Fine-tuning of large language models for obstacle circumvention in path planning via chain-of-thought reasoning

Autonomous path planning for aerial vehicles in complex environments is essential for flight safety and mission success. Practical systems must achieve both real-time obstacle avoidance and long-range navigation. However, existing methods show limited robustness, weak generalization, and low computational efficiency when facing dynamic obstacles, environmental uncertainty, and physical constraints. In this paper, we propose a two-stage path planning framework driven by a large language model (LLM). The framework exploits the spatial reasoning and generalization capability of LLMs. In the first stage, the model generates feasible detour paths directly from structured environment descriptions and task prompts. This process does not rely on traditional mapping or search modules. In the second stage, we construct high-quality chain-of-thought reasoning samples for obstacle circumvention. We perform supervised fine-tuning to improve spatial understanding and to stabilize path generation. This stage enables a smooth transition from a general-purpose model to a task-specific planner. Experimental results show clear performance gains in highly complex scenarios. The proposed method significantly improves obstacle circumvention success rate, path rationality, and adaptability to unknown environments. These results demonstrate the feasibility of using LLMs for structured spatial planning tasks. The method also provides a practical solution for intelligent aerial navigation under real-world constraints.