Dynamic Multi-Agent Drone Path Planning Using Conflict-Based Search with YOLOv12

Abstract

The growing deployment of autonomous unmanned aerial vehicles (UAVs) in applications such as surveillance, logistics, and disaster response necessitates robust multi-agent navigation in dynamic environments. A critical challenge in such systems is ensuring collision-free path planning while adapting to unforeseen obstacles. This paper proposes a vision-integrated multi-agent path planning framework that combines Conflict-Based Search (CBS) with real-time object detection using YOLOv12, an attention-enhanced deep learning model. Initially, optimal paths are computed using A* and refined through CBS to resolve inter-agent conflicts. During execution, onboard vision systems detect dynamic obstacles using YOLOv12, which are incorporated into the environment model, triggering real-time replanning. The proposed approach is validated in a 3D simulation environment. Experimental results demonstrate that the framework achieves zero collision instances, maintains near-optimal path lengths (19, 19, 14 steps for three drones), and introduces only 0.005 seconds of additional computation time compared to standard CBS, while successfully adapting to dynamically appearing obstacles.