A satellite has independently identified a target in orbit without waiting for human analysts on the ground, marking a new step for artificial intelligence in space. The demonstration, which took place in April, is the first reported use of a vision-language model on a spacecraft in orbit, according to the companies and researchers involved.
The experiment points to a potential shift in how Earth observation satellites work. Instead of sending large volumes of raw imagery back to Earth for later review, future spacecraft may be able to inspect data onboard, decide what matters and only relay the most relevant results. That could reduce delays and help customers make faster use of satellite imagery.
The test ran on YAM-9, a spacecraft built by Loft Orbital. The satellite used software developed by NASA's Jet Propulsion Laboratory to respond to natural language queries and find areas of interest in sensor data. The AI model behind the demo was Google DeepMind's Gemma 3, a vision-language model designed for edge computing, meaning it can run on limited hardware far from a data center.
Researchers used prompts that asked the system to identify specific kinds of scenes, such as natural environments near human development or infrastructure around railway hubs. The model was able to classify the imagery and return results without a person first reviewing the data on the ground.
Loft said the goal is not just to save time, but to open up more interactive use cases for satellites. Paul Lasserre, Loft's head of AI, said the technology could support continuous monitoring and two-way interaction with spacecraft, such as asking a satellite to watch a border area for unusual activity.
Earth observation companies currently rely on a workflow that sends large datasets to Earth for processing, either by machine learning systems or by human analysts. Moving some of that work into orbit could help triage data before it ever reaches the ground. That would be especially useful as satellite constellations grow and produce more imagery than teams can easily handle.
Loft views the demo as an early proof point for a broader strategy around orbital AI and compute. The company’s spacecraft are designed as infrastructure for third-party customers rather than as one-off satellites. YAM-9, launched in the fall of 2025, was built as a pathfinder for those efforts and carries a Nvidia Jetson Orin AGX GPU.
The software package, called NAVI-Orbital, was led by NASA JPL technical leader Juan Delfa Victoria. Although Gemma 3 was an off-the-shelf model, the team had to streamline the software stack so it could run within the satellite’s limited memory and library constraints.
Other companies are also exploring similar hardware. Planet Labs has satellites with Jetson Orin processors and says it is studying additional AI applications, including vision-language models. Kepler Communications, which operates a large group of GPUs in space, said it has had several undisclosed uses of its compute environment, though it did not confirm VLM deployment.
Loft said the demonstration is a sign of where the field is headed. The company wants to expand its constellation to provide more continuous coverage of Earth, a network it says would require roughly 50 to 100 satellites like YAM-9. It currently operates 12 spacecraft.
Beyond commercial monitoring, the work could also inform future tools for exploration. The idea behind NAVI-Space began as a concept for astronaut assistants on the moon or Mars, where typing on a keyboard may not be practical inside a pressurized suit.
For now, the result is a limited but important milestone: a satellite that could understand a request, inspect what it sees and identify a target on its own, without waiting for the ground station to catch up.