Deep-Learned Autonomous Landing Site Discovery for a Tiltrotor Micro Aerial Vehicle

This paper addresses the need for autonomous Micro Aerial Vehicle (MAV) landing in unstructured environments. For aerial robotics to become truly ubiquitous, the requirement for operation over generic environments where a priori knowledge of landing site locations is unavailable has to be met. We propose the utilization of a Deep Learning-based classification framework deployed on aerial view image captured by the aerial robot, to perform the functionalities of object detection and classification below the aircraft, in order to autonomously characterize potential sites that are safe for landing. The proposed solution employs a small-sized companion computer equipped with a Neural Processing Unit, capable of handling both the high-level autonomy tasks, as well as executing the Neural Network pipeline onboard a MAV– class aircraft. We demonstrate the system’s effectiveness by deploying it onboard the MiniHawk-VTOL, a custom-developed hybrid flight envelope aerial robot with solar energy harvesting capabilities, designed to accommodate the aforementioned needs with the ultimate goal of enabling autonomous migratory missions (via anytime/anyplace land-to-recharge over unknown, unstructured environments).