Task Selection by Autonomous Mobile Robots in A Warehouse Using Deep Reinforcement Learning

Abstract

We introduce a deep Q-network (DQN) based model that addresses the dispatching and routing problems for autonomous mobile robots. The DQN model is trained to dispatch a small fleet of robots to perform material handling tasks in a virtual, as well as, in an actual warehouse environment. Specifically, the DQN model is trained to dispatch an available robot to the closest task that will avoid or minimize encounters with other robots. Based on a discrete event simulation experiment, the DQN model outperforms the shortest travel distance rule in terms of avoiding traffic conflicts, improving the makespan for completing a set of tasks, and reducing the mean time in system for tasks.