In April 2026, the YAM-9 satellite operated by Loft Orbital became the first spacecraft to run a vision-language model (VLM) in orbit. The NAVI-Orbital software harness developed by NASA Jet Propulsion Laboratory, running Google DeepMind's Gemma 3, autonomously answered natural language queries about Earth sensor data. The milestone was reported by TechCrunch on June 15, 2026.
Key takeaways
- First documented deployment of a vision-language model (VLM) in Earth orbit.
- Mission conducted by Loft Orbital's YAM-9 satellite; NAVI-Orbital software developed by NASA JPL.
- Model used: Gemma 3 by Google DeepMind, designed for edge deployments on constrained hardware.
- Compute hardware: NVIDIA Jetson Orin AGX installed onboard YAM-9.
- The satellite answered natural language queries such as identify natural environments bordering human-developed areas.
VLM in Space: What Happened
The standard Earth observation pipeline works like this: a satellite collects large volumes of data, transmits it to ground stations, and analysts or ML algorithms determine what is relevant. NAVI-Orbital proposes inverting this sequence. Instead of transmitting raw data, YAM-9 processes it autonomously onboard and sends only the results to Earth: answers to posed queries or flagged areas of interest.
Vision-language models (VLMs) combine text understanding with image analysis. Unlike classical object detection models trained to recognize predefined classes, VLMs allow descriptive natural language queries. This is a significantly more flexible approach to satellite imagery analysis.
Gemma 3 as an Edge Model
Not every model is suitable for space operation. YAM-9 has constrained power and memory. Google DeepMind designed Gemma 3 as an edge model: compact and capable of running on limited compute. Juan Delfa Victoria from NASA JPL additionally optimized the NAVI-Orbital software layer, reducing library count and memory footprint to fit within the NVIDIA Jetson Orin AGX compute budget.
The Jetson Orin AGX chip is one of the most widely used processors in space applications. Planet Labs uses Jetson Orin in its satellites for simpler object detection tasks and is researching VLM applications. Kepler Communications, operator of the largest GPU cluster in orbit, confirmed several undisclosed use cases.
What the Test Means for Earth Observation
Paul Lasserre, Loft Orbital's head of AI, described the direction: if you have a VLM, you can assign a task such as monitor this border and alert me when something looks suspicious.
According to Loft Orbital, full real-time global coverage would require a constellation of 50-100 satellites similar to YAM-9. The company currently operates 12 spacecraft. This is a clear gap between a technology demonstration and commercial observation infrastructure.
It is important to separate what was proven from what is vision. The April 2026 test demonstrated that a VLM can operate aboard a satellite and correctly process natural language queries on sensor data. It did not demonstrate commercial reliability, constellation-scale deployment, or resistance to the space environment over extended operation.
Why This Matters
The YAM-9 test confirms that the technical barrier for AI in orbit is no longer hardware: Jetson Orin AGX and Gemma 3-class models are sufficient for what a year ago would have required dedicated accelerators. The barrier now is architectural and commercial: how to integrate onboard inference into customer operational pipelines? How to price orbital compute versus satellite bandwidth?
For the Earth observation sector the change is potentially significant. An orbital VLM can serve as a filter: instead of downloading everything and searching later, you query the satellite upfront and receive an answer. The military and regulatory implications are clear: satellites that autonomously identify objects from natural language queries enter geopolitically sensitive territory.
What Next?
- NASA JPL and Loft Orbital plan to extend the NAVI-Orbital test to a broader range of query types; commercial timeline not disclosed.
- Planet Labs has declared research on VLMs across its Jetson Orin fleet; results expected in H2 2026.
- Loft Orbital targets a constellation of 50-100 AI-capable satellites by 2028; current fleet is 12 spacecraft.
Sources
TechCrunch: A satellite just learned to find things on its own
Loft Orbital: loftorbital.com
