3D path planning of UAV based on improved reinforcement learning
Heng Dan, Hui Hua Peng · 2023
Aiming at the problem of slow convergence rate of reinforcement learning (RL), an improved RL algorithm based on A* algorithm and velocity obstacle (VO) is proposed. Firstly, a node path obtained by A* algorithm is used to initialize a part of Q table. Then, in the process of path planning by RL, the VO model is applied to output the minimum steering angle required by the obstacle-avoiding flight. The action which is greater than the angle will be taken to interact with the environment. Finally, in order to satisfy UAV performance constraints, B-spline curve approach is applied for path smoothing. The experiment results show that, compared with the standard Q-learning algorithm, the path length of improved RL algorithm is reduced by 26.37%, the success rate of algorithm planning path is increased by 30.76%, and the convergence round is advanced by 88.66%.