Eyes in the Sky Get Smarter: How Orbiting Software Thinks Without Human Help

For the first time, an Earth observation satellite successfully located a target on the ground entirely on its own, without waiting for human analysts to tell it what to look for. This milestone happened back in April, marking the very first time a spacecraft ran a vision-language model while floating in orbit. This technical leap gives us a clear look at how artificial intelligence can reshape space hardware and redefine the value of orbital intelligence.

Standard observation satellites operate as simple data collection tubes. They record giant chunks of raw images and beam everything down to analysts on Earth, who then use ground-based algorithms or manual review to piece together the picture. This new test flips that model completely. Onboard a spacecraft named Yam-9, which space infrastructure firm Loft Orbital built, a custom software package managed by NASA’s Jet Propulsion Laboratory successfully read natural language queries and pinpointed specific targets directly from the live camera feed.

The engine steering this orbital test is Google DeepMind’s Gemma 3. This vision-language model is custom-built for edge applications, meaning it can run directly on compact, power-constrained hardware rather than relying on a giant, air-conditioned data center. These models combine the deep contextual understanding of text algorithms with the ability to inspect complex images. During the flight trial, researchers asked the model to look at the sensor data, identify where human development pushed into natural environments, and call out specific infrastructure like railway hubs. The system handled the request perfectly.

This test holds massive implications for global monitoring for two main reasons. In the short term, it makes space hardware far more efficient. Satellites can triage data right on the chip, filtering out useless images of clouds or open ocean and drastically reducing the flood of raw files that analysts currently have to filter through. In the long term, it serves as a concrete proof of concept for running massive, interconnected computing networks in space.

Paul Lasserre, head of AI at Loft Orbital, explained that this software introduces permanent automated patrols in space. Instead of scheduling rigid photography passes, you can feed a satellite a conversational prompt, like asking it to watch a specific border and send an alert if it detects unusual movement. This sets up an ongoing, interactive dialogue between the ground team and the spacecraft.

Loft Orbital builds its business model around providing infrastructure as a service, bypassing traditional custom satellite manufacturing routes. They construct and operate modular platforms for third-party clients, like their recent deal to run a constellation for EarthDaily. Yam-9 launched in the fall of 2025 to act as a pathfinder for these orbital computing projects, packing an Nvidia Jetson Orin AGX GPU, which is the current industry standard chip for heavy processing in space.

Juan Delfa Victoria, a technical leader within the AI group at NASA JPL, guided the development of NAVI-Orbital, the specialized software bridge that links Gemma 3 to the spacecraft hardware. Because they pulled a standard, off-the-shelf model into a highly restricted environment, engineers had to aggressively strip down the code package to fit within the limited memory and power grids of the satellite.

While this represents the first confirmed live deployment of a vision-language model in orbit, rival space firms are already moving in the same direction. Planet Labs currently runs a fleet of satellites equipped with identical Nvidia Orin processors, though they currently use them for simpler object detection tasks instead of conversational reasoning. Kepler Communications, which commands the largest cluster of GPUs in orbit, declined to say if they are running similar models due to strict non-disclosure agreements, though they confirmed multiple unmapped enterprise trials running inside their space environment since January.

Loft Orbital currently operates twelve spacecraft, and the team intends to use the data from Yam-9 to scale up a larger constellation of fifty to one hundred smart satellites to secure total, real-time tracking across the entire planet. Beyond earth monitoring, this edge computing model lays the groundwork for deep space travel. The original idea for the project started when researchers envisioned conversational digital assistants to guide astronauts exploring the moon or Mars. When wearing heavy, pressurized spacesuits, typing on a keyboard is impossible, so crews will need interactive, intelligent voice assistants to manage complex tasks in deep space environments.