SatNav
Team led by a Caltech alumnus and Adobe ML Engineering Manager with deep expertise in local RAG, Python, Swift, and on-device copilot development.
YouTube Video
Project Description
Problem statement
Cubesats deploy different sensors to guage where they are in space. Sun sensors, star trackers, gyros and magnetometers are expensive and also take up space and add weight. If an off-the-shelf camera that does earth imaging can do double duty and help with attitude control, it would lower the expense of satellite development.
Solution
A local-first, agentic system that predicts CubeSat telemetry (geolocation, altitude, datetime) from six-camera imagery. It uses FunctionGemma on Cactus for fast on-device tool calling and Gemini as an intelligent cloud fallback when local confidence is low, with a web UI to run analysis and compare predictions to ground truth on a map.
Prior Work
n/a