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

Researchers introduce CoRe-MoE, a reinforcement learning framework enabling humanoid robots to seamlessly transition between walking and running while adapting to complex terrains. The two-stage approach decouples gait generation from terrain adaptation using a contrastive learning mechanism, with successful zero-shot deployment on a Unitree G1 robot across varied outdoor environments.

CoRe-MoE addresses a fundamental challenge in humanoid robotics: enabling unified locomotion policies that handle both gait transitions and terrain variability without sacrificing stability or natural movement patterns. Traditional Mixture-of-Experts architectures struggle with expert specialization when trained naively on multi-skill tasks, creating gradient interference that undermines performance. This research decouples the problem into two stages—first establishing a stable baseline gait policy for walking and running transitions, then introducing terrain-aware adaptation through contrastive learning on the gating network.

The technical innovation lies in how the framework preserves locomotion stability while enabling dynamic terrain adaptation. By using contrastive objectives to shape expert specialization and maintaining weighted fusion between the base policy and terrain branch, CoRe-MoE avoids the performance collapse that plagues multi-task learning in robotics. The approach directly addresses distribution shift caused by terrain-dependent visual and dynamic variations, a persistent obstacle in sim-to-real transfer.

For the robotics industry, this work demonstrates practical progress toward deployable humanoid systems capable of real-world navigation without task-specific retraining. Zero-shot deployment on physical hardware across stairs, slopes, obstacles, and unstructured terrain validates the robustness of the approach. This has implications for autonomous systems development, search-and-rescue applications, and industrial inspection tasks requiring adaptable locomotion.

Future developments should focus on extending this framework to additional locomotion skills, improving energy efficiency metrics, and scaling to more complex environmental interactions. The success on Unitree G1 suggests these methods may generalize across different humanoid platforms, establishing a foundation for more capable autonomous robots.

• →CoRe-MoE's two-stage framework successfully decouples gait generation from terrain adaptation, solving gradient interference in multi-skill learning
• →Contrastive learning enables clear expert specialization in Mixture-of-Experts architectures, improving multi-terrain adaptability
• →Zero-shot deployment on Unitree G1 demonstrates robust performance across diverse real-world terrains without retraining
• →The approach maintains natural locomotion patterns and stability while dynamically adjusting to environmental variations
• →Framework directly addresses sim-to-real transfer challenges through structured terrain representations and expert specialization