AI-Powered Satellite Learns to Find Targets on Its Own in Orbit

An Earth observation satellite has identified areas of interest on its own while in orbit, marking a major step toward satellites that can analyze what they see without waiting for human teams on the ground.

The milestone happened in April aboard YAM-9, a spacecraft built by Loft Orbital. The satellite used a vision-language model running in space to interpret sensor data and respond to natural-language prompts. Instead of simply collecting imagery and sending large amounts of raw data back to Earth, the system could identify what it was being asked to find while still on orbit.

The demonstration is important because it changes the role of a satellite. Traditionally, Earth observation satellites act mainly as data collectors. They capture images, transmit them to ground stations, and then analysts or machine learning systems on Earth process the material. With onboard AI, part of that analysis can happen immediately in space.

That could make satellites faster, more useful, and more valuable, especially for missions where time matters. Disaster response, border monitoring, infrastructure tracking, environmental observation, maritime activity, agriculture, and defense could all benefit from satellites that can sort through data before sending it down.

The Satellite Used a Vision-Language Model

The onboard system used Google DeepMind’s Gemma 3, a vision-language model designed to run on limited hardware. Vision-language models can understand both images and text, allowing them to connect visual data with natural-language requests.

That is different from older satellite analysis tools that may be trained to detect one specific object or pattern. A vision-language model can respond more flexibly. Researchers can ask it to identify a natural environment near human development, detect infrastructure around railway hubs, or classify areas that match a broader description.

This makes satellite AI more interactive. Instead of building a separate detection system for every task, operators can ask a model what to look for using language. That could make orbital sensing more adaptable and easier to use.

The demonstration also shows that advanced AI does not always need a huge data center. Gemma 3 is built for edge applications, which means it can run in places where power, memory, and connectivity are limited. A satellite is one of the clearest examples of that kind of environment.

Loft Orbital’s YAM-9 Became the Test Platform

The test took place on Loft Orbital’s YAM-9 satellite, which launched in the fall of 2025 as a pathfinder for the company’s orbital AI work.

Loft Orbital operates satellites as platforms for customers that want to deploy sensors, software, or missions without building an entire spacecraft themselves. Its model is closer to space infrastructure as a service than traditional satellite manufacturing. Customers can use Loft’s spacecraft, operations, and onboard systems to run missions in orbit.

That makes the company a useful testbed for onboard AI. If satellites become software platforms, customers may eventually upload and run AI applications in space the way developers run applications in cloud infrastructure.

YAM-9 includes an Nvidia Jetson Orin AGX GPU, a chip commonly used for edge AI applications. That onboard compute made it possible to run the vision-language model in orbit rather than relying entirely on ground-based processing.

NASA JPL Built the Software Harness

The model did not operate alone. NASA’s Jet Propulsion Laboratory developed a software package called NAVI-Orbital to help run the system aboard the satellite.

The software acted as the harness for the vision-language model. Engineers had to streamline the package so it could operate within the satellite’s limits, including memory, libraries, power, and compute constraints.

That work is important because running AI in orbit is not the same as running AI in a data center. Spacecraft have limited resources. Hardware must survive radiation, temperature swings, communication gaps, and strict power budgets. Every library, model size, and memory requirement matters.

The success of the test suggests that AI engineers are learning how to adapt modern models for space environments. That could open the door to more capable onboard systems in future missions.

Onboard Triage Could Reduce Data Overload

One of the biggest near-term benefits is data triage.

Earth observation satellites collect huge amounts of imagery and sensor data. Sending all of it to Earth can be expensive, slow, and inefficient. Analysts may then have to review large volumes of material to find the small portion that matters.

Onboard AI could change that. A satellite could scan its own data, identify the most relevant frames or regions, and prioritize what needs to be sent back. That would reduce the flood of raw data and make ground teams more efficient.

For time-sensitive missions, this could be crucial. If a satellite spots signs of a wildfire, flood damage, illegal fishing activity, damaged infrastructure, troop movement, or environmental change, faster detection could lead to faster response.

The satellite would not need to replace human analysts. It could simply make sure the most important data reaches them first.

Satellites Could Become Always-On Watchers

Loft Orbital’s head of AI, Paul Lasserre, described the technology as opening the door to always-on patrol layers in space.

That idea is powerful. A satellite with onboard AI could be assigned a mission in plain language, such as watching for suspicious movement near a border, tracking changes around infrastructure, or monitoring environmental conditions in a specific region. The system could then alert operators when it finds something that matches the request.

This would make satellites more active. Instead of waiting for humans to search through imagery, the spacecraft could participate in the search process itself.

That has obvious commercial and public-interest uses. Governments could use it for emergency response, border awareness, and infrastructure monitoring. Companies could use it for supply chains, energy assets, mining, insurance, agriculture, and logistics. Researchers could use it to observe climate, forests, coastlines, and disaster zones.

