CTS-MoE: Implicit Terrain Adaptation via Mixture-of-Experts for Perceptive Locomotion

The article introduces CTS-MoE, a novel framework for perceptive legged locomotion that leverages a mixture-of-experts architecture to adapt to discontinuous terrain. It employs a dense mixture-of-experts actor with perception-based gating and a multi-critic structure with task-specific value heads, facilitating shared behaviors while minimizing value interference. Experimental results demonstrate that CTS-MoE outperforms traditional monolithic policies in terms of tracking accuracy and success rates on both seen and unseen terrains, offering a robust solution for adaptive locomotion in complex environments.