Autonomous Indoor Rescue Drone Guided by Sound and LLMs

In high-risk indoor emergencies such as fires, autonomous drones can play a vital role in accelerating rescue operations while reducing risk to human responders. This paper presents a sound-guided indoor rescue drone enhanced by Large Language Model (LLM)-based multimodal reasoning to enable adaptive perception and decision-making in GPS-denied environments. The proposed system integrates sound strength estimation, vision-based scene understanding, and LLM-driven navigation to guide the drone toward a target sound source while avoiding obstacles. The approach was validated using a real DJI Tello drone in indoor corridor environments. Experimental results show that the drone successfully navigated toward the target sound source in 75% of test scenarios. On average, 6–10 LLM-generated navigation decisions were required to localize the sound source. Each perception–decision cycle required approximately 20–30 seconds, primarily due to LLM inference latency. Despite this limitation, the system demonstrated robust multimodal reasoning and safe navigation without collisions in all successful trials.