Since it mentions sensor fusion, is this not just* combining the data from the GPS, compass, accelerometers (and any other sensor input from the car e.g. comparing photos, IR, wifi signals) to smooth out any irregularities in a single sensor and making better guesses where they all agree. If I'm right then the demo video is cheating a bit by only showing the GPS input initially and the rest of the info seems to come from thin air, which might be why people think it's working with the underlying street data. Dead reckoning with a compass and accelerometer and/or wheel odometer/tachograph might also have produced a similarly wonky line, but then you have two different wonky lines to average together to get a better approximation of the true path.

Note this supposition is based almost entirely on doing the first 3 weeks of the Udacity course CS373: Programming a Robotic Car, but it seems to fit quite neatly with the stuff they talk about i.e. a car/robot moving, guessing how far it's moved, taking sensor snapshots of its environment then cleverly combining the data to figure out accurately where it is.

Off-topic: scrubbing back and forth in the blog video (to get a better idea of what was happening) seemed incredibly smooth for me, is this a side effect of Youtube/Firefox using HTML5 video? (edit: trying it out with Flash, I get a fancier, but less useful pop-up when I try to scrub, a tech limitation of Flash or just a design decision?)

* not meaning to underestimate the hard work and genius that underlies this technology, but the same concept is used in such everyday items as Wii remotes and suchlike.

is any geometry point?

I mean, when you have a loop (say, around the manhattan grid), then I can understand how everything works when you have one hard(er) constraint such as, "the two farthest points visited are 2nd street at avenue C and 112th street at 12th avenue" - and get everything else aligned accordingly.

But if no external geometry constraint is involved (roads being the most abundant, but not the only type of external geometry constraint), then this reader has been unsuccessful with the exercise.