Onboard Vision-Language Models

Loft Orbital demonstrated the first reported use of a vision-language model aboard an Earth observation satellite, enabling its YAM-9 spacecraft to identify targets directly in orbit from natural-language prompts. The system combined Google DeepMind’s Gemma 3 with NASA JPL’s NAVI-Orbital software, optimized to run on constrained hardware including an Nvidia Jetson Orin AGX processor.

The onboard AI successfully classified landscapes where natural and developed environments meet and identified infrastructure around railway hubs, reducing the need to transmit large volumes of raw imagery for ground-based analysis. Loft developed YAM-9 as a pathfinder for its infrastructure-as-a-service model, allowing third parties to deploy computing and sensing workloads in orbit, while the mission also provided insights into power and memory management for future space-based AI systems.

For researchers and commercial users, onboard AI promises faster, more actionable Earth observation with lower bandwidth demands, marking a significant step toward autonomous satellite constellations capable of continuous monitoring and intelligent, real-time analysis.