Research on Multi-AGV Rail Transportation Path Planning and Dynamic Obstacle Avoidance Based on Multi-Sensor Environmental Perception and Deep Reinforcement Learning

Sensor-equipped Automated Guided Vehicles (AGVs) are widely used in material distribution and other logistics processes. Path planning, a critical component of rail transport material handling systems, ensures the efficiency and safety of AGV material transport. Traditional fixed-path AGV transport is costly to maintain and prone to collisions due to path conflicts in large-scale operations. Addressing these challenges, this study first considers the characteristics of real-world rail transport material handling environments. It utilizes the A* algorithm as the foundation to develop a path planning method for multiple AGVs, enabling global path planning for each AGV. Further, it employs advanced AGVs equipped with multiple sensors to collect and process sensor data, developing a multi-sensor feature fusion environmental perception algorithm. This algorithm effectively identifies other dynamic AGVs, static obstacles, and targets within rail transport environments. Additionally, this research constructs a deep reinforcement learning model based on Deep Q-Networks, designs a novel reward function, and, integrating environmental perception capabilities, achieves dynamic obstacle avoidance for complex multi-AGV path planning. Finally, a simulation environment was established to verify the applicability of the proposed path planning and dynamic obstacle avoidance methods for multiple AGVs in rail transportation scenarios, providing experimental support for railway logistics handling requirements. Benchmark tests demonstrated that the proposed GG-DRL method significantly outperforms the traditional Artificial Potential Field (APF) algorithm, with a 70% reduction in collisions, a 25%–30% decrease in planning time, a 10%–15% reduction in path length, and a 20%–25% decrease in root mean square (RMS) curvature. Further comparative experiments conducted using a standardized Maze Benchmark Map and dynamic obstacle scenarios indicated that the GG-DRL algorithm is markedly superior to both traditional path planning methods such as Rapidly-exploring Random Tree (RRT) and classical deep reinforcement learning approaches such as Deep Deterministic Policy Gradient (DDPG), in terms of path quality, computational efficiency, obstacle avoidance performance, and trajectory smoothness. Specifically, GG-DRL achieved the shortest average path length of 81.12 units, the fastest average computation time of 11.94 seconds, the fewest obstacle avoidance maneuvers with an average of 5.2, and the highest average trajectory smoothness of 0.86. In increasingly complex environments containing 10, 20, and 30 dynamic obstacles respectively, the GG-DRL algorithm consistently exhibited outstanding robustness, planning speed, and trajectory quality. These results confirm the effectiveness of multi-sensor fusion and deep reinforcement learning in enhancing the reliability, efficiency, and safety of AGV transportation, indicating broad application prospects in railway logistics and intelligent navigation.