M.Sc. Thesis
2026.03.19 13:39
Seong-Won Choi, "An Improved Method for Ship Collision Avoidance Based on Deep Reinforcement Learning Using Attention Module and Mixture-of-Experts", M.Sc. Thesis, Seoul National University, 2026.02.25
조회 수 66
| 첨부 '1' |
|---|
| Abstract | In complex maritime environments, numerous ships operate simultaneously with overlapping routes, and the presence of confined waters, such as ports and coastal areas, increases the risk of collisions. In such environments, inappropriate maneuvers caused by human factors can lead to severe maritime accidents, including collisions and groundings. Therefore, autonomous navigation technology is essential for safe collision avoidance even in complex situations. Previous studies on traditional collision avoidance methods have explored rule-based methods such as RVO (Reciprocal Velocity Obstacles). However, in crowded environments with multiple ships, available maneuver options are limited, leading to excessively conservative decision-making. Furthermore, uncertainty due to sensor errors may degrade performance. To overcome these limitations, this study proposes an improved collision avoidance method based on DRL (Deep Reinforcement Learning) using the SAC (Soft Actor-Critic) algorithm. Previous studies using CPA (Closest Point of Approach)-based collision risk assessment often encountered numerical instability when the relative velocity between a TS (target ship) and OS (own ship) approached zero. Furthermore, processing multiple obstacle inputs using CNN (Convolutional Neural Network)-based methods was inefficient. Another issue involved discontinuities in action, as smooth transitions between path following and collision avoidance were not achieved. To address these challenges, this study introduces an AR (Approach Rate)-based collision risk assessment method to ensure numerical stability. The inputs from multiple obstacles are processed using an attention module, enabling the proposed method to efficiently reflect variable obstacle information and facilitate decision-making focused on critical data. Additionally, a MoE (Mixture of-Experts) structure is applied to flexibly manage transitions between path following and collision avoidance, thereby enhancing learning efficiency. The proposed method was validated through real-time simulations in a virtual environment and evaluated under conditions similar to those in actual sea areas by accounting for sensor error uncertainty. The validation results demonstrated that, compared with previous studies and the RVO method, the proposed method achieved a high collision avoidance success rate even in complex encounter situations and uncertain environments. Furthermore, it ensured robust, safe collision avoidance by maintaining a clearance distance during the approach to arrival. |
|---|---|
| Publication Date | 2026-02-25 |
