matter how you define "self-driving car". Glorified cruise-control that can steer? We're almost there. Better than an average human in most situations? We need some breakthroughs, at least in computer vision. No need for licensed drivers anymore? I don't see it happening in the predictable future.
I'll admit my AI and computer vision background is way out-of-date. I did both in college, but that was 20 years ago. The real world has moved me from fanciful cutting-edge inventions to real-world not-quite-as-cutting-edge engineering. (16nm mixed-signal IC design, if anyone cares.) Which is why I'm trying to patch in some links to current discussions on the topics.
Since I was the one pointing out that current computer vision
probably can't handle driving in a real Minnesota winter, and real-world conditions can affect the behavior of cameras/radar/etc, I was interested in the comment upthread that "since Volvo is doing it, they're probably working on that." So I found some articles on the Volvo autonomous cars. They're putting 100 of them in consumer hands in 2017! They'll handle "heavy traffic and emergency situations"! They can "read" road signs! That sounds awesome!
...but the system depends on a preloaded "high definition 3D map," rather than being able to identify "this is a road and this is not" by camera & radar. It'll only run on "30 miles of roads around Gothenburg", and "its test roads will be without pedestrians, cyclists or on-coming traffic." So, it's still just freeway-only cruise-control, plus braking and steering; the same level of incremental change that keithl was talking about on the first page of this thread. "In cases where the autopilot must shut off due to weather or malfunction, the car will prompt the driver to take over." So no workarounds for for seeing snowy roads, or using muddy radars or road-salt-glazed or insect-splattered cameras yet. (Quotes from here
, and here
And, just today, there's news about Ford
working on autonomous cars that can
drive in the snow. But, again, "To deal with [snow flurries and other adverse weather conditions], Ford is having its LiDAR sensors look for landmarks that are higher up, like road signs, and combining that information with stored high-resolution 3D maps of the environment that were captured using LiDAR in better weather conditions." So, a preloaded "google-street-view camera car"-gathered 3D map, taken when there's no other vehicles/people/etc around to mess up the view. That's not what I'd consider real-world-applicable technology. "We eventually want our autonomous vehicles to detect deteriorating conditions, decide whether it's safe to keep driving, and if so, for how long." So even Ford's look to the future is not for "fully autonomous", but for "how much precipitation can we handle before we have to give up".
If we were on the edge of more-than-just-a-glorified-cruise-control, or the-car-sees-as-well-as-a-human, the big marketing press releases would be about that
. Instead, they're about "we can handle driving on 30 miles of studiously pre-mapped roadways in pleasant conditions if we know there's nothing but same-direction vehicle traffic around." So, pick your definition of "self-driving car", and tune your expectations appropriately. Define it generously, and we're already there. Define it strictly, and... maybe your kids will have it.
And there is sort of a parallel to the "uncanny valley" problem... the closer you get to perfect, the more dangerous a self-driving car becomes. A car that can drive itself 20% of the time means all the cars still have drivers who drive almost all the time, and still do things by muscle memory or well-trained instinct. The roads will be about as safe as they are today, or safer. A car that drives itself 99% of the time, and only requires human intervention in the worst 1% of situations, is very dangerous. Now you have a road full of complete newbies dealing with only the hairiest of traffic and weather situations. (I would have only driven ~100 miles last year, not ~10,000). If you're doing this for safety reasons, you need to make the jump from 80% (or 66%, or 90.3%, or pick-your-own-number, whatever you think leaves the driver with "enough" experience), to 100%, in one swell foop. No insurance company is going to insure a driver with a 99.44%-self-driving car, unless the other 0.56% is "shut it down" and not "now let the human do it".