A Deep Q-learning Approach for Bin Relocation in Robotic Compact Storage and Retrieval Systems

Abstract

Robotic compact storage and retrieval systems (RCS/RS) offer space-efficient storage by stacking bins densely and using robots to retrieve them via a grid-based system. While existing operating strategies give fix guidelines on how to store and relocate blocking bins, the literature lacks learning-based strategies. This work closes that gap by applying deep reinforcement learning to optimize bin retrieval and relocation with respect to cycle time. A Deep Q-Learning agent, trained using Double-DQN with prioritized experience replay in a simulated RCS/RS, is evaluated across diverse scenarios. Results show performance gains regarding the cycle time of up to 36.98% over existing operating strategies. These findings demonstrate the potential of reinforcement learning for relocation decisions and suggest promising transferability to real-world systems.