An Obstacle Avoidance Method Using Asynchronous Policy-based Deep Reinforcement Learning with Discrete Action

摘要

With the increasing application of mobile robots in manufacturing, service, and military fields, the demand of intelligent autonomous decision is also growing. In this paper, the state-of-the-art policy-based deep reinforcement learning (DRL) algorithm is applied to the mobile robot obstacle avoidance task. To solve the strong coupling in the existing DRL-based training process of mobile robot decision, an asynchronous decoupling architecture is proposed in this paper, which greatly improves the scalability and sample generation efficiency of the DRL algorithm. At the same time, based on the asynchronous decoupling architecture and Soft Actor-Critic (SAC) algorithm, we design an obstacle avoidance method named asynchronous dueling-based discrete action SAC (ADDSAC). Experiments show both action discretization and dueling network are simple and powerful techniques to improve obstacle avoidance performance. Finally, the reasons are also analyzed from different perspectives.