Self-Organizing Navigation System for Drones Based on Multi-Agent Reinforcement Learning

Efficient self-organized navigation of Unmanned Aerial Vehicles (UAVs) in dynamic, unknown environments is crucial for enhancing their autonomy and collaborative capabilities in complex missions. This study introduces an innovative UAV navigation system based on multi-agent reinforcement learning (MARL) to overcome the limitations of traditional methods in handling environmental uncertainties and multi-UAV coordination. By employing a centralized training and decentralized execution framework, the proposed system enables individual UAV agents to learn effective decision-making and path planning from local observations without relying on a central control unit, thereby facilitating advanced autonomous navigation and coordinated task execution. Furthermore, the system’s decision-making performance is optimized through a carefully designed reward function, which elevates the overall efficacy of the multi-UAV network. Simulation results validate the feasibility and potential of the proposed approach.