Deep Reinforcement Learning for Task Assignment and Shelf Reallocation in Smart Warehouses

Abstract

With the rapid development of online shopping and the prosperity of the e-commerce industry in recent years, traditional warehouses are struggling to cope with increasing order volumes. Accordingly, smart warehouses have gained considerable attention for their relatively high efficiency and productivity. In such warehouses, robots transport shelves to picking stations on the basis of tasks assigned to them and then return to the inventory area. An accurate task assignment method must be developed to achieve high efficiency in smart warehouses; however, existing task assignment methods use limited information, resulting in a lack of insight regarding future tasks in warehouses. This paper proposes a method based on the deep Q-network (DQN) that considers inventory for task assignments. The developed DQN-based model determines shelf return locations on the basis of current states to improve warehouse performance. The proposed method was compared with a traditional task assignment method, namely regret and marginal-cost based task assignment algorithm (RMCA); the results indicated that compared with the RMCA method, the proposed approach is more efficient and faster and can accommodate more robots.