Multi-Agent Reinforcement Learning for Multi-UAV Pursuit with Full Planar Motion and a Limited Detectable Region

초록

Although previous studies have considered sensing constraints and UAV dynamics, most of them have used unrealistic sensing limitations and simplified dynamic models. Thus, these approaches can suffer from a significant discrepancy between simulation results and real-world deployment. To address this issue, this study incorporates high-fidelity sensing constraints and UAV dynamics into a multi-agent reinforcement learning approach, focusing on the practical interplay between FOV limitations and pursuit strategies. First, the proposed reward considers the sensing constraints via a gaze-alignment reward, which varies with the field-of-view condition, and a capturability reward that encourages transitions toward a capturable region. Second, realistic UAV dynamics, including lateral motion, forward motion, and yawing, are modeled in a simulation environment to reduce the sim-to-real gap. Quantitative evaluations demonstrated that the proposed formulation significantly improved the capture performance under diverse sensing conditions. The capturability reward increases the capture success rate by 11.4%. When the maximum speed of the evading UAV was 2 m/s faster than that of the pursuing UAVs, all capture trials failed when lateral motion was not used. However, when lateral motion was enabled, the success rate increased to 99.2%, highlighting the need for lateral motion.

키워드