Overview of Meta-Reinforcement Learning Methods for Autonomous Landing Guidance

This paper presents a vision-based method for autonomous planetary powered descent and landing using meta-reinforcement learning. The goal is map observations to thrust commands directly using a deep neural network. Two iterations of the method are presented. First the 3-degrees-of-freedom powered descent pinpoint landing is solved using images, altitude and vertical rate as inputs. Then the problem of landing site selection, solved using a secondary neural network that performs semantic segmentation, is introduced. The final model is capable of autonomously select a landing area and guide the spacecraft to the designated landing site using only images as input.