Satellites Go Solo: The New Era of Space Discovery

In a groundbreaking achievement this April, an Earth observation satellite successfully identified its targets autonomously, without requiring human analysts on the ground. This milestone marks the first documented deployment of a vision-language model (VLM) in orbit, offering a compelling preview of how artificial intelligence could fundamentally transform the capabilities and intrinsic value of space-based sensors.

Typically, satellites transmit vast quantities of raw data to Earth, where analysts utilize machine learning algorithms or manual visual inspection to interpret the information. However, onboard Yam-9, a spacecraft developed by space infrastructure company Loft Orbital, a specialized software package from NASA’s Jet Propulsion Laboratory (JPL) successfully pinpointed areas of interest in direct response to natural language queries.

The vision-language model powering this demonstration, Google DeepMind’s Gemma 3, is purpose-built for "edge" applications, meaning it is optimized to run efficiently on limited hardware far from conventional data centers. VLMs integrate the contextual understanding inherent in large language models with advanced image analysis capabilities. During the test, researchers effectively tasked the model with classifying sensor data where natural environments converge with human development, and also to identify infrastructure around railway hubs — tasks which it successfully completed.

This demonstration holds significant implications for two primary reasons. In the near term, it promises to vastly enhance the utility of space sensors by enabling initial data triage directly in orbit, thereby significantly reducing the overwhelming volume of raw data that analysts currently process. Over the longer term, it serves as crucial proof of concept for the eventual deployment of larger-scale AI infrastructure in space.

“It opens the door to always-on, patrol layers in space,” Loft’s head of AI, Paul Lasserre, informed TechCrunch. He further elaborated, “If you have a VLM, you can have logic—like ‘monitor this border for me, and let me know when something is suspicious,’ and interact back and forth with the satellites.”

Loft Orbital designs its spacecraft to serve as versatile platforms for third-party customers, operating under a business model more akin to infrastructure-as-a-service than traditional satellite manufacturing. A recent agreement involved the construction, launch, and operation of six new satellites for EarthDaily, which will analyze and market the data collected onboard. Yam-9, launched in the fall of 2025, acts as a pathfinder for the company’s orbital AI projects and features an Nvidia Jetson Orrin AGX GPU, a leading chip for space computing applications.

Juan Delfa Victoria, a technical leader within NASA JPL’s AI group, led the development of NAVI-Orbital, the software package that effectively served as the integration harness for the Gemma 3 VLM. While Gemma 3 is a commercially available model, software engineers dedicated significant effort to streamline NAVI-Orbital, reducing its library dependencies and memory footprint for optimal orbital performance.

While this marks the first reported instance of a VLM being used in orbit, it is anticipated that other companies will soon follow suit. Planet Labs, for example, operates satellites equipped with Jetson Orin processors. Currently, these are employed for simpler object detection tasks, but a company spokesperson confirmed that research is actively underway into other advanced AI applications, including VLMs.

Kepler Communications, which operates the largest collection of GPUs in space, declined to confirm whether it had deployed VLMs due to non-disclosure agreements with partners. However, the company did note that there have been “several undisclosed use cases of our compute environment” since their spacecraft launched in January.

“Now that we’ve proven the concept, that’s really the direction of travel,” Lasserre stated. The overarching goal is to expand the constellation to ensure continuous, real-time coverage across the entire Earth, a feat he estimates would require between 50 and 100 satellites similar to Yam-9. (Loft currently operates 12 spacecraft in orbit.)

The invaluable lessons learned from deploying these smaller models in orbit will inform how companies approach the deployment of larger-scale computing infrastructure in space, particularly addressing the critical, though often less glamorous, challenges of power and memory management.

These advancements could also pave the way for novel scientific instruments. The initial concept for NAVI-Space originated with JPL Researcher Taran Cyriac John, who envisioned digital assistants to support astronauts exploring the Moon or Mars.

Delfa Victoria elaborated on this vision: “We’re thinking, okay, you have astronauts with pressurized suits, and you know they cannot be tapping on a keyboard, whatever they want to do is complex.” He continued, “So, how about we provide an assistant, like in video games and in movies, where you see an AI which is interactive?”

Just don’t call it HAL 9000.