A VLM in orbit starts shrinking the satellite data bottleneck

The bottleneck in satellite imagery is not only taking pictures. It is deciding which data is worth sending back to Earth. A new on-orbit vision-language model demonstration points to a future where that first triage step happens in space.

TechCrunch reports that YAM-9, a spacecraft built by Loft Orbital, used software from NASA’s Jet Propulsion Laboratory to identify areas of interest in response to natural-language queries. The April demonstration is described as the first reported case of an Earth-observation satellite finding what it was looking for without ground analysts doing the initial search.

The model behind the demo was Google DeepMind’s Gemma 3, running through a JPL software package called NAVI-Orbital. Researchers asked it to classify sensor data around natural environments near human development and to identify infrastructure around railway hubs. Instead of sending large volumes of raw imagery down first, the spacecraft performed the first pass onboard.

The hardware detail matters because orbit is a constrained edge environment. YAM-9 launched in fall 2025 and carries an Nvidia Jetson Orin AGX GPU. JPL engineers had to streamline the software package to reduce libraries and memory use so the VLM could run under those limits.

The near-term value is better data triage. The longer-term implication is more autonomous space sensing, where satellites can respond to plain-language monitoring tasks and alert operators when something changes. Loft currently operates 12 spacecraft and estimates that real-time global coverage would require roughly 50 to 100 YAM-9-like satellites.