A Decentralized Path Planning Model Based on Deep Reinforcement Learning

With the increasing usage of automated guided vehicles on a large scale,the scalability of traditional path planning algorithms are limited in the central control systems. To alleviate the pressure on the system, the agent should be equipped with the ability to make independent decisions. In this paper, we propose a decentralized path planning model based on deep reinforcement learning (DPPDRL). The structure of DPPDRL is inspired by human path planning manner and dynamic couple methods. The AGVs controlled by DPPDRL can make decision independently and adjust their path in real time based on their local observations. In addition, DPPDRL encourages agents to adopt more diverse policies (such as going back or go around) to avoid collisions and explore potential routes. From the simulation results, we demonstrate that DPPDRL achieves a dramatic improvement in training speed and scalability. The policies of obstacle avoidance can adapt the changes of environment and the number of agents. Finally, we experimentally validate the learned policies of DPPDRL are available to any number of AGVs with a reasonable distribution of tasks