Nebula Public Library

In this activity, Mission Control Space Services deployed a low-level implementation of the OPS-SAT SmartCam model using a Field Programmable Gate Array (FPGA), comparing against a high-level CPU model using Tensorflow Lite. Experiments showed that the FPGA implementation reproduced the precision and accuracy of the high-level model, while running at a slower speed. Further optimisations of the FPGA are expected to close the gap in timing and unlock new methods for deploying deep learning on spacecraft.

ESA's OPS-SAT mission is a key element for operational validation in flight (or IOD/IOV) of the DeepCube service of Deep Learning at the edge that Agenium Space is preparing with the support of the GSTP programme (GSTP Make, ESA-CNES) and the R&D performed in the CORTEX project (Permanent Open Call EOEP-4, ESA/Phi-Lab, https://esacortexproject.agenium-space.com). The goal of this activity was to execute on board the inference of the simplified models defined in those projects.

Early space missions consisted of bespoke components communicating with facilities on Earth. Today this model has been updated with off-the-shelf components and reusable launch systems that lower mission cost into low-Earth orbit. Extending this model to develop a sustainable off-Earth economy requires bringing the concept of reconfigurable computing agents to deep space.

Model based approaches have proved to be efficient in supporting engineering activities, models replacing traditional document-based approaches. Nevertheless even in most advanced deployments, a lot of engineering artifacts are textual either because the return on investment of introducing models is too low on this particular case or because the information is more efficiently expressed in natural language even if consistency and correctness issues appears.