Researchers at Jeonbuk National University linked a simulated highway emergency call to a four-drone response in the Gazebo simulator. Separate AI agents parsed the call, assigned drones to incidents, examined aerial images, wrote scene reports and routed audio summaries to rescue centers. The authors describe a feasibility demonstration and say larger-scale, real-world validation remains future work.
The system began with a simulated call transcript. A planning agent based on GPT-4o mini extracted the incident coordinates and mission intent. A coordination agent then assigned up to four autonomous drones to four simulated incidents. The drones used ArduPilot and MAVROS, with ArduPilot’s GUIDED mode providing target-directed movement commanded by companion software.
Each agent handled a different part of the response
At each assigned site, a custom-trained YOLOv11n model looked for accidents and fires in the drones’ simulated aerial imagery. When it detected an event, a fine-tuned BLIP-2 model turned the visual evidence into a structured incident description. If several drones reported the same accident, the system chose the highest-confidence view as its primary account and added distinct details from the other reports.
A dispatch agent used Piper text-to-speech to turn the summaries into audio alerts and route them to a rescue center based on drone proximity. The publisher’s abstract does not explain which drone’s location controls that choice when multiple reports are combined.
The demonstration has not yet tested a real rescue operation
The study connects functions that are often evaluated separately: call interpretation, drone coordination, accident detection, report fusion and dispatch. But its reported configuration ran in Gazebo, so it does not establish faster responses or better outcomes on an actual highway. The authors identify larger-scale and real-world validation as future work. A useful next evaluation would report end-to-end timing and missed or false alerts, and test the system under communication or image-quality failures before drawing conclusions about emergency use.
Sources
- Research paper, Scientific Reports, published Oct. 14, 2026
- ArduPilot documentation on GUIDED mode
- Illustrative CSIRO UAV photo and license information
Photo credit: CSIRO (photographer Stefan Hrabar), CC BY 3.0, via Wikimedia Commons. The image shows a CSIRO unmanned helicopter developed for remote-area weed mapping; it is illustrative and is not the system tested in the paper.