A Loft Orbital Yam-9 satellite identified areas of interest without the usual direct assistance from ground control by using an onboard artificial intelligence model to interpret images and sensor data. TechCrunch reports that this could be the first instance of a vision-language model (VLM) being used in Earth orbit.
Systems and method used
The satellite ran Google DeepMind’s Gemma 3 model, designed to operate effectively away from data center environments on more limited hardware. The Yam-9 worked with software from the NASA Jet Propulsion Laboratory to pinpoint relevant locations, and it received only natural-language queries from the ground to guide the tasking.
Researchers asked the model to perform tasks such as classifying sensor data where natural environments meet human activity and identifying infrastructure around rail junctions. According to the report, Gemma 3 satisfied all requests made of it.
Why this matters
Today, most satellite observation workflows transmit large volumes of raw data to Earth, where humans or machine-learning systems analyze it. Using onboard VLMs can reduce that data flow by producing preprocessed, contextual results at the sensor, improving bandwidth efficiency and speeding up actionable insight.
Moreover, the successful demonstration represents a step toward building more substantial AI infrastructure in space, enabling more complex local decision-making and automation on orbiting platforms.
Conclusion
The Yam-9 experiment shows that VLMs can materially increase the utility of spaceborne sensors while lowering the need for extensive ground-side analysis. If further tests corroborate these findings, this approach could fundamentally change how data is collected and processed in Earth orbit.
