YAM-9 Satellite Demonstrates On-Board AI Image Interpretation

The YAM-9 satellite, launched under NASA’s NAVI-Orbital program and operated with Loft Orbital, carries Google DeepMind’s Gemma 3 vision-language model. Unlike traditional Earth-observation platforms that downlink raw imagery for ground processing, YAM-9 can accept natural-language prompts—e.g., “find all railway hubs”—and return classified results directly from orbit. Baseline ground tests on 7,960 images achieved 88.2 % accuracy across categories such as residential, agricultural and mountainous terrain. To date the satellite has performed two live in-orbit captures, and the team envisions a constellation of roughly 100 similar units to deliver continuous global monitoring.

ScienceAlert

Loft Orbital’s senior marketing manager Sarah Preston described the system as a continuous lookout that can be tasked to monitor specific phenomena—such as oil spills or new construction near borders—and report only when thresholds are met. Senior system engineer Juan Delfa Victoria framed the interface as an interactive assistant comparable to those in video games, allowing operators to ask questions without writing command sequences. Head of AI Paul Lasserre said the capability opens the door to always-on patrol layers in space. The researchers emphasized that the NAVI-Orbital paradigm replaces traditional software uploads with prompt editing, shortening retasking cycles and supporting a planned marketplace of AI agents across an eventual 100-satellite constellation.

ScienceAlert

Analysts warn that constant AI-driven observation could enable pervasive surveillance, raising privacy and civil-rights concerns at borders and ports. Delegating high-stakes image interpretation to algorithms may obscure accountability, and the system’s resilience to adversarial prompts remains untested, highlighting ethical and security gaps.

ScienceAlert

Performance metrics derive from ground-based tests; the two in-orbit captures offer no independent validation. The research appears only as a non-peer-reviewed arXiv preprint, and the authors note the study does not assess robustness against malicious inputs.

ScienceAlert

“This AI can actually 'see' what's in the image and identify what the analyst is looking for, such as bridges, highways, specific bodies of water, or signs of natural disasters like flooding and wildfires.” — Sarah Preston, Senior Marketing Manager, Loft Orbital

ScienceAlert

“So, how about we provide an assistant, like in video games and in movies, where you see an AI which is interactive?” — Juan Delfa Victoria, Senior System Engineer, NASA JPL

ScienceAlert

“It opens the door to always-on, patrol layers in space,” — Paul Lasserre, Head of AI, Loft Orbital

ScienceAlert

“Tasking a satellite to recognize a new feature has historically required writing command sequences, revalidating onboard software, and uplinking new binaries,” — Researchers (Delfa Victoria et al., arXiv 2026)

ScienceAlert