In-station UAV path planning based on multi-agent reinforcement learning and dynamic environment modeling

To address the problems of UAV path planning in complex railway station environment, such as difficulty in avoiding dynamic obstacles and low efficiency of multi-machine collaboration, this paper constructs a path planning method that integrates dynamic environment modeling and multi-agent reinforcement learning. This method builds the environmental infrastructure through coordinate system unification and regional division, and combines the sensor field of view model and the random walk probability model to achieve multi-scale motion prediction of dynamic obstacles. It uses the A* algorithm to generate the initial path, introduces a deep Q network for online optimization, and fuses it with a multi-objective reward mechanism to comprehensively optimize path length, energy consumption and obstacle avoidance effects. Experiments showed that when the number of obstacles was 5, the path search frequency reached 125 times/s. When there were 15 task points, the energy consumption per unit distance was 1.67 kJ/m. When the dynamic obstacles increased to 30, the collision rate was controlled at 12.1%. When the number of UAVs reached 20, there were only 6.2 path conflicts per mission. When there were 30 obstacles, the path adjustment delay was only 48 ms. When there were 20 mission points, the minimum energy consumption per unit distance was 0.85 kJ/m, and the mission completion rate still reached 88.2% when there were 20 UAVs. This method has excellent performance in planning efficiency, energy consumption control, safety and collaborative capabilities, and provides an effective solution for intelligent collaborative operations of drones in the station.