AI System Enables Satellite To Locate Targets

A remote sensing satellite, Yam-9, autonomously identified points of interest in April 2026 — marking the first instance of AI conducting triage while in orbit. Using Google’s open source Gemma 3 vision-language model, Yam-9 classified images captured by its sensors based on natural language queries (e.g., identify borders between nature and civilization, nearby infrastructure at train stations). Developed by Loft Orbital, the onboard software was authored by NASA’s Jet Propulsion Laboratory, and the vision model provided by Google DeepMind.

Gemma 3, an open source vision-language model developed by DeepMind, was deployed using JPL’s NAVI-Orbital software package on an Nvidia Jetson Orin AGX, a compact GPU similar to those found in robots but not in data centers. The goal is to enable edge AI, which executes on the satellite rather than in a data center hundreds of kilometers below. “It allows us to start thinking about creating these patrolling layers in space,” explained Paul Lasserre, CEO of Loft. “So if I say to you, ‘Hey, keep an eye on this border and tell me if there’s something funny going on,’ it can do that.”

1Why doing it onboard changes everything

Today, most remote sensing is performed slowly, with satellites capturing images of large portions of Earth, transmitting the data back to Earth, and then human operators analyzing the images hours or days after they were collected. But the bandwidth available to satellites is constrained and costly, resulting in a lot of wasted downlink bandwidth as satellites transmit irrelevant data (e.g., water, clouds, etc.). With onboard AI, the satellite can make decisions about which data to send back based on what it identifies in real-time, effectively turning the downlink into a question-answering pipeline instead of a pixel-transmission pipeline. “Instead of a passive camera, you get an interactive assistant in space,” noted Juan Delfa Victoria, the technical lead at JPL. Loft eventually hopes to deploy a constellation of 50–100 similar satellites to provide near real-time surveillance of Earth; other companies such as Planet Labs and Kepler Communications are pursuing similar objectives.

2Why it matters to you

What’s more interesting than the satellite is the fact that the model powering it is publicly available, and can be deployed on your laptop. Yes, we hear a lot about large models running in big data centers, but the real revolution is happening with small, free models running on low-cost hardware in disconnected environments (e.g., phones in the middle of nowhere, sensors on farms, chips on satellites). AI is quickly becoming ubiquitous and accessible.