Related Research
Jaeger
PDF Paper arXiv arXiv github Project

B u m b l e B e e

From Experts to a Generalist:
Toward General Whole-Body Control for Humanoid Robots


PKU    BeingBeyond

PDF Paper arXiv arXiv github Code

From Experts to a Generalist:
Toward General Whole-Body Control for Humanoid Robots

Yuxuan Wang1,2*, Ming Yang1,2*, Ziluo Ding2*, Yu Zhang2*, Weishuai Zeng1, Xinrun Xu2,
Haobin Jiang1,2, Zongqing Lu1,2†


1Peking University     2BeingBeyond


*Equal Contribution   †Corresponding Author

Paper arXiv Project

Abstract

    Achieving general agile whole-body control on humanoid robots remains a major challenge due to diverse motion demands and data conflicts. While existing frameworks excel in training single motion-specific policies, they struggle to generalize across highly varied behaviors due to conflicting control requirements and mismatched data distributions. In this work, we propose BumbleBee (BB), an expert-generalist learning framework that combines motion clustering and sim-to-real adaptation to overcome these challenges. BB first leverages an autoencoder-based clustering method to group behaviorally similar motions using motion features and motion descriptions. Expert policies are then trained within each cluster and refined with real-world data through iterative delta action modeling to bridge the sim-to-real gap. Finally, these experts are distilled into a unified generalist controller that preserves agility and robustness across all motion types. Experiments on two simulators and a real humanoid robot demonstrate that BB achieves state-of-the-art general whole-body control, setting a new benchmark for agile, robust, and generalizable humanoid control in the real world.

B u m b l e B e e

BumbleBee

    Our approach begins with a data curation stage that combines motion retargeting with PHC-based filtering to construct a clean and transferable dataset. The curated motions are then organized through an autoencoder-driven clustering process, which groups samples according to both semantic and kinematic similarities. For each resulting cluster, we train expert policies through iterative delta fine-tuning, enabling them to specialize in the distinct motion patterns represented. In the final stage, these specialized experts are distilled into a unified general Whole-Body Control (WBC) policy by means of a Transformer-based architecture, thereby integrating diverse skills into a single robust model.

BibTeX

@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}
}