Obstacle Avoidance of Multiple UAVs Based on Reinforcement Learning

Improving the obstacle avoidance capability of Unmanned Aerial Vehicles (UAVs) is crucial for maintaining their operational safety. UAVs with autonomous driving ability is the development trend of future aircraft. This paper introduces a novel UAV obstacle avoidance approach utilizing the Artificial Potential Field-Dueling Deep Q-Network (APF-Dueling DQN) method. According to the dynamic model of UAVs, the three-dimensional dynamic equation of UAVs is established, and the motion space of UAVs is constructed by combining pitch and heading angles. In order to improve the obstacle avoidance performance of UAVs, an improved Deep Reinforcement Learning (DRL) algorithm is designed, and the algorithm is used to improve the reward potential function of DQN algorithm. Simulation studies demonstrate that the APF-Dueling DQN approach surpasses the traditional DQN in performance, exhibiting robustness against local minima and yielding efficient, smooth flight paths. This underscores the efficacy of the APF-Dueling DQN in addressing UAV path planning problem.