To address the complex problem of cooperative collision avoidance among multiple intelligent ships in intersection waters, this paper proposes a decision-making method for multi-ship cooperative collision avoidance driven by enhanced large language models (LLMs).

By analyzing the navigational characteristics of intersection waters, this study formulates the multi-ship cooperative collision-avoidance problem as a partially observable Markov decision process (POMDP), thereby providing a formal mathematical foundation for decision-making. The cooperative process is decomposed into four causally linked modules—state perception, intent sharing, conflict coordination, and avoidance decision—to structure both information flow and reasoning. Guided by the embodied practices of human seafarers, a novel central-distributed dual-layer architecture is proposed. At the central layer, an LLM-based coordinator aggregates multi-ship situational data, applicable navigation rules, and conflict severity metrics to infer passage-priority sequences. At the distributed layer, individual ship agents leverage LLMs in combination with chain-of-thought prompt engineering to perform progressive, stepwise reasoning. These agents synthesize structured scene descriptions, coordination directives and retrieved navigation experience to generate executable avoidance maneuvers along with accompanying semantic explanations. To address known limitations of LLMs in precise numerical computation and continual learning, and to reduce potential hallucinations, the architecture incorporates two complementary augmentation mechanisms. A lightweight mathematical engine is employed to update kinematic states and compute deterministic conflict metrics, providing rigorous quantitative inputs to the reasoning pipeline. A retrieval-augmented generation (RAG) navigation knowledge base integrates a static corpus of navigation rules with a dynamic repository of historical scene−decision−evaluation tuples, enabling case-based grounding and continuous learning from past interactions. By embedding formal computation and evidence-based verification into the LLM reasoning loop, the proposed framework preserves the interpretive and interactive strengths of large models while ensuring verifiable, rule-compliant, and practically executable collision-avoidance decisions in complex intersection waters.

Simulation experiments demonstrate that the proposed enhanced LLM-driven method, implemented in DeepSeek-v3, achieves safe and efficient cooperative collision avoidance in typical intersection scenarios involving two, three, and four ships. The system maintains a minimum maneuvering speed exceeding 3 knots throughout the simulations and ensures a safety margin greater than twice the ship length.

This method advances the engineering application of LLMs in maritime decision-making and provides a new pathway for realizing highly autonomous shipboard artificial intelligence in complex operational environments.

To address the challenge of ineffective communication regarding avoidance intentions between autonomous ships （AS） and traditional manned ships (hereafter referred to as TS) in a mixed traffic environment, this study proposes an interactive collision avoidance decision-making method based on Stackelberg game (S-G) theory and Chain of Thought (COT). The aim is to enhance the interactive collision avoidance decision-making capabilities of ships operating in mixed environments.

First, ship collision avoidance scenarios in mixed environments are defined, and relevant research hypotheses are proposed. AS and TS ships are modeled using a leader-follower S-G game framework, with strategy spaces and payoff functions designed from a navigational practice perspective. Next, considering the interaction process between ships, a COT-game collision avoidance (COT-GCA) algorithm is developed, consisting of four sub-modules: state perception, intention sharing, strategy negotiation, and collision avoidance decision-making. Finally, the effectiveness of the proposed method is verified through experiments involving three-ship and four-ship encounter situations.

The experimental results demonstrate that ships in both groups can efficiently understand the avoidance intentions of other ships and successfully avoid collisions. The response time, steering range, and resumption of collision avoidance behavior exhibit timeliness, precision, and stability. The average output efficiency evaluation values before and after decision-making, calculated using the decision unit evaluation method, are 1 and 0.993, respectively, indicating the high efficiency of the S-G model in solving ship interaction collision avoidance problems.

The proposed model and algorithm effectively enhance the interactive collision avoidance decision-making capabilities of ships in mixed environments, providing significant theoretical insights for future practical applications.

This paper introduces a novel data-driven approach for generating realistic and hazardous overtaking scenarios. These scenarios are crucial for rigorously evaluating the autonomous collision avoidance capabilities of autonomous ships. Existing methods often struggle to balance scenario diversity, realism, and the representation of hazardous situations. To overcome this limitation, our method leverages the rich information embedded in automatic identification system (AIS) data to generate diverse and realistic overtaking encounters.

Specifically, we propose a hybrid model that integrates a sequence generative adversarial network (SeqGAN) with a self-attention mechanism (SAM). The SeqGAN captures the complex patterns and dynamics in AIS-based ship trajectories, enabling the generation of novel, yet plausible, overtaking maneuvers. The incorporation of a SAM further enhances the model's ability to capture long-range dependencies in ship trajectories, resulting in more realistic and nuanced simulations. To ensure that the generated scenarios accurately reflect hazardous situations, we have developed a constraint model based on longitudinal and lateral safety distances between vessels to define realistic initial conditions. This model dynamically adjusts the initial positions and velocities of both the target vessel and the autonomous ship under test, ensuring that each generated scenario presents a genuine collision risk.

The results show that the effectiveness of our approach is validated through extensive simulations. A total of 500 hazardous overtaking scenarios were generated, significantly improving the coverage of test scenarios. Notably, 97.3% of these generated trajectories fall within a predefined buffer zone that encompasses real-world trajectories, demonstrating the high fidelity of our model. Furthermore, the speed distributions of the generated target vessels closely match those observed in real-world AIS data, further validating the realism of our approach.

The enhanced realism and diversity of scenarios generated by this method significantly improve the efficiency of autonomous collision avoidance testing. This allows for a more precise definition of safety performance boundaries and accelerates the development and optimization of autonomous collision avoidance algorithms. Ultimately, this work contributes to the development of safer and more reliable autonomous maritime systems capable of navigating the complexities of modern maritime environments.