Agile Control For Quadruped Robot In Complex Environment Based on Deep Reinforcement Learning Method

Abstract

Using deep reinforcement learning for a quadruped robot to complete complex tasks will lead to a dimension explosion, which will make it is more difficult for designing and training the network. In this paper, a hierarchical learning (HL) framework based on distributed proximal policy optimization (DPPO) algorithm is proposed to find strategies for realizing agile control in complex environment. The framework is divided into a low-level gait generation network and a high-level environmental adaptation network. In the low-level network, some open-loop signals of different gaits is introduced to improve the training efficiency of the DPPO algorithm, which can generate a stable initial gait faster. The experimental results show that the learning effect of the proposed method is obviously better than that without a hierarchical network.