This successful experiment marks the transition from passive Earth imaging to autonomous space AI, which is able to independently interpret what it sees without sending gigabytes of raw data to ground stations. The experiment was conducted on board the YAM-9 commercial satellite, operated by Loft Orbital. A unique digital architecture was deployed on it.
Engineers integrated the multimodal Google Gemma model 3 (vision-language neural network Vision-Language Model), which can understand graphics at the same time, and natural human language. Software – special NAVI-Orbital platform, which was developed by the famous NASA Jet Propulsion Laboratory (JPL), optimized to work under the strict limitations of the satellite's on-board memory.
All calculations were performed by a compact on-board Nvidia Jetson Orin AGX processor, which has high energy efficiency. Researchers sent text messages to the satellite (requests) in plain English. Example, The AI was asked to “find infrastructure around railway junctions” or “identify zones, where human construction borders on wild nature".
The satellite was taking a picture, the on-board processor instantly scanned it, found the necessary objects according to the description and sent only a ready report and target coordinates to Earth, instead of transmitting a massive output frame.
Why this approach of passively capturing the Earth has a perspective:
Saving traffic: Satellites no longer need to waste scarce energy and bandwidth on gigabytes of "empty" images (example, with clouds or an empty ocean).
Constant patrol: The spacecraft turns into an autonomous "guard". He can be given a task: "Watch the coast and signal, if an oil slick or military equipment appears”.
Reaction speed: Time from the moment the event was recorded (fire, flood, troop movement) to alert the services on Earth is reduced from hours to seconds.
Source: https://universemagazine.com