A learning-based control approach for blind quadrupedal locomotion with guided-DRL and hierarchical-DRL

Abstract

Control parameters play an important role on the locomotion performance of quadruped robot system. In this paper, a learning-based control method is proposed, where the parameters of controller are learned by deep reinforcement learning (DRL). The proposed control system consists of a hierarchical controller and an agent, in which the agent learns the parameters in the upper layer of the controller. In the learning process, the guided-DRL and the hierarchical-DRL were used to solve the exploration problem and reward sparse problem, respectively. The former learns good expert trajectories with controller, which can realize the transition from supervised learning to self-learning, whereas the latter divides the task into a series of sub-goals to ensure that the difficulty of each subtask matches the decision-making ability of the agent in the learning stage. Finally, the controller with the trained policy is deployed to a real robot without manual tuning, and experimental results prove the effectiveness of the proposed method.