Satellites can now see and think — is the era of human analysts over?
a loft orbital satellite used google deepmind's gemma 3 to autonomously find objects in space for the first time, proving vlm edge inference works in orbit.
a loft orbital satellite named yam-9 successfully executed the first on-orbit demonstration of a vision-language model, using google deepmind's gemma 3 to identify infrastructure and classify terrain based on natural language commands. developed with nasa jpl's 'navi-orbital' software, the system runs on an nvidia jetson orin agx gpu to process data locally rather than bottlenecking ground stations with raw imagery. this shift allows for intelligent data triage in space, enabling constellations to function as autonomous patrol layers that alert users only to significant anomalies. loft aims to scale this capability to a 100-satellite constellation for real-time global coverage, potentially revolutionizing how defense and commercial entities utilize earth observation data.