Action Correction-Enhanced Multi-Agent Reinforcement Learning for Path Planning in Urban Environments

In urban environments, the path planning (PP) of unmanned aerial vehicles (UAVs) presents significant challenges, particularly since they are tasked with executing various operations in crowded areas. This scenario can be framed as a Multiple Traveling Salesman Problem (MTSP), where multiple drones must efficiently visit a set of target locations while ensuring safety and collision avoidance. The high density of obstacles, such as buildings, trees, and other aerial vehicles, increases the risk of collisions, making effective PP essential for operational safety. This paper proposes a two-stage PP approach to address these challenges. In the first stage, we introduce an improved Particle Swarm Optimization (PSO) algorithm for task allocation (TA), assigning each UAV a unique task queue to minimize the overall flight distance while ensuring efficient coverage of the target area. In the second stage, we employ a multi-agent reinforcement learning algorithm for PP and incorporate a safe action correction module that operates independently to adjust actions, thereby enhancing collision avoidance capabilities. Experimental results demonstrate that our approach reduces the probability of collisions with obstacles by 9% compared to the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm while also increasing the success rate of drone mission execution by 10%. This validates the effectiveness of our strategies for safe and efficient multi-drone operations in urban environments.