[jandrewrogers]

This is a Hard Problem because no fixed reference frame exists for registering position. For many applications we pretend that exists but with sufficient resolution (centimeter) the illusion of a fixed reference frame is shattered. Many companies, like Microsoft, want/need near pixel perfect registration. The challenge is worse than people imagine.

GPS positioning does not provide a fixed reference frame, even when it works as advertised, as it assumes some properties of reality are constant that are actually variable. But let's assume that it does provide a fixed reference frame for the sake of argument.

Physical objects are not fixed in any global reference frame. They can move quite a bit throughout the day, exhibiting significant Brownian and regular displacement relative to their mean position. No big deal, we'll just use a local reference frame, like the geometry of buildings and objects, right?

Local geometric relationships we treat as fixed are also quasi-randomized throughout the day. For example, the distance between two buildings can vary by centimeters over a day. With enough measurements you can sort of average out the local noise, but the precision is much worse than people find desirable.

We can't precision measure our way out of this problem because the things we measure don't sit still.

High-precision registration in physical reality is generally believed to be an AI-complete problem. This is a major hurdle for the vision of AR most companies have. You have a huge number of contradictory positioning cues, all of which are constantly changing, from which you need to synthesize a coherent positioning model that matches the one humans naturally perceive.

[Animats]

No, it's not an "AI-complete problem". It's just hard. With GPS for coarse position, inertial sensors for movement, depth sensing, and SLAM for fine position, it can work.[1] The drone industry and DARPA are working hard on this.[2] Right now, you can do it, but not with cell phone grade hardware.

[1] https://www.youtube.com/watch?v=iZ1psxcMvrQ [2] https://www.spar3d.com/blogs/the-other-dimension/nanomap-sla...

[jandrewrogers]

You can't build repeatable models of space for high-accuracy registration with big drone or car hardware either, I've worked with both. The geometry of space may rhyme but it never repeats. Those links don't address registration.

If you measure the environment with high-precision and use that to construct a geometric model of the space, and then come back a week later and measure it with the same instruments, the two spaces won't be congruent even for objects we normally think of as invariant, and the variability is sometimes surprising in magnitude. The noise floor for repeatable measurement out in the physical world is centimeters in most cases, regardless of the instrument precision used to measure it. This isn't a problem if you don't need particularly high-precision but people are inventing applications that do.

The software challenge is trying to position relative to previous measurements of the same space when the myriad positioning cues are contradictory. Knowing which of the totality of cues are relevant in context so that the software can appropriately adapt its positioning behavior to the change in geometry is the part that is usually deemed AI-complete by the people I know that have been working in the space a long time. There are many infamous example cases of humans being able to correctly register contradictory positioning information in context that we don't know how to algorithm our way out of currently.

Some of the drone work we did was actually measuring how the geometry of "fixed" spaces varies over time. The world around us moves a lot more than humans can perceive.

[Animats]

OK. I've never worked tighter than 15cm, for automatic driving, so I haven't seen that.