CMoE: Contrastive Mixture of Experts for Motion Control and Terrain Adaptation of Humanoid Robots

Shihao Ma ¹ , Hongjin Chen ¹ , Zijun Xu ¹ , Yi Zhao ¹ , Ke Wu ¹ , Ruichen Yang ¹ , Leyao Zou ¹ , Zhongxue Gan ^1,† , Wenchao Ding ^1,†

¹ Fudan University

^† Corresponding Authors

Paper arXiv Video Code (Coming Soon)

Overview

Abstract

For effective deployment in real-world environments, humanoid robots must autonomously navigate a diverse range of complex terrains with abrupt transitions. While the Vanilla mixture of experts (MoE) framework is theoretically capable of modeling diverse terrain features, in practice, the gating network exhibits nearly uniform expert activations across different terrains, weakening the expert specialization and limiting the model’s expressive power. To address this limitation, we introduce CMoE, a novel single-stage reinforcement learning framework that integrates contrastive learning to refine expert activation distributions. By imposing contrastive constraints, CMoE maximizes the consistency of expert activations within the same terrain while minimizing their similarity across different terrains, thereby encouraging experts to specialize in distinct terrain types. We validated our approach on the Unitree G1 humanoid robot through a series of challenging experiments. Results demonstrate that CMoE enables the robot to traverse continuous steps up to 20 cm high and gaps up to 80 cm wide, while achieving robust and natural gait across diverse mixed terrains, surpassing the limits of existing methods. To support further research and foster community development, we will release our code publicly.

Framework

Our main contributions are summarized as follows:

We propose a single-stage, end-to-end framework that directly maps multimodal sensory inputs to robot actions, integrating estimators to process both historical information and terrain-specific data.
We introduce a novel MoE actor-critic model, augmented with a contrastive learning objective to enhance the robot’s ability to adapt its terrain response strategy across complex, heterogeneous surfaces.
Extensive experiments conducted on the Unitree-G1 robot demonstrate our approach’s state-of-the-art performance, and we release our code to further contribute to the research community.

BibTeX citation

If you find our work useful, please consider citing our paper:

@article{ma2026cmoe,
  title={CMoE: Contrastive Mixture of Experts for Motion Control and Terrain Adaptation of Humanoid Robots},
  author={Shihao Ma and Hongjin Chen and Zijun Xu and Yi Zhao and Ke Wu and Ruichen Yang and Leyao Zou and Zhongxue Gan and Wenchao Ding},
  journal={arXiv preprint arXiv:2603.03067},
  year={2026},
  url={https://arxiv.org/abs/2603.03067}
}