Deep Reinforcement Learning for Robust USVs Navigation in Diverse Environmental Scenarios

Collision avoidance is essential for the safe operation of unmanned surface vehicles (USVs) in marine environments. While existing studies have addressed USV collaboration and navigation, they often overlook environmental challenges. In this research, we develop a novel deep reinforcement learning model and train USVs to navigate safely under different situations, such as rain, wind, and their combined factors that can disrupt control and increase collision risk. We apply Generative Adversarial Imitation Learning (GAIL) and Proximal Policy Optimization (PPO) to improve the agent's performance using expert demonstrations. The experimental results demonstrate that the proposed framework significantly outperforms the baseline across multiple metrics: the maximum episode length increased from 15 to 350 steps, the cultivated reward improved by 158%, the extrinsic reward increased by 175%, and the number of collisions decreased by 75-78% across all environmental conditions. Moreover, the policy loss stabilized after 1.1 million training steps, confirming efficient convergence. These results show that our model performs better than the baseline model in mean reward, episode length, value estimates, and collision reduction. In the end, some concluding remarks and future directions are added by the authors.