level: technical
a software system called navi-orbital was deployed on a low earth orbit spacecraft and operated on april 16, 2026. it performed the first known in-orbit use of a vision-language model for fully autonomous onboard inference. the system uses a local model, gemma 3, to look at captured images, classify them, and write text descriptions of what it sees and how features relate to each other. operators can also ask follow-up questions and get answers through natural language dialogue.
instead of sending traditional command sequences, operators retask the system using plain english prompts. a graph-based state machine called langgraph coordinates separate agents for detection and dialogue. this approach lets the satellite process earth observation data on its own, without needing to send everything to the ground. it helps close the gap between how fast data is collected and how fast it can be turned into useful information.
the demonstration shows that vision-language models can work in the constrained environment of a spacecraft. by handling inference onboard, the system reduces reliance on limited downlink bandwidth and human processing. this could lead to faster responses for time-sensitive tasks like disaster monitoring or environmental tracking. the use of natural language also makes it easier for operators to interact with the satellite.
why it matters: onboard ai reduces the need to downlink all data, enabling faster autonomous decisions for earth observation tasks.