Graph-Based Communication Optimization for Multi-Agent Reinforcement Learning in Unmanned Warehousing

—With the advancement of the industrial Internet and the ongoing intelligent transformation of manufacturing, multi-robot cooperative operations in unmanned warehouse systems face critical challenges in communication efficiency and real-time decision-making. Conventional path-planning algorithms are insufficient for cooperative scheduling in dynamic and complex environments, while existing multi-agent reinforcement learning (MARL)-based communication approaches often fail to determine appropriate communication targets or when to broadcast messages, resulting in excessive overhead and low efficiency. To address these limitations, this paper proposes a MARL-based communication optimization algorithm with graph representations. A graph-structured encoder is designed to intelligently select communication partners and optimize the communication topology. In addition, a graph information bottleneck mechanism is introduced to guide the graph neural network in learning minimally sufficient representations of communication messages. This mechanism maximizes the relevance of the representations to the cooperative task while minimizing dependence on the original communication graph, thereby enabling effective compression of redundant information. Experimental validation on a cooperative transportation task with warehouse robots in the robot operating system (ROS) and Gazebo simulation environment demonstrates that the proposed method reduces communication overhead by 79.0% and improves efficiency by a factor of 3.5, while maintaining a task success rate comparable to that of full-communication schemes. These results provide an efficient communication solution for large-scale multi-robot cooperative systems in industrial Internet scenarios.