Learning a Push-Recovery Controller for Quadrupedal Robots

Reinforcement learning has been applied in the field of the legged robot controlling. However, tasks such as push-recovery often involve incompatible objectives, in which the general reinforcement learning framework does not take into account the hierarchical structure among the objectives. As a result, it is not easy to get a feasible solution. In this paper, we propose a Prioritized Hierarchical Reinforcement Learning (PHRL) framework that considers internal priorities and hierarchy of objectives in quadrupedal push-recovery. Our framework learns a controller that can resolve problems with incompatible objectives and manifest robustness when the robot is disturbed. We deploy the controller on a quadrupedal robot in simulation and show better results compared to baseline methods.