https://www.theguardian.com/technology/2018/aug/28/driverles...
Admittedly, on a very limited basis. But there is one thing to consider: There are a lot of city cores where average speeds are low enough that such specialised services are viable way of reducing the need for fully flexibly autonomous cars.
E.g. here's an autonomous shuttle manufacturer:
These are also in actual use some places (though lagging laws means that many places there's still an attendant), and there are other manufacturers.
So while the full general purpose replacement may take longer than hoped, a lot of more special purpose vehicles like that will nibble away at transport needs, and illustrates that if we look at it as not a problem of replacing cars, but a problem of replacing journeys, it's possible to solve parts of it piecemeal long before we can fully replace cars.
The main problem with having humans take over control in situations the car can't deal with is if the switch needs to happen at speed. If you're crawling along near some accident getting input from someone directing traffic, it doesn't seem like it should be a problem to get the human's input with voice commands, or touch on a dash-mounted tablet, as to what to do next.
It's even worse. I remember seeing prototypes of autonomous cars a little bit above 20 years ago. And I don't know when that research started.
I still don't understand why everyone is working so hard trying to make smart cars when it seems like smart roads that control dumb cars would be a simpler, if more expensive, solution. Why not make it a closed system?
If they don't broadcast it with a valid signature, just ignore them, they're not authorized to give any instructions.
> turn the road system into a closed system
You can turn it into a "closed enough" system, with every vehicle having a radio locator, and everything else being treated as an unknown obstacle... that someone could get fined for putting it on the road.
Here's the actual status quo: after a decade of research into autonomous cars specifically (not to mention the decades of AI and systems research before that), we are barely at the point where a car can function safely with Level 2 autonomy, in good conditions.
These prognostications about eliminating traditional cars are assuming reliable Level 5 autonomy in all conditions. And unlike most software, you can't just get 99% of the way there and call it done. If you want to remove the steering wheel, it has to work ALL the time (or close enough to make no difference.) They say that 80% of the work is in the final 20% of the project? We haven't even finished the first 30% of the work yet.
Even more than that, nobody seems to be talking about the fact that to get this to work reliably, you basically need full general intelligence (unless you can manage to turn the road system into a closed system, like rail is). For example: following verbal or gestured instructions from a police officer or construction worker. We don't even have _theories_ on how to build intelligence like that, yet.
Will it happen eventually? Sure, probably, eventually. Will we get a lot of use out of Level 2 and Level 3 autonomy in the mean time? Already happening. Will we be taking robotic taxis in the next 20 years? No way, not with current approaches absent a breakthrough in general AI.