Year: 2024
Type: conference-paper
Abstract: Aiming at the problems of low search efficiency and slow convergence speed of deep reinforcement learning algorithms in complex environments, an unmanned vehicle path planning method based on improved deep Q-learning algorithm (DQN) is proposed. The optimal path is obtained by jointly adjusting the action space, action selection strategy and reward of the unmanned vehicle. In the exploration process, different selection probabilities are set for different actions according to the heuristic function in the A* algorithm instead of the traditional average random action selection, and a time-varying dynamic greedy strategy is used to adaptively balance exploration and utilization according to the environment and the learning effect of the neural network, to improve the global search capability and avoid falling into the local optimal solution; combining with the obstacle environment and the current position of the unmanned vehicle, a new reward function is designed in combination with the obstacle environment and current position of the unmanned vehicle, which improves the convergence speed of the algorithm and at the same time ensures collision-free avoidance. Simulation results show that under the same environment, this method is better than DQN, A* and RRT algorithms, with better real-time performance, convergence and safety. The feasibility of the improved algorithm is further demonstrated by porting it to ROS robots.
Authors: Siyuan Liang, Kedi Wang, Guodong Duan, Lei Cao
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4411730094