Drone-based visual inspection has emerged as a crucial manner for infrastructure
inspection, owing to its mobility and potential for automated perception. However, from the
perspective of human-agent interaction, the predominant modes in this field currently involve either
full-process human intervention or end-to-end execution based on deep learning. The former
constrains the level of automation, while the latter overlooks considerations of interactivity and
controllability. To improve the level of automation and interactivity of the inspection process, this
research explores the integration of Large Language Models (LLMs) into the visual inspection of
infrastructure to utilize the capabilities of LLMs to understand human intentions and generate
control commands. Specifically, high-level function libraries for drone and sensor control, as well
as comprehensive and standardized prompts, are developed to fulfill the objective. The effectiveness
of the method is demonstrated in both simulated and laboratory environments.
