Research on Obstacle Avoidance Strategy of Grid Workspace Based on Deep Reinforcement Learning
Shuo Li, Jun Zhang, Bin Zheng · 2022 2nd Asia-Pacific Conference on Communications Technology and Computer Science (ACCTCS) · 2022
This paper proposes a multi-agent autonomous obstacle avoidance method based on the deep learning method. This method is used in grid workspace with strict position and motion constraints. The whole process uses Continuous Lidar data as the environment description to generate actions, and guides the agent to learn excellent strategies by continuously performing actions in the environment to obtain rewards. Based on the Proximal Policy Optimization(PPO) algorithm, we improved the design of the reward function, introduced density reward, distance reward and step penalty, which reduced the occurrence of congestion and improved the task efficiency of the agent. Finally, we conducted experiments in the simulation environment. Compared with the previous distributed methods, our method has smaller redundant paths and higher success rate of task completion. Compared with centralized methods, our method has advantages in terms of computational complexity, which can guarantee short planning times in large-scale environments.