Deep Reinforcement Learning-based policy for autonomous imaging planning of small celestial bodies mapping

This paper deals with the problem of mapping unknown small celestial bodies while autonomously navigating in their proximity with an optical camera. A Deep Reinforcement Learning (DRL) based planning policy is here proposed to increase the surface mapping efficiency with a smart autonomous selection of the images acquisition epochs. Two techniques are compared, Neural Fitted Q (NFQ) and Deep Q Network (DQN), and the trained policies are tested against benchmark policies over a wide range of different possible scenarios. Then, the compatibility with an on-board application is successfully verified, investigating the policy performance against navigation uncertainties.