Objective To address the limitations in generalization capability and interpretability of decision outputs in current intelligent ship navigation systems, this study explores the application of large model technology. The research specifically targets open-sea autonomous collision avoidance tasks.

Method First, the methodology begins by extracting ship encounter scenarios from pre-processed AIS (Automatic Identification System) data. Using quantitative collision risk assessment and COLREGs-compliant encounter classification, representative scenarios are systematically sampled to build a diversified scenario library, providing diverse training samples for the large model. Next, a scene-instruction mapping module is introduced to convert scene data into descriptions, forming the foundation for the large model. By analyzing key components of ship navigation decision-making tasks, we propose a large language model (LLM) tailored for ship navigation decision tasks. This model employs a multi-turn dialogue design and incorporates a hierarchical reasoning framework, which progresses through three phases: cognition, analysis, and decision. This structure enables the model to handle navigation scenarios progressively from simpler to more complex tasks. Additionally, SeaPilot is capable of directly supporting the planning of subsequent driving actions and trajectory generation. Finally, both qualitative and quantitative experiments are conducted to validate the proposed method.

Results The model demonstrates strong comprehension of navigation scenarios, as evidenced by its performance in scenario-based QA tasks. In decision-making and trajectory planning, the model achieves an F1-score of 0.92 for direction action classification, an F1-score of 0.82 for speed action classification, and maintains trajectory planning errors below 10 meters. Additionally, the time consumption of the algorithm is calculated and analyzed. These quantitative results confirm the model's effectiveness in ship navigation decision-making and path planning.

Conclusion The results indicate that the proposed method provides a novel technical approach for advancing ship autonomous navigation.