2004 Nobody won the DARPA grand challenge... no car completed the course.
2005 5 vehicles completed the course.
2007 they switched to an urban course having to obey the rules of the road and six teams finished the course.
And prior to 2004 there were already cars being demonstrated as autonomous with minimal human input. Poster above linked to one from the 90's.
That is rapid progress.
From 2007 to 2016 we have seen pretty steady progress with commercially available features for things like automatic parking, automated braking and collision avoidance, widespread use of GPS navigation (via smartphones and built-in) and more recently the fully autonomous highway driving from Tesla (yes I've seen the people reading books while "at the wheel").
Tesla themselves say it is not autonomous. Automated braking was demonstrated in the 70's on a Volvo. You list it as some sort of breakthrough.
And Google has been pretty open in their fully autonomous car project with two different cars one based on an off the shelf lexus and another custom built electric vehicle: 1.5 million miles driven and "currently out on the streets of Mountain View, CA, Austin, TX, Kirkland, WA and Metro Phoenix, AZ"
And we are seeing Uber's autonomous efforts play out in Pittsburgh. Multiple companies, multiple projects, multiple on-street implementations that are getting better and better.
Firstly, Uber themselves say that the drivers in their autonomous cars will remain there for the foreseeable future. Their words, not mine. Secondly, you yourself pointed out that they are at the same place they were 10 years ago - that's not "getting better". There's been incremental improvements, sure, there's been more widespread takeup of the existing technologies, sure... but the state-of-the-art today in autonomous cars is basically the same as ten years ago. The cars are mostly autonomous.... up until they aren't! There's no new tech. The improvements have been tiny and incremental, and not enough to replace a human driver yet.
What people don't get is that we've seen this movie before, in the field of AI. The AI titans were 99% there with regard to thinking machines. Turned out that that 1% was unattainable. Same deal with SDC software: they may have checked off 99 out of 100 requirements for the software, but that remaining requirement (requiring corrections from an alert human at the wheel) may or may not get there.
The AI winter that followed the "we're 99% there!" AI boom was painful as many researchers had to admit that they had an intractable problem on their hands. We are currently seeing the same thing in autonomous cars: everyone is so certain that they're almost there and the remaining problems will be easy. Well, if the remaining problem was that easy, it would have been solved when it first came up in the 90's.
The problem isn't the hardware (GPS, Vision capture, distance sensor, lane detection, etc), it's the software. No company has demonstrated software-controlled cars that can function without regular human input.
And this problem isn't an easy one, which is why it hasn't been solved. The only improvements you can list came from better and more accessible hardware.