Unified Planning and Reinforcement Learning Control for Quadruped Robots in Off-Road Environments
Keywords: Off-road Autonomy; Trajectory Planning; Reinforcement Learning
Abstract: Trajectory planning for quadrupedal robots in complex unknown environments is an extremely challenging task due to the need to maintain balance, stability, and safe interaction with unstructured terrain while navigating efficiently. Existing methods often decouple planning and control, relying on computationally expensive environmental representations, or struggling with non-convergence in intricate scenarios. This paper presents a novel reinforcement learning(RL)-based approach that tightly integrates spatio-temporal trajectory planning and control for quadrupedal robots. The proposed method is validated on a RL-based locomotion controller which is tailored to challenging terrains. By unifying optimization-based planning and RL-based control in a unified framework, the quadrupedal robot can execute tasks intelligently while making real-time adjustments based on environmental feedback, resulting in improved overall performance and robustness. The proposed framework paves the way for robust and efficient navigation of legged robots in complex, unstructured, and off-road environments.
Submission Number: 19
Loading