Learning to Assist Drone Landings

Unmanned aerial vehicles (UAVs) are often used for navigating dangerous terrains, however they difficult to pilot. Due complex input-output mapping schemes, limited perception, the system dynamics and need maintain a safe operation distance, novice pilots experience difficulties in performing landings obstacle filled environments. In this work we propose shared autonomy approach that assists perform on one of several elevated platforms at proficiency equal or greater than experienced pilots. Our consists two modules, perceptual module policy module. The compresses high dimensionality RGB-D images into latent vector trained with cross-modal variational auto-encoder. provides assistive control inputs reinforcement algorithm TD3. We conduct user study (n = 33) where participants land simulated drone without use assistant. Despite goal platform not being known assistant, all skill levels were able outperform while assisted task.