AI/ML Inference Engine Software for High-Reliability applications on Space Qualifiable Hardware

This paper focuses on the need for increased autonomy in spacecraft operations and the role of advanced sensors and computational power in enabling autonomous navigation for space exploration, in-orbit servicing, and space debris management. It highlights the emergence of new-generation spacecraft equipped with improved sensors capable of generating high-quality data at unprecedented rates, alongside enhanced computational capabilities. The study also emphasizes the importance of complying with ECSS standards to ensure the reliability, safety, and performance of AI algorithms employed in space systems, while meeting specific space environment requirements and constraints. The presented research builds upon the previous work conducted by Klepsydra with support from ESA in the field of AI for flight software, particularly the KATESU project. This work involved evaluating OBPMark-ML Deep Neural Network models on Teledyne e2v’s LS1046 and the ZedBoard (Xilinx ZYNQ 7000) onboard computers, utilizing the embedded Linux as operating system. The findings demonstrated substantial performance improvements, including increased data processing rates and power savings, compared to standard AI solutions, thereby enabling real AI capabilities in space.To make AI applications viable for institutional missions, the project outlines future research efforts within the KATESU initiative. These include porting the novel inference engine software to space-qualified hardware systems such as the GR740 and GR765 on-board processors, developing an extension to accelerate AI algorithms via the HPDP co-processor, and creating another extension to expedite AI algorithm implementation on CPU + FPGA boards. Additionally, the ongoing support for two operating systems, namely RTEMS5 (ECSS-qualified real-time operating system) for GR740, GR765, and HPDP, and FreeRTOS for the FPGA extension, is discussed. This article provides the current development status pertaining to the target hardware platforms and outlines the planned validation process, along with the designed performance benchmarks utilizing OBMark-ML.