An agentic multi-drone framework for autonomous multi-accident detection, reasoning, and response

Rapid and effective response to highway accidents is critical for reducing response delays and improving situational awareness. In this study, we present an agentic multi-drone, end-to-end incident management framework, developed and validated in the Gazebo simulation environment, in which specialized planning, coordination, perception, scene-description, and dispatch agents communicate through ROS/MAVROS and MQTT interfaces. A simulated emergency call transcript is processed by a GPT-4o mini-based planning agent to extract incident coordinates and mission intent, which the coordination agent resolves into an optimal drone-to-incident allocation. Up to four autonomous drones, each managed via MAVROS and ArduPilot, navigate under GUIDED mode to their assigned sites, where a custom-trained YOLOv11n perception agent performs real-time accident and fire recognition on live aerial imagery. Detected events trigger a fine-tuned BLIP-2 description agent, which converts visual evidence into structured natural-language incident reports. When multiple reports correspond to the same incident, confidence-prioritized fusion selects the highest-confidence view as the primary description and supplements it with non-duplicated salient information from secondary views. A Piper TTS-based dispatch agent then converts these summaries into audio alerts and autonomously routes them to the nearest rescue center based on drone proximity. The description agent was fine-tuned on 1,783 image–caption pairs derived from the curated incident dataset, with reference captions generated using moondream2, holding its visual encoder and Q-Former frozen to preserve pre-trained feature representations, and evaluated on a held-out set of 254 samples using BLEU-4, ROUGE-L, METEOR, CIDEr, and SPICE. Results demonstrate the feasibility of the proposed modular agentic architecture for autonomous multi-incident perception, reasoning, and coordinated drone-based emergency response in the evaluated four-drone, four-incident simulation configuration. The framework provides a basis for future validation in larger-scale and real-world emergency-response scenarios.