Action-limited, Multimodal Deep Q Learning for AGV Fleet Route Planning

Abstract

The superiority of Automated Guided Vehicle (AGV) fleet management system is often reflected in the time-efficient on overall dispatch/navigation. The reinforcement learning can then be applied to help provide an optimal route planning for such fleet. In this study, in order to obtain suitable navigation strategies for certain specific momentary road conditions, we propose an improved deep Q network. It modifies the regression loss calculation method by bounding the Q output of certain actions, so that the network can focus on actions that are more in line with current road conditions. Moreover, multimodal deep Q learning is adopted to further improve fleet efficiency, owing to the help of multi-source monitoring data. Such learning collects action suggestions from each unimodal learning, and integrates their results through experience-based pooling calculations. The simulation results show the proposed method can optimize fleet management efficiency on time consumption level.