UAV Agile Navigation Method for Unknown Environment via Deep Reinforcement Learning

This paper mainly considers the navigation problem of unmanned aerial vehicle (UAV) in an unknown environment. Traditional path planning method relies on accurate model parameters and environment maps, which has poor adaptability. Therefore, this paper adopts the deep reinforcement learning algorithm to accomplish the navigation task. The classical proximal policy optimization (PPO) algorithm lacks the perception of the correlation between UAV action and state makes difficult for the UAV to choose the optimal path, thus affecting the success rate and speed of navigation. To solve this problem, this paper adds a long short-term memory (LSTM) network to the policy and evaluation network of the PPO algorithm so that the UAV can refer to the preceding status and action information during path planning. The method is extended to three-dimensional motion space. Simulation results demonstrate that the LSTM-PPO algorithm designed in this paper can complete navigation tasks in unknown environments, and show stability in continuous state space and continuous action space. Meanwhile, compared with the PPO algorithm, the success rate of navigation and average arrival time is significantly improved.