Multi-Agent Deep Reinforcement Learning-Based Collaborative Robot Scheduling for Concurrent Processing Optimization in Semiconductor Cluster Tools

Scheduling semiconductor cluster tools is particularly challenging in concurrent processing environments, where wafers with diverse characteristics are processed simultaneously. Effective cooperative scheduling between the vacuum transfer module (VTM) robots and atmospheric transfer module (ATM) robots is essential because it directly impacts overall system throughput. However, traditional rule-based scheduling methods often degrade in performance under dynamic and complex conditions. To address this issue, this study proposes a multi-agent deep reinforcement learning (MADRL) framework in which the VTM and ATM robots are modeled as independent agents that jointly learn collaborative scheduling strategies without relying on expert-designed rules. Experimental results show that the learned scheduling strategies consistently outperform rule-based baselines-including heuristics commonly used in practice-in terms of throughput. Moreover, the proposed framework exhibits strong generalization capabilities under increased process complexity. These results highlight the potential of the MADRL-based approach as a rule-free, adaptive, and generalizable solution for optimizing scheduling in complex semiconductor manufacturing systems.