Self-Coordinating Multi-AGV Scheduling in Self-Organizing Manufacturing Shop Floor via Multi-Agent Reinforcement Learning

With the increasing automation and digitalization of cognitive manufacturing systems (CMS), efficient scheduling of automated guided vehicles (AGVs), which is an integral component of CMS, has become crucial for achieving high operational efficiency. Traditional rule-based and centralized control methods struggle to adapt to dynamic and uncertain shop floor environments. To address these challenges, a novel multi-agent reinforcement learning (MARL) to optimize AGV coordination in next-generation self-organizing manufacturing systems. Our method jointly coordinates order dispatching and path planning to achieve cost-effective and timely material delivery. The proposed MARL framework consists of three key components: the Order Dispatching Attention Network (ODAN), the Path Planning-Repositioning Dual-Action Network (PPRDAN), and the Centralized Value Network (CVN). Experimental results in grid-based manufacturing scenarios demonstrate significant improvements in on-time delivery rate, total travel distance, and waiting time compared to baseline models. This study highlights the potential of MARL for enhancing AGV coordination in intelligent manufacturing environments.