Ship collision avoidance method in starboard-to-starboard head-on situations

Among maritime incidents, ship collision accidents account for the majority. According to the report by International Maritime Organization, more than 80 percent of maritime accidents are attributed to inappropriate decisions due to the misunderstanding of the situations or the instructions by International Regulations for Preventing Collisions at Sea (COLREGs). Due to the complexity and variability of encounter situations, officers on board may have different understandings, and the resulting actions will also be different. Among the three encounter situations, the collision avoidance responsibility in crossing and overtaking situations is identified by defining stand-on and give-way ships. But in head-on situations, two ships share equal avoidance responsibility, i.e., both should take appropriate actions. Therefore, the coordination of the actions is crucial. As required by COLREGs, both ships turn starboard side. In port-to-port situations, such actions enlarge the distance and probably lead to the goal. However, in starboard-to-starboard situations, such actions may even cause a closer distance and a higher risk of collision. In this paper, Deep Q Network is adopted to solve the collision avoidance problem of two ships in starboard-to-starboard head-on situations. By setting different initial ship speed and different initial distances, the encounter scenarios are depicted in detail. Based on the instructions by COLREGs and the navigational experience, the set of collision avoidance actions and the reward functions are proposed, to determine the optimal collision avoidance actions of ships in specific situations. Besides, the coordination between the ships is also considered. The research results provide theoretical references for the autonomous collision avoidance scheme of ships in specific scenarios and can be applied to practical ship collision avoidance in the transition period from fully manned ships to intelligent unmanned ships.