One of my favourite AI deployments isn't on Earth. It's on Mars.
The rovers — Curiosity, Perseverance — have onboard AI systems that help them decide, moment to moment, where to drive, what rocks to point their instruments at, and what's worth telling Earth about. There's a reason for this: the radio round-trip from Mars to Earth is up to 22 minutes. You can't joystick a rover. So you give it judgment, and you trust it.
Perseverance has a feature called AEGIS that uses computer vision to identify interesting geological targets while the rover is driving. It points its laser spectrometer at them without waiting to ask humans. When scientists on Earth wake up, they find data about things they never told it to look at — things it found interesting on its own.
That's a lovely image to me. A small, patient, curious machine on another planet, deciding what to notice.
Where else AI is earning its place in space
Earth observation. Satellites generate more imagery than any human team can process. AI systems now handle the first pass: identifying deforestation, monitoring crop health, tracking illegal fishing, spotting early signs of forest fires. The scale is extraordinary — some tasks that used to take months of analyst time are now completed overnight.
Autonomous docking and rendezvous. SpaceX's Dragon and other modern spacecraft use computer vision and machine learning for autonomous docking with the ISS. This is harder than it sounds — they have to track each other, adjust for tiny gravitational perturbations, and bring two vehicles travelling at 28,000 km/h into contact at a few centimetres per second, all while both are moving. AI handles the final approach.
Astronaut assistance. The ISS has tested systems like CIMON (an astronaut-assistant robot) and more recently voice-operated AI assistants that help with experiments, procedures, and communication. Useful for crew cognitive load in long-duration missions.
Exoplanet discovery. We've discovered thousands of planets around other stars. Many were first spotted by AI systems trained to notice the tiny, periodic dips in a star's brightness that indicate a planet transiting in front of it. The Kepler and TESS missions wouldn't have produced anything like the catalog they have without machine learning.
Scheduling and operations. Space missions have extraordinarily complex constraints — power budgets, communication windows, thermal limits, instrument priorities. AI scheduling assistants help mission operations teams figure out how to squeeze the most science out of limited resources.
Spacecraft fault diagnosis. When something goes wrong on a spacecraft, the engineers on the ground have to diagnose the problem from a trickle of telemetry. AI tools that correlate anomalies across subsystems are speeding up this diagnosis — sometimes dramatically.
Some unexpected applications
Monitoring the ocean from space. AI models looking at satellite imagery can now detect ships, even small ones, even at night using thermal imaging. This is being used to track illegal fishing, piracy, and the otherwise-invisible 80% of global shipping that doesn't transmit AIS.
Mapping the world. Not just Google Maps. Humanitarian organisations use AI on satellite imagery to rapidly map disaster-affected areas after earthquakes and hurricanes — finding which buildings are damaged, which roads are blocked, where people might be trapped. This has saved lives.
Pollution monitoring. Satellites watching methane plumes, sulphur dioxide from industrial sites, and CO₂ from power stations — with AI doing the heavy lifting of analysis — are making climate accountability possible at a scale that wasn't practical before.
The hard parts
Radiation-hardened compute is limited. The AI hardware that works on Earth won't last long in the radiation environment of deep space. Onboard AI has to run on much more modest processors than a modern data centre. This forces elegant, efficient models — which is often a good thing.
You can't update in flight (easily). If a spacecraft's AI has a bug, you're not downloading a patch over the air. Missions pick conservative, well-validated models and live with them for years.
Sim-to-real is especially brutal. You can't really test a Mars rover on Mars before it gets there. Models trained on simulations and Earth analogues have to generalise to environments they've never seen. This is the hardest version of a problem we discussed in the robotics post.
Why I love this field
There's something about applying AI to the universe at large that I find deeply moving. The tools we've built to understand our own planet and solar system are helping us understand others. Humans have always looked at the stars and tried to figure out what was there. The machines we've sent are, in a small way, part of that looking.
Perseverance's AEGIS system, deciding on its own to shine a laser on an interesting rock, is very much in that tradition. Not a replacement for the scientists on Earth. A patient assistant in a place we can't easily go, doing its best to notice things we might want to know about.
If you're working on anything in this space (so to speak) — or just curious about it — we'd love to chat. Space-adjacent conversations are some of our favourites.
