In a historic milestone reached in April 2026, a satellite named Yam-9 became the first to autonomously identify objects of interest on Earth using an onboard vision-language model, bypassing the need for human analysts to review raw data. This breakthrough signals a transformative shift in how satellites process data and could enhance real-time space-based monitoring capabilities.
- Yam-9 satellite uses vision-language AI to independently find targets on Earth.
- Onboard AI reduces data sent to Earth and enables real-time monitoring.
- This technology paves the way for autonomous satellite constellations with broad surveillance capabilities.
What happened
In April 2026, the Yam-9 satellite achieved a first-ever feat by autonomously identifying its targets on Earth without human intervention. The satellite integrated a vision-language model (VLM) called Gemma 3 developed by Google DeepMind, running on an Nvidia Jetson Orin AGX GPU tailored for limited hardware environments in space. This feat involved interpreting natural language requests and analyzing sensor data directly onboard.
This onboard AI allowed Yam-9 to classify complex scenes where natural landscapes intersect urban and infrastructure elements, significantly streamlining data analysis processes. Typically, satellites would downlink vast amounts of raw data for ground-based analysts, but Yam-9’s demonstration showed the potential for independent, in-orbit data triage and interpretation.
Why it matters
The successful deployment of a vision-language model in orbit marks a fundamental step towards more intelligent satellites capable of real-time decision making. This development reduces reliance on downlink bandwidth and extensive ground-based processing, enabling satellites to deliver higher-value insights more rapidly. It could revolutionize not only Earth observation but also how space assets are operated in the future.
Moreover, the technology introduces possibilities for continuous, automated monitoring applications—such as border surveillance or infrastructure inspection—where satellites respond dynamically to natural language commands. The proof of concept by Loft Orbital and NASA JPL demonstrates the practical viability of embedding sophisticated AI in space hardware, encouraging broader adoption across the aerospace industry.
What to watch next
Further development of onboard AI in satellites will likely focus on scaling these models for larger constellations to provide global, real-time coverage. Loft Orbital aims to expand its fleet to between 50 and 100 AI-enabled satellites similar to Yam-9. Other industry players, including Planet Labs and Kepler Communications, are exploring their own onboard AI initiatives, portending a race to deploy advanced edge AI in space.
Additionally, advancements in efficient power and memory management for AI models in orbit will be critical to supporting the next generation of space compute infrastructure. Beyond Earth observation, applications may extend to assisting astronauts on the Moon or Mars by interpreting commands without the need for physical input devices, thus broadening the scope and impact of AI in space missions.