DQN-Based Adaptive Intersection Control for Warehouse AGVs in SUMO with a Proposed VLC Architecture

Congestion at shared intersections can limit the scalability of Automated Guided Vehicle (AGV) fleets in high-density warehouses. This study evaluates a Deep Q-Network (DQN) controller for adaptive intersection management in a SUMO-based warehouse model with five crossings, including the shared central crossing CP1. After one 200-episode training run, the DQN was evaluated in three seeded 6200 s simulations per demand scenario. Against a study-specific cyclic fixed-time controller at 250, 350, and 400 AGVs/h, the DQN achieved unweighted mean reductions of 62.9% in signal queue, 59.1% in average waiting time, and 18.0% in AGVs waiting to load, while loading/unloading occupancy changed by less than 1%. A matched 750 AGVs/h stress test showed that the benefit did not persist after saturation: at CP1, the mean signal queue increased from 13.15 to 30.68 AGVs and the mean number waiting to load from 9.64 to 15.93 AGVs. Sustained CP1 congestion began between 350 and 400 AGVs/h for the tested layout. Retained test outputs quantify test-level dispersion, but the single training run cannot estimate training variability. These results quantify the DQN traffic-control policy under ideal availability of SUMO state information and do not validate a VLC system. VLC is presented separately as a proposed future implementation layer; communication delay, packet loss, occlusion, and other physical-channel effects were not modelled.