and demonstrates stable, agile real-world motion on a Unitree G1 robot.
Key insights
- Decomposing diverse motion control into specialized expert learning through meaningful semantic and kinematic clustering effectively mitigates data conflicts and improves training efficiency.
- Sim-to-real transfer is significantly enhanced by applying cluster-specific delta action models, accounting for varying dynamic features and simulator discrepancies across different motion types.
- A final distilled generalist policy, utilizing a Transformer architecture, can consolidate diverse expert knowledge and occasionally outperform individual experts by leveraging stable control behaviors learned across various skills.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.