Research on Target Capturing of UAV Circumnavigation Formation Based on Deep Reinforcement Learning

Abstract

In order to improve the adaptability and intelligence of multiple unmanned aerial vehicles (UAVs) in complex dynamic environment, the paper proposes an UAV circumnavigation formation target capturing method based on multi-agent deep reinforcement learning (MADDPG) algorithm. Firstly, the limitations of distributing expected capturing points by angular position are analyzed, and the capturing radius, angular spacing and angular velocity of circumnavigation are selected as the capturing control indicators of the UAV formation. Then the two control indicators of capturing radius and angular spacing in the polar coordinate system are transformed into one in the Cartesian coordinate system, and adopting the dynamic adaptive allocation strategy to complete the allocation of round-up points. Finally, the adversarial training is performed between the UAV formation and the target to be captured in the multi-agent particle environment (MPE). The results demonstrate that the UAVs can cooperate with each other in a dynamic confrontation environment, and can successfully round up escaping targets in a circumnavigation formation in pursuit and encirclement situations.