Thanks for the summary.

I suspect you could get this to FAR higher accuracy if you combined it with a recent upload of Starlink et al LEO constellation ephemera, an initial GPS fix at launch, and a planned flight path, because LEO constellations are bright foreground objects (high location-specific parallax differences against background stars) at apparent magnitude of about 5.0.

This is simultaneously not reliant on perfect vertical attitude sensing coming off the autopilot IMU, you can do it purely photometrically.

The limitation is that this is a dawn/dusk thing, in the middle of the night there isn't a ton of light reflected and in the day you're limited by scattered daylight.

EDIT: Medium orbit satellites outside Earth's umbra but within view still provide some sort of visual fix. I wonder what the math is like for the GSO belt at midnight?

EDIT2: Or the Moon.

I would assume the same. Operation in GNSS-denied environments is critical for military navigation systems. Comparatively, for civilian uses, it's an addon that provides low accuracy, and potentially high development or equipment cost (Maybe not for a cel nav camera, but for Ring Laser Gyro INSs etc)

GNSS is very accurate, and receivers are cheap, but its reliant on satellite signals makes relying on it a liability in adversarial uses.

Cel nav isn't self-contained in the way an INS is, because you need a clear LOS to the stars. But, it's useful on a clear night when your GPS is jammed.

Ukrainian and Russian drones already do that. They use simple visual navigation for terminal guidance: https://www.rockingrobots.com/ukraine-drones-able-to-navigat...

At this point, it's pretty clear that this type of functionality is out of the bag. Any significant actor can easily replicate this with minimal effort, given the advances in AI.

Don't get distracted https://news.ycombinator.com/item?id=42388354