Decentralized Multi-Uav Target Encirclement Under Communication Constraints Using Multi-Agent Deep Reinforcement Learning

The increasing demand for multi-UAV collaborative systems in real-world scenarios highlight the limitations of conventional approaches, which mostly assumes centralized coordination or interference-free communication. This paper investigates the problem of decentralized multi-UAV target encirclement in a battlefield-like environment characterized by communication constraints, such as limited range and the absence of a central controller. To address these challenges, we propose an approach based on Multi-Agent Deep Reinforcement Learning (MADRL) with LSTM networks, enabling UAVs to achieve effective path planning and state prediction in obstacle-rich environments. Each UAV is equipped with exploration capabilities, utilizing simulated multi-line LiDAR sensors to perceive the environment and navigate in real time. Simulation experiments are conducted in an environment with obstacles; and they also validate the effectiveness of the proposed algorithm in achieving coordinated target encirclement under constrained conditions.