CoRe-MoE: Contrastive Reweighted Mixture of Experts for Multi-Terrain Humanoid Locomotion with Gait Adaptation

Kailun Huang1,*,  Zikang Xie1,*,  Yanzhe Xie1,*,  Panpan Liao3,*,  Fanghai Zhang1,  Yanheng Mai1 Wenhao Xu1,  Yunheng Wang1,  Renjing Xu1,†,  Haohui Huang3,†,  Chenguang Yang2,† Fellow, IEEE 1 Hong Kong University of Science and Technology (Guangzhou) | 2 Hong Kong Polytechnic University 3 Guangdong University of Technology | * Equal contribution | † Corresponding authors
Paper arXiv Code(coming soon)

Abstract

Bringing humanoids into unstructured everyday scenes requires a single controller that combines natural walking and running driven by speed with reliable multi-terrain traversal. Existing systems can already produce multiple gaits, yet many rely on an explicit mode label for style selection, weakening the direct correspondence between speed commands and locomotion behavior. Selecting walking and running from commanded velocity provides a simpler control interface, but joint optimization with terrain adaptation can alter the learned gait styles. We present CoRe-MoE, a residual architecture anchored by a frozen adversarial motion prior (AMP) conditioned on velocity for walking and running. Terrain adaptation is learned only through a mixture of experts (MoE) branch driven by depth history, with contrastive prototype routing and action fusion specific to each body region. Under a unified retraining and evaluation protocol, CoRe-MoE produces stable bidirectional transitions between walking and running and achieves a 93.6% mean success rate across nine simulated terrain families, exceeding CMoE and MoRE by 1.1 and 2.3 percentage points, and deploys zero-shot on a Unitree G1 with onboard depth sensing.

Abstract figure

CoRe-MoE pipeline. Stage 1 learns a walking and running AMP prior conditioned on velocity. Stage 2 freezes πb and learns a residual MoE driven by depth history with contrastive prototype routing. Gains specific to each body region fuse ut and Δat before clipping and PD control.

BibTeX

@misc{huang2026coremoecontrastivereweightedmixture,
      title={CoRe-MoE: Contrastive Reweighted Mixture of Experts for Multi-Terrain Humanoid Locomotion with Gait Adaptation},
      author={Kailun Huang and Zikang Xie and Yanzhe Xie and Panpan Liao and Fanghai Zhang and Yanheng Mai and Wenhao Xu and Yunheng Wang and Renjing Xu and Haohui Huang and Chenguang Yang},
      year={2026},
      eprint={2606.04718},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2606.04718},
}