Learning to Solve Pod Retrieval as Sequential Decision Making Problem

The problem of pod retrieval in Robotic Mobile Fulfilment System (RMFS) is a key problem to improve the order picking efficiency. In such system, each robot needs to complete a set of retrieval requests, including bringing each pod from a retrieval location to a picking station and return the pod to a storage location. The objective is to minimize the total cost for each robot with all retrieval requests completed. In the previous literature, the problem was viewed as a static combinatorial optimization problem, which was commonly solved by heuristic methods. This kind of approachs often face with computational efficiency problems and are hard to satisfy the real-time requirement in complex real scenes. In this paper, we formulate the problem as a Markov Decision Process, a kind of Sequential Decision Making Problem, and then using Transformer with reinforcement learning to learn an efficient retrieval policy. The effectiveness of the method is verified by experiments.