An Earth-observation satellite captures images, but a useful observation still needs to reach someone who can interpret it. What if part of that interpretation could happen before the image leaves orbit?
NAVI-Orbital explores that possibility. Developed by researchers at NASA’s Jet Propulsion Laboratory and Loft Orbital, the system runs Google’s Gemma 3 vision-language model aboard Loft’s YAM-9 satellite. Loft has confirmed the collaboration and its deployment of a vision-language model in orbit.
The experiment offers a concrete example of AI operating close to the instrument that produces the data.
What did the researchers demonstrate?
The research paper reports an onboard demonstration on April 16, 2026. The system classified satellite imagery, generated text descriptions and answered follow-up questions. Its model processed images locally on the spacecraft.
The authors describe two live image captures alongside tests using other imagery. They frame the results as a demonstration of feasibility, with further work needed to establish robustness.
This is an explanation of an earlier experiment, rather than an announcement of a new launch today. Google highlighted the deployment again in its October 8 Gemini at Work presentation.
Why analyze an image before transmitting it?
Consider a hypothetical operator looking for a particular type of scene among many images.
Receiving a brief description first could help the operator decide which full images deserve closer inspection. The summary would act as an index: a way to select material, rather than a replacement for the evidence.
The NAVI-Orbital paper describes this approach as semantic compression. Its proposed workflow sends compact descriptions to the ground, allowing operators to request full imagery for selected scenes.
The distinction matters. A smaller image file preserves a visual representation. A written description selects information and expresses an interpretation.
For example, “a coastal settlement beside a road” may help someone find a relevant scene. It cannot support every conclusion that a specialist might draw from the original pixels.
A summary can save attention—and omit something important
Our interpretation is that the strongest use case is prioritization.
A summary could help an operator identify an image worth reviewing. But if it overlooks the feature the operator needs, a useful observation might receive too little attention.
A sensible evaluation would therefore measure both directions of error:
- False positives: scenes flagged as relevant that do not contain the target.
- False negatives: relevant scenes that the system fails to flag.
Which error matters most depends on the task. For an exploratory search, reviewing a few extra images may be acceptable. For a time-sensitive monitoring task, missing an important scene could carry a much higher cost.
That trade-off should determine how the system is used.
Natural-language questions need precise expectations
The paper describes operators defining observation tasks through text prompts and asking questions about the results.
An accessible interface could make it easier to express a research question. It also creates a need to define what counts as a supported answer.
Imagine asking:
Identify images that may contain coastal settlements. Explain which visible features support that description, and flag ambiguous scenes for review.
This is an illustrative request, not a quoted experiment result.
“May contain” leaves room for uncertainty. A stronger statement would require stronger evidence. The same distinction applies to ordinary AI document analysis: the wording of a conclusion should reflect what the source actually establishes.
Our guide to checking AI answers with sources explains this relationship between evidence and interpretation.
What remains to be proven?
The authors explicitly distinguish their demonstration from a broader characterization of reliability. Adversarial dialogue testing was outside its scope, and the live-image sample was small.
For a future operational system, useful questions would include:
- Does performance hold across different landscapes and image conditions?
- Can an operator trace a description back to the original acquisition?
- What happens when the answer is ambiguous or unsupported?
- Can important imagery still be retrieved when the summary is wrong?
These are evaluation questions, not claims that the experiment has already answered them.
What this teaches us about useful AI
Our reading is that NAVI-Orbital illustrates a valuable design choice: put analysis where it can help people select and understand information.
The benefit would come from improving the path between observation and review. A fluent description is only one part of that path; the system also needs traceability, uncertainty handling and access to the original material.
For satellite operators, the experiment provides evidence that onboard vision-language processing is feasible. The next question is how reliably it can support a specific mission.
For everyone else, it is a reminder that useful AI extends beyond chat interfaces. Sometimes the most interesting application is helping people find the evidence they should examine next.
Sources
- Delfa Victoria, John and Herson — NAVI-Orbital: First In-Orbit Demonstration of a Zero-Shot Vision-Language Model for Autonomous Earth Observation, research paper.
- Loft Orbital — From One Satellite to Constellations: Loft’s First 5 Years in Space.
- Google Cloud — Gemini at Work 2026, October 8, 2026.