MoE-Loco: Mixture of Experts for Multitask Locomotion

What is worse, training a policy across multiple terrains with different gaits poses further challenges, leading I. INTRODUCTION to model divergence. Robots are often required to traverse diverse terrains and In this work, we enable a quadruped robot to traverse demonstrate various skills [1], [2]. Recent advancements in various terrains--including bars, pits, stairs, slopes, and baffles--while reinforcement learning (RL) algorithms and physics-based also supporting gait switching between bipedal simulators have enabled RL-based approaches to become the and quadrupedal modes, using only one policy.