Autonomous Obstacle Avoidance and Target Tracking of UAV Based on Deep Reinforcement Learning

When using deep reinforcement learning algorithm to complete Unmanned Aerial Vehicle (UAV) autonomous obstacle avoidance and target tracking tasks, there are often some problems such as slow convergence speed and low success rate. Therefore, this paper proposes a new deep reinforcement learning algorithm, namely Multiple Pools Twin Delay Deep Deterministic Policy Gradient (MPTD3) algorithm. Firstly, the state space and action space of UAV are established as continuous models, which is closer to engineering practice than discrete models. Then, multiple experience pools mechanism and gradient truncation are designed to improve the convergence of the algorithm. Furthermore, the generalization ability of the algorithm is obtained by giving UAV environmental perception ability. Experimental results verify the effectiveness of the proposed method.