Title: CoRe-MoE: A Contrastive Reweighted Mixture of Experts Framework for Adaptable Humanoid Locomotion Across Diverse Terrains

Abstract: To navigate intricate environments efficiently, humans instinctively utilize walking and running without employing overly complicated movement strategies. In the same vein, humanoid robots must be capable of seamlessly shifting between these gaits while ensuring their movement remains both natural and stable. However, integrating gait transitions with multi-terrain adaptation into a single policy is difficult, largely due to gradient interference and distribution shifts caused by terrain-specific variations in dynamics and visual input. While Mixture-of-Experts (MoE) structures can help mitigate interference between multiple skills, standard joint training often results in a lack of distinct expert specialization, thereby reducing their utility.

To overcome these obstacles, we introduce CoRe-MoE, a two-stage reinforcement learning approach that separates gait generation from terrain adaptation. During the initial phase, the system learns a robust locomotion policy designed to generate natural walking and running motions with smooth transitions. In the subsequent phase, a terrain-aware MoE branch is incorporated and optimized using a contrastive objective. This process refines the gating network, allowing it to extract structured terrain representations and encourage clear expert specialization. The final motor output is derived through a weighted combination of the base gait policy and the terrain-aware branch, which enables the system to retain stable locomotion patterns while adjusting to complex surfaces.

Comprehensive simulations show that our method surpasses baseline techniques regarding success rates, locomotion stability, and adaptability across various terrains. Moreover, zero-shot implementation on a Unitree G1 humanoid robot confirms the framework’s practical efficacy. The robot demonstrated robust walking and running capabilities over stairs, slopes, steps, obstacles, and unstructured outdoor areas, successfully maintaining dynamic stability and precise foothold placement even when subjected to external disturbances.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC