Risk-Constrained Multi-Objective Deep Reinforcement Learning for AGV Path Planning in Rail Transit

Sensor-rich Automated Guided Vehicles (AGVs) are increasingly deployed in logistics, yet large fleets relying on fixed tracks face high maintenance costs and frequent route conflicts. This study targets rail-based material handling and proposes an end-to-end multi-AGV navigation pipeline under realistic operational constraints. A conflict-aware global planner, extended from the A* algorithm, generates feasible routes, while a multi-sensor perception stack integrates LiDAR and camera data to distinguish moving AGVs, static obstacles, and task targets. Based on this perception, a Deep Q-Network (DQN) policy with a tailored reward function enables real-time dynamic obstacle avoidance in complex traffic. Simulation results demonstrate that, compared with the Artificial Potential Field (APF) baseline, the proposed GG-DRL approach reduces collisions by ~70%, lowers planning time by 25–30%, shortens paths by 10–15%, and improves smoothness by 20–25%. On the Maze Benchmark Map, GG-DRL surpasses classical planners (e.g., RRT) and deep RL baselines (e.g., DDPG) in path quality, computation, and avoidance behavior, achieving an average path length of 81.12, computation time of 11.94 s, 5.2 avoidance maneuvers, and smoothness of 0.86. Robustness is maintained as a dynamic obstacles scale up to 30. These findings confirm that combining multi-sensor fusion with deep reinforcement learning enhances AGV safety, efficiency, and reliability, with broad potential for intelligent railway logistics.