Maintaining Formation While Evading Attacks and Obstacles... Soongsil University Develops AI for “Collision-Free Drone Swarms”
- Input
- 2026-11-17 13:18:24
- Updated
- 2026-11-17 13:18:24

[Financial News] Soongsil University said on the 17th that a team led by Professor Choi Dae-sun of its AI Safety Research Center had developed AI control technology for drone swarms that maintains formation without collisions even when faced with external attacks or obstacles. By reducing the computational burden on the small computers used in small drones by 99.7% compared with existing methods, the technology increased response speed and more than doubled the success rate of collision-free flights in virtual tests. In a virtual test in a complex forest environment, the research team deployed two attack drones to pursue the swarm. The collision-free flight rate for the swarm drones using the new technology was 51.7%, 28.7 percentage points higher than the 23.0% achieved with the existing AI control method. The average number of collisions per flight also fell sharply, from 2.77 to 0.82.
Until now, when multiple drones flew in a swarm, they had to calculate simultaneously how to evade sudden threats and how to restore their disrupted formation. This could overload the small computers inside the drones and slow their response.
The research team divided drone behavior into three stages—“normal flight,” “threat evasion” and “formation recovery”—and designed the system to focus solely on evasive maneuvers when danger arises. In particular, it changed the system so that the time-consuming calculations needed to restore the formation are performed separately only when necessary. As a result, it reduced the total number of calculations by about 99.7% compared with the existing method, which recalculated the formation at every moment.
With fewer calculations, the processing time for a drone to assess a situation and move fell to 0.0069 seconds (6.9 milliseconds). That is much faster than the benchmark of 0.01 seconds (10 milliseconds) for performing 100 control operations per second. In tests with the number of drones increased to 6–8, processing time also remained stable at around 0.0068–0.007 seconds.
The technology is expected to be useful when deploying multiple small drones, which cannot easily carry expensive, high-performance equipment, for search operations at disaster sites or facility inspections. Having demonstrated the technology’s performance in a virtual environment, the research team plans to apply it to real drones and assess their ability to respond to a range of unexpected situations.
The research, conducted as a project of the Institute for Information and Communications Technology Planning and Evaluation (IITP) with support from the Ministry of Science and ICT, was accepted for presentation at the international academic conference ESWEEK 2026 and was also published in a special issue of the international journal IEEE TCAD.
[email protected] Kim Man-gi Reporter