But it also raises serious questions. Always-on monitoring from space can become a surveillance tool. As satellites get smarter, policymakers may need to think more carefully about privacy, military use, and how autonomous detection is governed.

Other Space Companies Are Likely to Follow

Loft Orbital is unlikely to remain alone in this area.

Planet Labs already operates Earth observation satellites with onboard Jetson Orin processors and is reportedly researching more advanced AI applications, including vision-language models. Kepler Communications operates a large group of GPUs in space and has indicated that undisclosed compute use cases are already happening in orbit.

The trend is clear. Satellites are moving from simple sensors toward compute platforms. As chips become more capable and models become more efficient, more analysis can happen onboard.

This mirrors a broader shift in technology. Phones, cars, drones, robots, cameras, and factories are all becoming edge AI devices. Satellites are now joining that movement.

The difference is that space adds higher stakes. Bandwidth is limited, hardware is hard to repair, and missions are expensive. That makes onboard intelligence especially valuable if it can reduce communication load and increase mission usefulness.

The Commercial Value of Satellites Could Rise

If satellites can find valuable information on their own, their business value could increase.

Today, many Earth observation companies sell imagery, data feeds, analytics, or monitoring services. The more processing they can do onboard, the faster and more targeted those services can become.

Instead of selling large datasets for customers to analyze later, a satellite operator could sell alerts, detections, summaries, and answers. That moves the business closer to intelligence as a service.

For example, an energy company may not want thousands of images of pipelines. It wants to know whether a pipeline corridor has changed. A shipping company may not want raw ocean imagery. It wants to know where certain vessels are. A government agency may not want every frame from a disaster zone. It wants the damaged roads, flooded areas, and blocked bridges.

Onboard AI could help satellite companies deliver those answers faster.

The Technology Points Toward Space-Based AI Infrastructure

The demonstration also has a longer-term implication: it is a proof point for running AI in space.

The current model is small compared with the huge AI systems running in data centers on Earth. But learning how to run efficient models in orbit could inform larger space-based compute projects later.

This matters because several companies are now exploring whether space could eventually host data centers or AI infrastructure. The idea is still early, but the logic is that satellites can access solar power, operate beyond some land constraints, and potentially support workloads that do not need immediate terrestrial response.

Before that can happen at large scale, companies need to solve basic problems around power, memory, cooling, radiation, communication, and model deployment. Small onboard AI demonstrations help build that experience.

In that sense, YAM-9 is not only about Earth observation. It is part of a larger experiment in making space a computing environment.

Scientific Missions Could Also Benefit

The work may also help future science and exploration missions.

NASA JPL’s interest in onboard AI is partly connected to the idea of digital assistants for astronauts and explorers. On the Moon or Mars, astronauts may not be able to type complex instructions while wearing pressurized suits or operating in difficult environments. An interactive AI assistant that understands images, sensor data, and natural language could help them navigate, inspect equipment, identify geological features, or manage tasks.

Robotic missions could benefit too. Rovers, landers, and orbiters often operate with communication delays. If they can analyze their surroundings and make some decisions locally, they may become more capable and efficient.

This is especially important beyond Earth orbit, where sending all data back for human review is slower and more limited. Smarter spacecraft could help scientists find interesting targets faster.

The Risks Are Not Only Technical

The rise of autonomous satellite detection also brings policy and security risks.

A satellite that can identify objects or activity on its own could be used for disaster response or environmental protection. It could also be used for military surveillance, border monitoring, protest tracking, or commercial intelligence gathering.

The technology itself is neutral, but the applications are not. As AI makes satellite monitoring more responsive and scalable, governments and companies will need clearer rules about acceptable use, data sharing, and accountability.

There is also the question of model reliability. If a satellite AI misidentifies an object, misses a critical event, or generates a false alert, the consequences could be serious. Human review will remain important, especially in security, defense, and emergency contexts.

Onboard AI should improve speed, but it should not remove the need for verification.

Satellites Are Becoming Smarter Sensors

The YAM-9 demonstration shows that satellites are beginning to move beyond passive observation.

For decades, the main challenge was getting better sensors into orbit and returning data to Earth. The next challenge is making those sensors understand what they are seeing while they are still in space.

That shift could make satellites more useful in real time. It could reduce data overload, speed up decision-making, and create new commercial services. It could also make space-based monitoring more pervasive and raise new governance questions.

The milestone is early, but the direction is clear. Satellites are becoming smarter, more software-defined, and more autonomous.

A spacecraft that can find what it is looking for without human analysts on the ground is not just a better camera. It is the beginning of a new kind of orbital intelligence layer around Earth.