Published in Submitted to IEEE RAM and Under Review, 2023
The capability of a legged robot varies in line with its structure, topology, and locomotion controller. With different capability, legged robots can walk past terrains of varying complexity. However, existing legged robots’ motion controllers and path planners neither consider how to maximize the robot’s capability according to the current environment, nor give a global guidance path that meets the robot’s capability. In this letter, we propose a hierarchical learning based motion control and planning system for legged robots. It consists of three parts: capability maximizing, capability abstraction and capability based path planning. Capability maximizing uses reinforcement learning to choose best control strategy and robot topology which maximizes the robot’s traverse capability; capability abstraction network gives the robot’s capability under current terrain and footholds state; capability based path planning gives long-range guidance path that conforms our robots traverse capability learned by capability abstraction network. We trained our framework in simulation and did long-range locomotion experiments both in simulation and real world on our hexapod robot Qingzhui. The results shows that our method greatly improves the global locomotion performance of our legged robot.
Recommended citation: Changda Tian and Yue Gao. Learning Capability to Enhance Locomotion Control and Planning for Legged Robots. Submitted to IEEE RAM and under review, Sept 2023.
