NASA Tests Google's Gemma LLM in Orbit for Satellite Image Analysis
NASA's Jet Propulsion Laboratory has successfully demonstrated the use of Google's Gemma 3 large language model (LLM) in orbit, marking a significant advancement in how researchers interact with spacecraft. The NAVI-Orbital system, developed by NASA JPL and Loft Orbital, utilized a compressed, 4-bit version of Gemma 3 to analyze images captured by Loft Orbital's YAM-9 satellite. This demonstration is the first in-orbit use of a vision-language model to process imagery directly from a satellite's sensors.
This new approach allows scientists to send natural language prompts to the spacecraft, enabling direct interaction and analysis, a departure from previous methods requiring highly structured commands and dedicated operations teams. NAVI-Orbital achieved 88 percent accuracy in classifying images from a benchmark dataset, notably without specific fine-tuning on that data. The system ran on an Nvidia Jetson Orin AGX module, requiring only 8GB of memory, highlighting the model's lightweight nature.
Beyond image analysis, the technology offers a potential solution to bandwidth limitations by enabling "semantic compression," where satellites transmit text summaries of key information instead of large raw image files. This could expedite critical processes like wildfire detection, reducing delays from up to 90 minutes to near real-time. While currently isolated from flight controls, the NAVI-Orbital system's success paves the way for future applications, including potential AI companions for astronauts, enabling natural language interaction within spacesuits.
The successful orbital deployment of Google's Gemma 3 LLM by NASA's NAVI-Orbital system demonstrates a pragmatic approach to leveraging advanced AI in space constrained by power and bandwidth. By utilizing a compressed model on specialized hardware, the project overcomes significant logistical hurdles, enabling on-board data processing and reducing reliance on costly, slow downlinks. This shift from raw data transmission to semantic compression offers a compelling economic and operational advantage for satellite missions, potentially democratizing access to timely insights from space. The system's ability to interpret natural language prompts also hints at future human-AI collaboration in space exploration, aligning with the long-term trajectory of AI integration into complex operational environments.
AI-generated to prompt reflection — not editorial opinion, not advice, not a statement of fact. How this works.