A Robust and Learning Approach for Multi-Phase Aerial Search with UAVs

Abstract

Unmanned aerial vehicles (UAVs) have been attracting more and more attention in the research and industry field. Aerial search is a common mission and is intrinsically fit for UAVs, e.g. disaster rescue, remote sensing and environmental monitoring. With the improvement of UAV hardware and software, UAVs tend to achieve better autonomy and accomplish more complex tasks. However, current UAV aerial search is usually hardcoded, which limits their adaptability, autonomy and robustness in realistic scenarios. In this paper, we propose to address this problem by leveraging reinforcement learning (RL) and a recent control architecture, behavior trees (BTs). We develop robust and adaptive UAV systems that can automatically conduct multi-phase complex aerial search, including search, communication and refueling. Experimental results in a 3D robot simulator verify the effectiveness and robustness of the proposed approach, which achieves better performance than the baseline.