Reinforcement Learning-Based Autonomous Collision Avoidance System for Maritime Vessels Under Dynamic Sea Conditions

Abstract

The paper presents an autonomous framework with reinforcement learning for collision avoidance between vessels working under dynamic sea conditions. The algorithm tabular Q-learning is implemented having a state representation, which is discretized, according to relative distance, bearing, relative velocity, time to closest point of approach, and encounter type. Rule-compliant reward shaping with safety supervision control at the higher layer guarantees safe and legal navigation. The MATLAB simulation environment is offered to evaluate system performance under varied traffic densities and environmental disturbances. The proposed method was compared against rule-based baseline and reinforcement learning model not integrated with regulations on indicators concerned with decreasing collision risks, shortening closest point of approach, improving separation distance, and overall performance index. I observed a significant improvement in both overall performance and separation behavior while keeping it within compliance with the navigation rules. However, collision risk reduction was highly dependent on traffic density.