Abstract
The collaborative aspect of drone swarms endangers smooth functioning of services and security of national facilities. Multi-Agent Deep Learning (DL) Coordinating swarms of drones in dynamic systems are not simple tasks, but learning is proving to be a legitimate solution. The paper introduces a new end-to-end UAV swarm intelligence system that combines DL and multi-agent reinforcement learning (MARL) to achieve autonomous and coordinated actions of drones. The system uses a new UAV swarm intelligence system that is based on YOLOv8 detection, DeepSORT-like tracking, and multi-agent PPO reinforcement. YOLOv8n model has the following performance: 0.91 precision, 0.83 F1-score, and 6.6 ms processing time per frame. The tracker is running at 151.5 FPS, ensuring the identities of the UAVs are the same throughout the movie. A specialized DroneSwarmEnv trains drones in formation control as well as in collision avoidance and cooperative navigation thus achieving an average reward of 1,886 with minimal collisions. In order to encourage generalization, real and synthetic UAVSwarm datasets are employed, consequently, training diversity and adaptability are multiplied. An in-depth analysis and visualization have indicated that the detection, tracking, and swarm behavior performance is outstanding and proves the policy convergence and policy stability. The system is of low weight, scalable, and is applicable for real-time deployment, thus providing a huge potential for uses such as autonomous surveillance, disaster monitoring, and aerial mission planning.