Learning Robust and Flexible Locomotion of Wheel-Legged Quadruped Robots in Complex Terrains

Abstract

The wheel-legged quadruped robot, equipped with leg and end-wheel structures, possesses the capability to traverse continuous surfaces at relatively high speeds while also being able to navigate unstructured terrains. However, designing its controller using traditional methods presents significant challenges, particularly under conditions of limited or lost external environmental perception and highly variable terrain complexity. In light of this issue, this paper proposes a novel, concise, and effective reinforcement learning framework. The framework employs an asymmetric actor-critic structure incorporating a velocity estimation network and leverages multi-contact states generated by a central pattern generator for fusion, thereby training a single control policy to address the robust and flexible traversal of complex terrains by wheel-legged robots relying solely on an inertial measurement unit and joint sensors. Our method enables the modified Unitree Go1-based wheel-legged robot to traverse various challenging terrains, such as steps, high obstacles, rough terrain, and low-adhesion surfaces, while ensuring efficient locomotion performance on smooth and continuous surfaces. The effectiveness of the framework’s training results has been validated through testing in both simulation and real-world environments.