Hierarchical Reinforcement Learning for Quadruped Locomotion and Path Planning

Learning robust robotic legged locomotion control requires a combination of locomotion generation as well as adaptive behavior for environment changes. We apply a hierarchical reinforcement learning framework described in a paper by Jain et al. to train a multi-level policy with the single goal of reaching a target. By decomposing the task into two levels, we train the 12-DOF quadruped robot, Laikago, to walk along a curved path and reach the target. We show that the hierarchical policy structure is capable of naturally separating steering and general locomotion into the two levels and learning both simultaneously. (Github Source Code) (Poster)