QlevEr Sat: demonstrating an edge AI for EO mission on a 2U payload ground model
- Paper ID
80957
- DOI
- author
- company
Universite Grenoble Alpes; Universite Grenoble Alpes, CNRS
- country
France
- year
2024
- abstract
Testing and prototyping phases are crucial to any engineering discipline, and it is especially true in Space engineering. It is even more true with the arrival of Artificial Intelligence in Space. In this paper, we present how the Grenoble University Space Center (CSUG) has demonstrated that the QlevEr Sat 2U payload ground model was able to acquire 16Mpix images, with a high-performance colour CMOS sensor (such as Emerald by Teledyne e2v) and directly process them by a lightweight AI algorithm specifically designed for edge computing, on a quad-core “radtol” processing module (such as Qormino by Teledyne e2v). Moreover, this innovative optimised AI model made in Grenoble can be indefinitely retrained to adapt to various use cases and hardware constraints. In order to achieve such a demonstration on the ground, a sensor board and a microprocessor motherboard at CubeSat scale were designed, then successfully assembled, integrated and tested with a 300mm space-ready lens in a low-volume payload (2U). On the software side, the various software modes of the payload, including Acquisition mode, AI mode, and transfer modes, were specified, developed, integrated and tested, while the AI solution was being assessed by the European Space Agency. This payload ground model was then connected to an external graphical user interface prototype emulating both the 6U platform and the ground segment in order, on one hand, to send telecommands to the payload, and, on the other hand, to transfer and display mission data and telemetries from the payload, like a mission centre would do. Mission data included both live images and a realistic space dataset, together with their associated AI-generated maps. The advantage of such AI-interpreted maps is that they are hundreds of times lighter than raw images, while containing only useful information. The demonstration was itself placed into the whole mission context thanks to a mockup-storyboard, produced during a collective intelligence workshop involving rapid prototyping. Overall, the CSUG’s QlevEr Sat team was able to run a functional edge-AI payload model capable of onboard processing to deliver the specified EO mission.