Feasibility of a COTS GPU-Based Onboard AI Platform for Robotic Space Missions

Kurzfassung

Future robotic space missions demand advanced on-board artificial intelligence (AI) capabilities to support autonomy in servicing, assembly, and planetary exploration. Traditional space-qualified processors and field programmable gate arrays (FPGAs), while reliable, impose long development cycles and high costs, limiting their suitability for rapid mission preparation. At the same time, commercial-off-the-shelf (COTS) GPUs have become central to terrestrial robotics, offering high parallel performance and mature software ecosystems. This paper investigates the integration of COTS GPUs into the Scalable On-board Computing for Space Avionics (ScOSA) architecture, a hybrid platform combining radiation-hardened reliable computing nodes with high-performance nodes. By leveraging ScOSA's fault-tolerant middleware and established machine learning toolchains, we demonstrate how GPU-enabled nodes can bridge the gap between laboratory AI prototypes and reliable in-orbit execution. The approach enables faster iteration cycles, reduces time-to-orbit, and provides scalable performance across mission classes.