Hierarchical Safe and Robust UAV Path Planning Based on Large Language Models

Unmanned aerial vehicle (UAV) path planning is widely applied in fields such as Internet of Things (IoT) environmental monitoring, intelligent transportation, low-altitude logistics, and autonomous inspection. With the rapid development of UAV intelligence, natural language task planning powered by Large Language Models (LLMs) has emerged as a promising research hotspot. However, in safety-critical flight missions, LLMs suffer from issues such as unstable structured outputs, insufficient understanding of complex constraints, and unreliable responses to dangerous inputs, which can jeopardize the safety and robustness of UAV path planning. To address these challenges, this paper proposes an LLM-based hierarchical safe and robust UAV path planning framework to achieve co-optimization of natural language task understanding and deterministic path search. We design a Prompt–RAG–Validator safety enhancement mechanism that improves the LLM’s safety decision-making capabilities through prompt constraints, Retrieval-Augmented Generation (RAG), and rule validation, while constructing a structured constraint representation to convert natural language requirements into parsable planning constraints. Furthermore, we combine a grid cost map with the A* algorithm to achieve autonomous obstacle avoidance and safe grid-path generation. In addition, a two-stage evaluation system targeting both the constraint generation layer and the path-planning layer is constructed to verify model output reliability and actual planning performance. We validate the effectiveness of the proposed method across various complex scenarios, including multi-constraint tasks, risk-zone planning, incomplete input information, and prompt injection attacks. Notably, on the current template-derived benchmark, the deterministic regex/slot parser provided a stronger baseline with no dangerous outputs, whereas the LLM-based pipeline offers greater flexibility in handling diverse linguistic expressions.