Abstract: Deep space missions will increasingly require spacecraft capable of operating autonomously in environments with communication delays and bandwidth limitations. A key enabler of such autonomy is the ability to run learning-based algorithms directly on board. In this work, we present a flight demonstration on a 6U CubeSat of training and inference of a segmentation model on a Jetson TX2 GPU in space on thermal images, and closed-loop attitude control using the inference model on a feed of real-time images to track the Moon. Apart from this demonstration, we implement ROS2 Foxy as a data transport middleware and integrate an open source file system that has strong guarantees around data corruption and repair. Over the course of one year, we performed 12 experiments using the spacecraft to demonstrate inference and training with an onboard NVIDIA GPU, as well as Moon tracking. This demonstration advances the state of the art in autonomous spacecraft systems by showing how modern learning-based pipelines, commercially available hardware, and open-source software can operate reliably under the compute and environmental constraints of Low Earth Orbit.

Sorina Lupu, Ph.D., is a Research Fellow at the Openmind Research Institute where she works on fundamental research in reinforcement learning, robotics, and space exploration. She has joint positions at NASA Jet Propulsion Laboratory and Caltech as a Visiting Researcher. She has a PhD from California Institute of Technology and is an affiliate of the Keck Institute of Space Sciences (KISS). She has master degrees from Caltech in Space Engineering and from the Swiss Federal Institute of Technology, Lausanne (EPFL) in Robotics and Autonomous Systems and holds a Bachelors of Science in Electronics Engineering from Polytechnic University of Bucharest, Romania. She worked previously at CERN (European Organization for Nuclear Research). She has won numerous awards and fellowships including the AI4Science Fellowship from Amazon and the Anita Borg Women Techmakers from Google. Her work was published in Science Robotics, IEEE Transactions of Robotics, etc.