Source linked

النقلة Yam-9 تعمل على Gemma 3 VLM في الأقمار الصناعية، وتجد الأهداف دون مساعدة بشرية

لأول مرة، تم تحديد الكائنات الأرضية بصورة مستقلة باستخدام نموذج لغة الرؤية، مما أدى إلى إغلاق بيانات المراقبين.

loft orbitalgoogle deepmindgemma 3nasa jplsatellite aivision language models

In April 2026, a 400-kilogram satellite named Yam-9 used a vision-language model to identify infrastructure around railway hubs and classify ecotones without a single human analyst looking at the raw data. That's the first time an Earth observation satellite has autonomously answered a natural language query from orbit.

What Yam-9 Did Differently

Typical satellites dump terabytes of imagery to ground stations, where teams run object detection algorithms or manually scan for targets. Yam-9, built by Loft Orbital under its infrastructure-as-a-service model, runs a Nvidia Jetson Orin AGX GPU paired with Google DeepMind's Gemma 3 vision-language model. NASA JPL's NAVI-Orbital software package, led by technical lead Juan Delfa Victoria, acts as the harness, trimming library dependencies and memory footprint to fit the edge environment.

Loft's head of AI, Paul Lasserre, told TechCrunch the VLM enables logic like "monitor this border for me, and let me know when something is suspicious." The satellite can now triage data on orbit, relaying only relevant observations. That changes the economics of space sensors: you pay for downlinked answers, not raw pixels.

From Pathfinder to Patrol Constellation

Yam-9 launched in fall 2025 as a pathfinder for orbital AI. Loft currently operates 12 spacecraft. Lasserre estimates that a constellation of 50 to 100 satellites like Yam-9 could provide real-time coverage of any point on Earth. That's an always-on patrol layer in space, not a store-and-forward imaging archive.

Competitors are moving. Planet Labs flies Jetson Orin processors on some satellites for simple object detection and confirms VLM research is underway. Kepler Communications, which operates the largest GPU cluster in space, declined to confirm VLM deployment due to NDAs but acknowledged "several undisclosed use cases" since its compute-heavy satellites launched in January.

The next step: deploying larger-scale AI infrastructure in orbit. Lessons from power and memory management on Yam-9 will inform how companies build future constellations. And the original spark for NAVI-Space came from JPL's Taran Cyriac John, who envisioned digital assistants for astronauts on the Moon or Mars. "We're thinking, okay, you have astronauts with pressurized suits, they cannot be tapping on a keyboard," Delfa Victoria said. "So, how about we provide an interactive AI assistant?" That assistant just proved it can work from a thousand kilometers up.


Source: A satellite just learned to find things on its own - here's what that means
Domain: techcrunch.com

Read original source ->

External source stays available while the OJO article and comment thread stay local.

Comments load interactively on the live page.