基于深度强化学习的未知环境下无人艇路径规划实时算法

For the Unmanned Surface Vehicle (USV) in unknown environment, the requirements of the adaptability and real-time are strongly demanding. To this end, this paper proposes a a path planning algorithm based on Deep Reinforcement Learning (DRL). For the request of plan-avoid-acclimate, on the basis of A3C, the proposed method optimizes net architecture, enriches navigation data and re-regulate the action space of the agent. Three kinds of maps are used for targeted training to improve the flexibility. By combining with the GPU platform, the pre-training data are collected with deep neural networks. In this way, the training efficiency is improved and the real-time requirement is guaranteed. Experimental results show that, in comparison with current methods, the training time reduces by 59.3% and the efficiency rises by more than 79.5%. Moreover, the performance of the trained model in unknown environment is effectively enhanced.