An Earth observation satellite has used a vision-language model to identify requested features in imagery without relying on analysts on the ground, marking the first reported use of this type of AI model in orbit.
The April demonstration took place aboard YAM-9, a satellite operated by Loft Orbital. Software developed by NASA’s Jet Propulsion Laboratory allowed Google DeepMind’s Gemma 3 model to respond to natural-language requests and search satellite imagery for matching areas.
AI Processes Images On The Satellite
Vision-language models combine the language understanding of large AI models with the ability to interpret images. During the demonstration, researchers asked the system to classify areas where natural landscapes met human development and identify infrastructure near railway hubs.
The model completed those tasks aboard the satellite rather than sending all the raw imagery to Earth for analysis. This could allow satellites to filter data in orbit and transmit only the most relevant findings.
Google designed Gemma 3 as a lightweight multimodal model that can operate on limited hardware. YAM-9 carries an Nvidia Jetson Orin AGX processor, which provides the computing power required to run AI applications in space.
NASA Software Prepared The Model For Orbit
Juan Delfa Victoria, a technical leader in NASA JPL’s AI group, led the development of NAVI-Orbital. The software package acted as the operating framework for Gemma 3 aboard YAM-9.
Although Gemma 3 was already available, engineers had to reduce the number of software libraries and the amount of memory required. Power, processing capacity, and memory are limited aboard satellites compared with ground-based data centres.
Loft launched YAM-9 in November 2025 as a test platform for orbital AI projects. Its official mission announcement said customers could deploy and update algorithms using the satellite’s onboard computing resources.
Loft Plans A Larger Satellite Network
Loft’s head of AI, Paul Lasserre, told TechCrunch that vision-language models could allow satellites to monitor locations continuously and alert users when they detect specified activity.
The company currently operates 12 spacecraft. Lasserre estimated that between 50 and 100 satellites similar to YAM-9 would be needed to provide real-time coverage of any location on Earth.
Other satellite companies are also examining onboard AI. Planet Labs uses Jetson Orin processors for object detection and is researching vision-language models, while Kepler Communications operates a group of graphics processors in orbit.
The project could also support AI assistants for astronauts. NASA JPL researchers have considered systems that could respond to spoken requests from astronauts wearing pressurised suits while working on the Moon or Mars.
Featured image credits: Wikimedia Commons
For more stories like it, click the +Follow button at the top of this page to follow us.