Britain’s space-AI roadmap finds an opportunity in limited onboard computing

The UK Space Agency’s AI roadmap, published on 27 October, identifies a national opportunity in systems that operate efficiently and can be trusted inside demanding space missions. Its 2026–2036 strategy emphasises low-power computing, compressed models, radiation-tolerant hardware and verification. Those are priorities for investment and development, rather than a newly funded launch schedule.[1]

The commercial question becomes concrete when a satellite has more information than it can promptly send home. Should it transmit every image, choose the useful ones, or decide to take a different picture in the first place? Each choice puts intelligence in a different part of the mission and creates a different product for an AI supplier.

Choosing the data before sending it

ESA’s Φsat-2 offers one established illustration. Launched in August 2024, the small demonstration satellite carries applications for tasks including cloud screening, vessel detection and image compression. Its cloud application is designed to discard obscured views before downlinking; other applications turn imagery into information such as street maps.[2]

This is a useful way to understand efficiency. The goal is not simply to run the largest possible model on a small computer. It is to spend limited processing on a decision that reduces a larger burden elsewhere. A compact answer about an image can sometimes be more useful to the next operator than the image arriving late.

Changing the observation itself

NASA reported a different approach in July 2025. In a flight test on CogniSAT-6, its Dynamic Targeting technology analysed a forward-looking image and planned cloud-free observations in roughly 60 to 90 seconds. The spacecraft tilted ahead, assessed the scene, then returned its camera toward the ground for the selected imaging. That first test targeted cloud avoidance; searching for fires and storms was future work in the report.[3]

Here the software influences what gets measured, rather than only deciding what to keep. A ground system receiving the same information later cannot recover an observation opportunity already passed. This gives response time a scientific value independent of the price of processing the image.

A longer-running example is JPL’s AEGIS system on Mars rovers. It identifies geological targets and ranks them against priorities supplied by scientists, then directs an instrument without waiting for Earth. JPL records routine use on Curiosity since 2016 and deployment on Perseverance in 2022. Scientists set the criteria; the rover applies them when a new scene becomes available.[4]

The product includes the evidence that it will work

The UK roadmap links efficient onboard hardware and software with shared test environments, representative data and space-specific assurance. It also calls for validation and performance standards in funding programmes. Its assessment describes space-grade neuromorphic deployment as nascent, keeping that longer-term prospect distinct from more mature applications.[1]

These examples suggest a demanding but intelligible route to market: deliver a bounded mission task, show that it fits the available resources, and make its behaviour examinable. The software supplier’s output may be a cloud mask, an observation plan or a ranked target list. In every case, the mission team needs to understand what the system will do with the next unfamiliar scene.

The opportunity is therefore larger than supplying a smaller model. It includes the interfaces, test evidence and operating rules that turn an algorithm into something a mission can rely on. Britain’s proposed niche makes sense precisely where a distant data centre cannot make the immediate decision for the spacecraft.

Report an error or suggest a correction