Enabling Autonomy with a Deep Learning Framework for Planetary Exploration

Abstract

Future space missions will require enhanced perception and autonomy capabilities in the ground and flight segment to explore areas out of direct contact with the earth, intelligently process increasingly large volumes of sensor data, and make decisions independent of direct human oversight. Recent advances in deep learning for terrestrial applications have demonstrated that algorithmic approaches can provide a new level of perceptual understanding and data processing capability. The development and deployment of deep learning algorithms in space missions requires a holistic approach to user needs, data curation, model documentation, and production on flight hardware. This paper presents a deep learning framework that methodically captures these elements and presents two examples of its use for space flight missions, reprogramming a neural network deep learning model onboard a satellite in low-earth orbit and the first demonstration of deep learning on the lunar surface.