In a groundbreaking development, an Earth observation satellite has autonomously identified specific objects without human intervention. This achievement, realized in April, marks the inaugural use of a vision-language model (VLM) in space, potentially revolutionizing the capabilities and value of space-based sensors.
Traditionally, satellites transmit vast amounts of data to Earth, where analysts employ machine learning algorithms or manual inspection to interpret the information. However, aboard the Yam-9 spacecraft, developed by Loft Orbital, a software package from NASA’s Jet Propulsion Laboratory (JPL) enabled the satellite to process and identify areas of interest in response to natural language queries directly in orbit.
The VLM utilized, Google DeepMind’s Gemma 3, is specifically designed for edge applications, allowing it to operate on hardware with limited resources far from centralized data centers. By integrating the contextual understanding of large language models with image analysis capabilities, the model successfully classified sensor data to distinguish between natural environments and human developments, and identified infrastructure around railway hubs.
This demonstration holds significant implications. In the short term, it can enhance the utility of space sensors by performing preliminary data analysis onboard, thereby reducing the volume of raw data that needs to be processed by analysts on Earth. In the long term, it serves as a proof of concept for deploying more extensive AI infrastructure in space.
Paul Lasserre, Loft’s head of AI, highlighted the potential of this technology, stating that it paves the way for continuous monitoring systems in space. With VLMs, satellites could be programmed to monitor specific areas, such as borders, and alert operators to any suspicious activities, facilitating interactive communication between ground teams and satellites.
Loft Orbital’s business model focuses on providing satellite platforms for third-party clients, akin to an infrastructure-as-a-service approach rather than traditional satellite manufacturing. For instance, they recently built, launched, and now operate six new satellites for EarthDaily, which analyzes and markets the data collected. Launched in the fall of 2025, Yam-9 serves as a testbed for the company’s orbital AI initiatives and is equipped with an Nvidia Jetson Orin AGX GPU, a leading chip for space-based computing.
Juan Delfa Victoria, a technical leader in NASA JPL’s AI group, spearheaded the development of NAVI-Orbital, the software framework that integrated the Gemma 3 VLM. While Gemma 3 is an off-the-shelf model, engineers streamlined the software to minimize the required libraries and memory, ensuring efficient operation in the satellite’s constrained environment.
Although this marks the first reported deployment of a VLM in orbit, other companies are likely to follow suit. Planet Labs, for example, operates satellites equipped with Jetson Orin processors. Currently, these are used for simpler object detection tasks, but research is underway to explore additional AI applications, including VLMs.
Kepler Communications, which manages a substantial array of GPUs in space, has not disclosed specific applications of VLMs due to non-disclosure agreements with partners. However, the company acknowledged multiple undisclosed uses of their compute environment since the deployment of these space-based GPUs.
This advancement signifies a pivotal shift in satellite technology, moving from passive data collection to active, intelligent analysis. As AI continues to evolve, the integration of such models into space operations could lead to more responsive and efficient monitoring systems, enhancing our ability to observe and understand Earth’s dynamic environment.
