In-Orbit AI: Satellite Identifies Targets Without Ground Control

Typically, Earth observation requires downloading massive raw data sets for human analysts to review. The Yam-9 mission, powered by Google DeepMind’s Gemma 3 model and NASA JPL’s NAVI-Orbital software, bypasses this bottleneck. By running on a Nvidia Jetson Orin AGX GPU, the satellite successfully responded to natural language queries, such as identifying infrastructure near railway hubs or distinguishing between natural environments and human development.

This capability transforms satellites from passive sensors into active observers. Paul Lasserre, Loft’s head of AI, envisions a future of persistent patrol layers where satellites monitor specific borders or events, providing alerts only when suspicious activity occurs. By performing initial data triage in space, these systems drastically reduce the bandwidth requirements and latency inherent in current orbital infrastructure.

While Yam-9 acts as a pathfinder, the technical architecture is already being scrutinized for wider adoption. Engineers from NASA JPL had to heavily optimize the Gemma 3 model to fit within the stringent memory and power constraints of space-grade hardware. Industry competitors are closely tracking these developments, with firms like Planet Labs and Kepler Communications already operating orbital compute environments that could soon support similar large-scale AI applications.

Beyond commercial surveillance, the technology has roots in space exploration. JPL researcher Taran Cyriac John originally conceptualized the software as a digital assistant for astronauts on the Moon or Mars, where hands-free, voice-interactive AI could manage complex tasks. Loft plans to scale its constellation to between 50 and 100 satellites, aiming for real-time, global coverage that relies on the satellite's ability to "think" before it transmits.