Meta-reinforcement learning-based collision avoidance for autonomous ship

Collision avoidance is critical for intelligent ship navigation. Ships encounter a variety of complex scenarios in real-world navigation environments, which requires improvements in the adaptability and effectiveness of collision avoidance policies. Therefore, we have innovatively proposed a meta-reinforcement learning method for solving ship collision avoidance. Inspired by meta-learning, we designed a two-layered recurrent model to enhance the adaptability and effectiveness of collision avoidance policies. Then, we created a task sampling method to train vessel agents in making collision avoidance decisions for high-risk encounter situations. The objective function and the policy gradient method for risk assessment are designed to enable vessel agents to thoroughly evaluate the risk situation of the current encounter scenario and optimize the collision avoidance policy. Lastly, we conducted simulation experiments to validate the feasibility of our work. The results indicate that collision avoidance policies outperform various comparative methods, exhibiting competitive advantages in adaptability, effectiveness, and safety in diverse encounter scenarios. Overall, our novel method provides a safer solution to enhance intelligent ship navigation.