LLM-Guided Exploration for Sample-Efficient UAV Navigation

Autonomous navigation of unmanned aerial vehicles in unknown and cluttered environments remains challenging due to inefficient exploration and high sample complexity. While Large Language Models (LLMs) offer strong commonsense and spatial reasoning, they are not suited for real-time continuous control because of limited action precision and inference latency. To bridge this gap, we propose LLM-Guided Exploration, a hybrid training framework that leverages LLM reasoning to bootstrap the training of off-policy RL agents. Our approach utilizes the LLM as an initial supervisor, providing high-quality demonstrations to guide the agent through scenarios before gradually handing over control to the RL policy for fine-tuning. By decoupling high-level exploration reasoning from low-level motion execution, our framework enables structured and adaptive exploration without sacrificing control stability. Extensive experiments in diverse environments demonstrate that the proposed method significantly improves convergence speed, navigation success rate, and overall sample efficiency compared to existing exploration and navigation baselines.