A Node-Based Nash-Inspired D3QN RL Framework for Traffic–Energy Coordinated AGV Scheduling in Smart Warehouses

Automated guided vehicle (AGV) scheduling in large warehouses becomes increasingly difficult as dense traffic and concentrated charging demands interact at shared intersections. Vehicle-centric multi-agent reinforcement learning (MARL) also faces increasing state dimensionality and non-stationarity as the fleet grows. This study therefore models controlled road-network nodes, rather than individual AGVs, as agents and develops a Nash-D3QN framework for traffic–energy coordinated scheduling. A* generates initial feasible paths, Nash-D3QN allocates intersection right-of-way using local and neighboring states, and a shared Petri-net layer enforces occupancy and safety constraints. State of charge (SOC), payload, start–stop losses, and charging queues are incorporated into the state, action, and reward designs to favor low-SOC and heavily loaded AGVs when necessary. In the 50-AGV scenario, Nash-D3QN shows numerically higher task completion rates than the compared baselines by 0.41–1.40 percentage points, together with 7.6–24.0% lower average waiting time, 4.7–46.0% higher SOC efficiency, and 8.38–32.42 percentage points higher coordination rates. Multi-seed analysis further evaluates task completion, coordination, SOC-based task efficiency, and waiting time at 50, 110, and 200 AGVs. With 200 AGVs, Nash-D3QN retains approximately 76.34% task completion, 2.68 tasks/SOC efficiency, and 86.17% coordination, indicating performance retention under larger tested system scales and operating loads.