We present BumbleBee (BB), an expert-generalist learning framework designed to achieve general agile whole-body control on humanoid robots. BB addresses the challenges of diverse motion demands and data conflicts by combining motion clustering and sim-to-real adaptation. Using an autoencoder-based clustering method, BB groups behaviorally similar motions and trains expert policies within each cluster. These experts are further refined with real-world data through iterative delta action modeling to bridge the sim-to-real gap. Finally, the expert policies are distilled into a unified generalist controller that maintains agility and robustness across all motion types. Extensive experiments on simulators and a real humanoid robot demonstrate that BB sets a new benchmark for agile, robust, and generalizable humanoid control. Real-robot videos are available on our [Website].
We will release our code soon.
If you find our work useful, please consider citing us!
@article{wang2025experts,
title={From Experts to a Generalist: Toward General Whole-Body Control for Humanoid Robots},
author={Yuxuan Wang and Ming Yang and Ziluo Ding and Yu Zhang and Weishuai Zeng and Xinrun Xu and Haobin Jiang and Zongqing Lu},
journal={arXiv preprint arXiv:2506.12779},
year={2025}
}