Designing an Adaptive AI System for Operation on Board the SpIRIT Nano-Satellite

Zaher Joukhadar, Jonathan Morgan, Christopher Bayliss, Miguel Ortiz del Castillo, Jack McRobbie, Robert Mearns, Krista A. Ehinger, Benjamin I. P. Rubinstein, Richard O. Sinnott, Michele Trenti, James Bailey

Abstract: This paper presents a detailed account of the adaptive AI system deployed on the recently launched SpIRIT nano-satellite. SpIRIT (Space Industry Responsive Intelligent Thermal) is a 6U CubeSat, an Australia-Italy mission led by the University of Melbourne. Launched in December 2023, SpIRIT is equipped with a payload designed to perform advanced computer vision experiments. The nano-satellite was launched with a pre-trained model for cloud detection, featuring the ability for onboard in-orbit fine-tuning. This paper highlights the software and system design choices made to address the unique challenges of running AI in space. The system is engineered for autonomy, robustness, and fault tolerance, allowing it to operate under the extreme conditions of space, including limited power, restricted communications, domain shifts, and constrained computing resources. The system consists of two distinct subsystems: one designed for space operations and the other for Earth-based activities. The space-bound subsystem handles inference and onboard fine-tuning of the model, while the Earth-based subsystem provides oversight and manages ground-truth labeling through a custom-built software called the Ground Truth Factory.