Reproduce “CoRe-MoE: Contrastive Reweighted Mixture of Experts for Multi-Terrain Humanoid Locomotion with Gait Adaptation” on OpenResearch

About this paper

Kailun Huang, Zikang Xie, Yanzhe Xie, Panpan Liao, Fanghai Zhang, Yanheng Mai, Wenhao Xu, Yunheng Wang and 3 more

The CoRe-MoE framework enables humanoid robots to perform smooth, automatic gait transitions between walking and running while adapting to diverse, complex terrains using a single vision-based policy. It demonstrated superior success rates, such as 94.3% on Up Stairs and a lower Cost of Transport of 0.61, achieving robust zero-shot sim-to-real transfer to a Unitree G1 robot.

Key insights

• A two-stage training strategy effectively decouples gait generation from terrain adaptation, mitigating gradient interference and distribution shifts for more stable learning.
• Employing a SwAV-based contrastive learning objective on the MoE gating network significantly enhances expert specialization by creating clearly separated clusters for different terrain types.
• Fusing residual terrain-aware actions from a vision-based MoE (Stage 2) with core gait actions from a proprioception-based MoE (Stage 1) enables fine-tuned adaptation while preserving fundamental gait stability.

Summary by alphaXiv.