Offline RL Optimizes Warehouse SLAM Throughput and Efficiency

Managing throughput in large-scale warehouse operations, particularly for Scan/Label/Apply/Manifest (SLAM) processes, is crucial for efficiency and avoiding congestion. Traditional control methods often struggle to adapt to dynamic conditions, leading to suboptimal performance. This research introduces an offline reinforcement learning (RL) framework designed to intelligently manage SLAM throughput. The system learns from de-identified historical warehouse operational data to recommend optimal throughput settings. It balances maximizing output with maintaining downstream stability by adaptively adjusting throttling behavior. The framework incorporates a history-informed state representation, an abstracted action space for delayed-impact control, and a reward function that considers both upstream and downstream operational metrics. Evaluated using various model-free and model-based strategies, the CQL policy within this framework consistently outperformed alternatives, improving overall system health by 22.97% and reducing average throttling duration by 3.18%. This demonstrates the potential of offline RL for robust and scalable warehouse control.