UAV Path Planning Based on Deep Reinforcement Learning

In order to further study the path planning problem of unmanned aerial vehicle (UAV) in complex flight environment, this paper explores the application of traditional path planning algorithm in three-dimensional space. The existing methods often have defects such as insufficient flexibility, large amount of calculation, and easy to fall into the local minimum trap. Aiming at these problems, this paper designs a simulation experiment environment based on AirSim, and proposes a new algorithm PPO-PointNet network based on deep reinforcement learning. In this algorithm, a new neural network architecture is designed to extract three-dimensional features from 3D point clouds. Compared with the method of extracting two-dimensional features from RGB images, this method can capture spatial information more accurately, thus providing more accurate navigation and path planning instructions for UAVs. Experiments show that compared with baseline, the agent can spend less time, learn the best strategy to cross the doorframe along the ideal trajectory during the training process, and show the path planning ability close to the human driver.