Satellite Uses Onboard AI to Find Targets Itself

In April 2026, an Earth-observation satellite picked out areas of interest on its own, without a human analyst telling it what to look for — the first time AI has done that kind of triage while running in orbit. The spacecraft, Yam-9, classified its own sensor data from plain-language queries, flagging things like the boundaries where nature meets human development and infrastructure near railway hubs. It was built by Loft Orbital, with the onboard software written by NASA's Jet Propulsion Laboratory and the vision model supplied by Google DeepMind.

The model doing the work is Gemma 3, DeepMind's open vision-language model, wrapped in JPL's NAVI-Orbital software package and running on an Nvidia Jetson Orin AGX — the kind of compact GPU you'd find in a robot, not a data center. That's the whole point: the AI runs at the edge, on the satellite itself, hundreds of kilometers up with no cloud to phone home to. 'It opens the door to always-on, patrol layers in space,' said Paul Lasserre, Loft's head of AI. 'You can have logic — like monitor this border for me, and let me know when something is suspicious.'

Why doing it onboard changes everything

Most Earth observation today works the slow way: a satellite photographs huge swaths of the planet, dumps the raw data to the ground, and human analysts comb through it hours or days later. Bandwidth from orbit is limited and expensive, so a lot of what gets downloaded is empty ocean, cloud cover, or nothing of interest. Running a model onboard flips that around — the satellite decides what's worth sending before it sends anything, so the link carries answers instead of raw pixels. Juan Delfa Victoria, the JPL technical lead, described the result as an interactive assistant in orbit rather than a dumb camera. Loft says it eventually wants a fleet of 50 to 100 such satellites for near real-time coverage of the planet, and rivals like Planet Labs and Kepler Communications are chasing the same idea.

Why it matters to you

Two things here are bigger than the satellite. First, faster orbital triage feeds the systems ordinary people quietly rely on — disaster response, wildfire and flood mapping, shipping and weather — because the useful frames reach analysts in minutes instead of after a manual scan of everything. Second, and the part that stuck with me: the model flying this mission is the same open Gemma family you can download and run on a laptop. The frontier story is all about giant models in giant data centers, but the quieter trend is small, free models doing real work on cheap hardware in places with no connectivity at all — a phone in a dead zone, a sensor on a farm, a chip on a satellite. That capability is becoming normal infrastructure, not a lab demo.