For the first time, an Earth observation satellite has independently identified what it was asked to find, without relying on analysts on the ground. The test, completed in April, marks the first reported orbital use of a vision-language model and points to a new era for space-based intelligence.
Instead of sending large data sets back to Earth for manual review, the Yam-9 spacecraft used software developed by NASA's Jet Propulsion Laboratory to process natural-language queries directly in orbit. Built by Loft Orbital, the satellite responded to requests by spotting areas of interest in sensor data on its own.
The system ran on Gemma 3, Google DeepMind's edge-focused model designed for hardware with limited computing power. By combining language understanding with image analysis, the model was able to classify scenes where nature meets infrastructure and detect features around railway hubs.
The achievement matters on two levels. In the short term, it can help satellites filter data before it reaches Earth, making observations faster and more efficient. In the longer term, it offers a practical step toward larger AI systems operating in space.
Loft Orbital says this approach could support always-on monitoring layers in orbit, where satellites respond to specific instructions and flag relevant changes in near real time. The company's model is built around flexible spacecraft platforms for customers, and Yam-9 serves as a testbed for orbital AI.
NASA JPL's NAVI-Orbital software helped adapt the model for space use by reducing memory and software demands. The project also used an Nvidia Jetson Orin AGX processor, showing how compact computing hardware is becoming central to next-generation missions.
Researchers expect similar systems to spread as satellite operators explore more advanced onboard AI. Beyond Earth observation, the same tools could support future astronaut assistants for lunar or Martian missions. The result may be a future where satellites do not just collect data, but actively interpret it in orbit.