Obstacle Avoidance for Flight Safety of Unmanned Aerial Vehicles Using Deep Reinforcement Learning

This paper presents an autonomous flight algorithm for unmanned aerial vehicles (UAVs) that relies on depth camera vision data and integrates Voronoi diagrams for efficient path planning. To enable the UAV to autonomously choose suitable paths and adapt to changing environments and moving objects, a deep reinforcement learning approach is proposed. A buffered α-predicted-vector weighted Voronoi diagram (BαPVWVD) is created to divide the environment into distinct regions, with each region's boundary representing potential flight paths. Subsequently, a graph is dynamically generated to depict these flight regions and their connections. A graph-based deep reinforcement learning technique is used to train the UAV in selecting the optimal safe flight path. The experimental results indicate that our algorithm achieves comparable success rates to the baseline methods. Additionally, our algorithm demonstrates advantages in terms of flight distance, obstacle avoidance, and path selection, effectively enhancing the autonomy and efficiency of UAV flight.