Navigation‐GPT: A Robust and Adaptive Framework Utilizing Large Language Models for Navigation Applications

ABSTRACT Intelligent navigation decision support systems are crucial for maritime safety, yet these systems frequently exhibit limited adaptability and reliability in novel, non‐predefined scenarios, constituting a persistent challenge. This study proposes Navigation‐GPT, a dual‐core large language model (LLM) agent designed for intelligent marine navigation. The framework leverages the strong generalization capability of LLMs in unfamiliar situations. It employs a large‐scale LLM with ReAct prompting as its control core, responsible for task parsing, planning, and orchestrating external tools to mitigate hallucinations. Furthermore, we fine‐tune a lightweight LLM in two stages: using LoRA and a novel rule‐controlled GRPO (RC‐GRPO) method to develop a specialized agent decision core. This core generates COLREGs‐compliant high‐level collision avoidance decisions, which are translated into dynamically feasible reference trajectories using a ship dynamics model formulated according to Fossen's equations. A PID‐based controller then tracks these trajectories to guide the ship through the resulting avoidance maneuver. Experimental results show that Navigation‐GPT completes the process from task reception to decision output in 11.13 s, remaining within the critical safety window for collision avoidance, though longer than the 0.73 s of traditional methods. In complex scenarios, it achieves an 86% collision avoidance success rate and a 90% behavioral compliance rate, outperforming its base model Qwen2.5‐7B by 38% and surpassing benchmarks including the dynamic window approach, artificial potential field, and other LLMs (Qwen2.5‐0.5B, Qwen2.5‐14B, DeepSeek, GPT‐4o). This work integrates LLM technology with traditional navigation systems, offering a comprehensive solution that enhances both safety and operational efficiency across diverse maritime scenarios.