Motion Control of Wheeled-Legged Quadrupeds on Complex Terrains via Integrated Latent Encoding

Effectively controlling wheeled-legged robots on complex terrains remains an ongoing challenge in robotics. To enhance the motion performance of wheeled-legged quadrupeds on complex terrains, this paper proposes a control method that leverages integrated latent encoding and teacher-student networks within a privileged learning framework to achieve robust adaptability. By leveraging a privileged learning framework, the robot autonomously acquires robust motion skills and adaptability on complex terrains. A regionalized curriculum learning strategy is designed to ensure the robot can move at high speeds while adapting to complex terrains. Simulation and real-world platforms using IsaacGym, Webots, and a physical robot have been established for complex terrain motion tests of wheeled-legged quadrupeds. Experiments have demonstrated the efficacy of the proposed control method in enabling wheeled-legged quadrupeds to robustly traverse complex terrains such as rugged grasslands, stairs, and slopes.