Autoregressive DRL for Multi-Robot Scheduling in Semiconductor Cluster Tools

Deep reinforcement learning (DRL) has been widely applied to multi-robot scheduling, including semiconductor cluster tools, where maximizing throughput is critical. These tools operate under strict constraints, requiring precise coordination for efficient operation. However, managing multiple robots in such complex environments remains challenging. We propose an autoregressive DRL framework that sequentially generates robot actions using dynamic action masking, enabling context-aware decision making in large discrete action spaces. The agent is guided by a reward function that promotes step-wise progress, task completion, and reduced travel distance, encouraging efficient use of resources. Our approach demonstrates strong performance across representative tool configurations, highlighting the value of policy decomposition in complex scheduling tasks.