Objective Existing unmanned surface vehicle (USV) path planning algorithms rely heavily on rule-based libraries and high-precision sensor inputs, making them ill-suited to “information-scarce” environments characterized by cognitive blind spots, perception loss, and semantic ambiguity, and thus unable to handle sudden unknown threats. This work proposes an intelligent agent framework based on large language models (LLMs) to investigate its feasibility and advantages for USV path planning under such information-scarce conditions.

Method The proposed intelligent agent framework utilizes a pre-trained LLM as its cognitive core, and is designed for high modularity and extensibility. The framework integrates a diverse library of tool functions (including sensor APIs for real-world data acquisition) and the human-computer interaction interface for natural language engagement, which significantly extends the LLM's perceptual and operational capabilities. To facilitate continuous learning and experience accumulation, the framework incorporates long-term and short-term memory mechanisms combined with distinct online and offline evolutionary strategies. This enables the agent to autonomously self-optimize through persistent interaction and accrued experience. Central to the framework is a highly structured prompting template. This template systematically organizes critical contextual information, including the agent's persona, environmental context, current states, sensor observations with associated confidence scores, task objectives, chain-of-thought (CoT) and ReAct prompts, prior knowledge, decision-making priorities, conditional logic rules and so on. This structured prompting effectively guides and constrains the LLM's cognitive processes, enabling it to perform complex tasks such as environmental perception, dynamic reasoning, interactive communication, and goal-directed path planning.

Results Experimental results demonstrate that, in environments with static obstacles, the proposed intelligent agent outperforms or is at least comparable to representative traditional algorithms such as A*, RRT, PSO, and Hybrid A* in key metrics including path length, number of turns, and average turning angle. The agent achieved a composite performance score of 1.00, significantly higher than those of the traditional algorithms (0.47, 0.47, 0.77, and 0.76, respectively). In dynamic environments, the LLM can effectively invoke tool functions, initiate human-in-the-loop interactions when appropriate, and perform real-time path planning based on sensor inputs. When faced with challenging conditions such as cognitive blind spots, perceptual gaps, and semantic ambiguity, the agent can adapt its planning strategy by leveraging confidence assessments, historical memory, and commonsense knowledge, demonstrating robust reasoning, generalization capabilities, and human-like intelligence.

Conclusion These findings indicate that the LLM-based path planning agent exhibits significant advantages in both simple obstacle avoidance and complex, information-scarce scenarios. This approach offers strong potential to advance USVs toward higher levels of autonomy.