I remember talking to Apple's Kim Silverman, ¿head? of speech recognition, somewhere in the end of the '90s. He said you had to spend a few hundreds of hours training the recognizer to get at a good level. That's not a lot more than one would spend learning to touch type at speed, but a large fraction of _that_ time is productive; training a speech recognized wasn't.
Also, touch typing, once you can do it, works everywhere; speech recognition didn't work as well when there was background noise or echos. So, few people were willing to invest the time. And they probably were right.
I also agree with your statement that we will solve restricted, but useful domains first, and that a general solution for autonomous driving will be a long way of.
I also agree roads will be adapted to the self-driving cars. It isn't rocket science to embed a steel wire in the center of each car lane that a robot car can detect, for example.
I remember a television program where the engineers of some European car manufacturer thought universally usable self-driving cars would be safer than human driven ones in 10 years or so, but their human factors specialist said it would be something like 40 years before they would hit the road because she knew you can't expect humans to fully attend to driving on very short notice for a few seconds each week.
(that show also said, and showed, that features such as lane assist are more limited than the cars are capable of because of human factors)
This could be a big improvement for both safety and driving comfort and seems as if it would be a much more amenable to solving over, say, a 10 year horizon than a cross-town Manhattan taxi ride at rush hour.
In the vast majority of the cases, that sign will have been seen by another autonomous car on a sunny day before it had graffiti, and stored in the map database. You should be comparing a whole fleet of autonomous vehicles learning from each other versus an individual human.
I've felt the same way after watching various autonomous driving demos (like Tesla's[1]). But then I remember my experience of actually trying the same speech recognition software myself; under unrehearsed real-world conditions with edge cases and human mistakes, the technology performed terribly. Any "intelligence" I'd seen in the demo was essentially smoke and mirrors.
Granted, current autonomous vehicle technology incorporates a lot more artificial intelligence than those speech recognition demos ever did, but the challenge is also significantly greater (and more life-critical). Sure, your sensors and cameras might be able to read signage correctly in 99% of conditions, but when there's graffiti on the sign at night during heavy rain, all bets are off. The human driver may have difficulty also, but the human driver has real intelligence and life experience in a variety of domains, enabling them to make inferences based on more than just statistical probabilities.
In my opinion, prior to solving all the edge cases at an acceptable level using AI, we'll solve a different problem allowing us to sidestep many of those challenges; we'll incrementally start building (and converting to) smart roads where only smart vehicles are allowed. Obviously it won't be everywhere, but the most important routes will be covered, and your safety on those routes will be much higher than it would be on traditional roads with AI or human drivers.
[1] https://www.tesla.com/videos/autopilot-self-driving-hardware...