ASA ICEYE Innovation Lab project
AI Models Onboard Satellites
ASA is assembling a team of five students to answer a real open research question for ICEYE, the company behind the world's largest SAR satellite constellation.
ASA ICEYE Innovation Lab
The ASA ICEYE Innovation Lab is a collaboration between the Aalto Space Association and ICEYE, created to connect ambitious students with real research problems from the space industry. It provides a framework for focused co-projects, in which a small team of university students takes on a research question drawn from the frontier of ICEYE's work. ICEYE guides the team throughout, so the students are not working on a hypothetical exercise but on a question the industry actually needs answered.
The pilot project: AI models onboard satellites
Satellites produce imagery faster than ground links can carry it. Running AI models onboard cuts delivery latency, eases bandwidth pressure, and lets satellites act on what they see, for example by cueing another satellite to take a closer look.
Foundation models are strong candidates for this role. Pretrained on vast and diverse data, they learn general representations that can be adapted to many tasks from a single shared backbone. The obstacle is that they get their capability from scale, while a satellite has hard limits on compute, memory, and power.
Current research shows what foundation models can do in Earth observation. It rarely addresses what happens when one has to fit onto a satellite. Which onboard tasks are worth a shared foundation model backbone, and which are better served by small models built for a single job? Nobody has measured it.
The question has only recently become answerable. Open SAR foundation models with downloadable weights now exist, and GPU-equipped satellites are entering operation. This project sits at that intersection, and the team will find out where the line falls.
What to Expect
- 1.A deep dive into SAR remote sensing and what satellites are actually asked to detect, the open foundation models available, the datasets to work with, and the lightweight alternatives they would compete against.
- 2.Compressing a foundation model down to something a satellite could plausibly run, and training lightweight specialists as the counterweight, under a matched compute budget.
- 3.Deploying the models on an NVIDIA Jetson and measuring what actually matters onboard.
- 4.Finally, working out what the numbers mean and presenting the findings to ICEYE. ICEYE guides the team throughout, and the work opens a door to continuing with them afterwards.
What is Expected
The team is built so that the pieces complement each other, and no one is expected to cover everything. Between them the five need to reach across foundation models, model training and evaluation, model compression, edge deployment, and the SAR domain, but depth in one area matters more than coverage of all of them.
What matters most is a background solid enough to absorb new material quickly and read recent research with a critical eye. Much of what the project needs will be learned during it. Genuine interest in the subject and a commitment to the workload count for more than a long list of prior experience.
Relevant backgrounds include machine learning and computer science, embedded systems and electrical engineering, and remote sensing or space technology.
Practicalities
Timeline. The project runs from the beginning of October 2026 to the end of February 2027.
Workload. 5 ECTS, which is roughly 135 hours per person, or about 7 hours per week across the project.
Credits. Participation is credited as 5 ECTS through Aalto University's CS-E4003 Special Assignment in Computer Science.
Working. Much of the work can be done remotely. ASA is arranging a space where the team can work together once a week.
Team. Five students, including a project lead.
Applications. The application period ends on 13 September at 23.59.
Questions.
