Multi-UAV Autonomous Obstacle Avoidance Based on Reinforcement Learning

Obstacle avoidance is a necessary behaviour to ensure the safety of UAV. In this paper, aiming at the problem of autonomous learning and obstacle avoidance of multiple UAVs in the multi-obstacle map environment, an obstacle avoidance method of UAVs based on an improved reward deep Q learning network is proposed. According to the dynamics model of UAV, the three-dimensional dynamic equation is established, and the combination of pitch angle and heading angle constructs the action space of UAV. a new reward evaluation method, which adaptively adjusts the weight of the reward according to the distance between the UAV and the obstacle, is developed thus improving the performance of the UAV. Thus, the obstacle avoidance performance of multiple UAVs in unknown environment can be effectively improved. Finally, by comparing the traditional deep Q learning network, the simulation results show that the algorithm in this paper is reasonable, and UAVs can successfully achieve obstacle avoidance in unknown environment.