Advancing space systems autonomy through onboard hardware-accelerated AI

The demand for autonomous space systems has increased significantly in recent years, driven by the growing volume of data produced by modern satellites and the need for real-time analysis and decision-making directly in orbit. While this abundance of data can greatly enhance the scientific and operational value of space missions, it also introduces major challenges related to onboard processing, storage, and transmission, particularly under the strict size, weight, and power constraints typical of small satellites. This thesis investigates the integration of artificial intelligence (AI) methodologies at the edge, onboard satellites, to enable more responsive, reconfigurable, and intelligent space systems. Edge AI represents a paradigm shift in how spaceborne data are managed, allowing satellites to autonomously filter, process, and prioritize information before downlink. In particular, the work explores how commercial-off-the-shelf system-on-chip technologies, especially those integrating field-programmable gate array (FPGA)-based accelerators, can support low-latency, low-power AI inference in orbit. Three key case studies are presented. First, two contributions address autonomous onboard decision-making for Earth Observation (EO) CubeSats through the implementation of deep learning (DL) techniques for onboard image processing and compression, aimed at reducing payload data transmission and optimizing bandwidth usage. Second, the thesis details a real mission case: the upcoming PRISMA Second Generation mission by the Italian Space Agency, which will deploy edge AI to optimize image acquisition strategies and provide onboard disaster detection capabilities for low-latency alert systems. Third, the study extends beyond EO and edge AI, demonstrating how computer vision techniques can be applied to analyze the massive datasets produced by modern ground-based telescopes, contributing to enhanced space situational awareness (SSA). The work addresses the “Edge AI Trilemma”, i.e., the trade-off between power, latency, and accuracy, and demonstrates that optimized, FPGA-accelerated DL models can be successfully deployed onboard without compromising mission constraints. Moreover, it shows how AI represents a powerful tool to advance system autonomy not only onboard but also in support of ground-based infrastructures, by strengthening SSA pipelines and thereby indirectly enhancing the autonomy, safety, and sustainability of orbital operations.