Method for collision avoidance based on deep reinforcement learning with path-speed control for an autonomous ship

Abstract

In this paper, we propose a collision avoidance method based on deep reinforcement learning (DRL) that simultaneously controls the path and speed of a ship. The DRL is actively applied in machine control and artificial intelligence. To verify the proposed method, we applied it to the Imazu problem. It provides benchmark scenarios for collision avoidance. In particular, we compared and analyzed the collision avoidance performance according to the level of learning and various parameters to ensure that the proposed method displays optimal avoidance performance. The results indicated that the proposed method can determine a safe avoidance path for a given situation. Finally, to compare the performance of the proposed method, we compared the collision avoidance method based on the path–speed control of the OS proposed in this study with the collision avoidance method that controls only the path of the OS (Chun et al., 2021). We observed that the proposed method failed in 6 out of 20 scenarios of the Imazu problem when only the path of the OS was controlled. However, it succeeded in collision avoidance in all the 20 scenarios when both path and speed were controlled simultaneously.