And that's one of the big problems with autonomous vehicles: they only know what the clear, defined rules are. Exceptions are difficult to understand, and how humans bend or elide over or shade or selectively interpret those rules are impossible to program in.
Remember: these aren't actual "artificial intelligences." They're sophisticated expert systems … which means they just plain old follow a set of programmed rules, which means messy real life and messy real life drivers are always going to be trouble.
It's an amusing side light for me, as I am currently in a household with a student driver who is very rule-abiding. One of the most difficult things for me as a teacher / passenger has been trying to convey when it's okay (or sometimes even preferable or safer) to work around those rules, or deal with ambiguous situations that require a judgment call.
My daughter is a lot smarter than an autonomous vehicle. I'll be curious to see how she — and such cars — evolve in their driving abilities over time.
Originally shared by +Les Jenkins:
The Real Problem With Self-Driving Cars: They Actually Follow Traffic Laws
In the century or so that people have been driving, two different sets of rules have developed: The official laws that we’re supposed to obey, and the unofficial code of the road that bends a…
I've long thought that the dangerous part of this is the transition. You need either all human or all self drivers. I expect to see segregation pretty shortly. Self only lanes, self only cities or zones, etc.
I'm curious how they handle left turns. Legally you're supposed to go forward until you are in the lane and then turn. However if the folks coming towards you are also doing a left turn, that doesn't work, you both need to cut the corner.
+Kee Hinckley It may vary by locality and instance. But there is usually enough space in the intersection to move forward and then turn, and proximity detectors should handle that.
I'm a bit more curious about double-left turns (which we have a plenitude of), as those often end p with very sharp turns when intersecting the other street (if people there have pulled fully up to the line), as well as ambiguity as to what the guy to your left/right is doing, esp. of the dotted lines crossing the intersection have worn off or are otherwise obscured.
Which raises the further question of how the autonomous vehicles handle snow-obscured roadways.
+Dave Hill The comments I've seen said that snow and construction are bad news. They really want those lines on the road. I think we saw that with a Tesla issue recently, where there were two lines, one new one on top for the construction change. Snow covered roads is really interesting, since it's not just a matter of eyeballing where the lanes ought to be–you want to drive where the previous cars have been driving–which means recognizing ruts in the snow and picking the best of them, and doing a lot of following the car in front of you.
I recently hired a car with "adaptive cruise control" for a week and I'm sure the gap it was leaving between me and the car in front was the recommended following distance for whatever speed I was travelling. What that meant, on multi-lane highways, was that people often cut in front of me, and then the car would brake to preserve that following distance, and then I would find myself going significantly slower than the other traffic, so more vehicles would pass, cut in, etc.
I believe that experience was firmly in the "uncanny valley" of autonomous driving: it all seems OK for a while, and then it does something that few human drivers would do.
The fully autonomous vehicles from Waymo and the like are obviously much more sophisticated about this kind of situation, and having driven over 3 million miles now they've had the benefit of way more miles on the road than I've managed since I first found myself behind the wheel in 1972 or so (I'd estimate I might have driven 700Mm by now). Not only do they have more experience than me, they have a far better sensory experience than I do: with information from all around the vehicle, in multiple spectra, in both active & passive.
The road rules are the expert system underpinning this stuff, but all of the little micro-negotiations that occur as we drive around that are layered onto those rules are what the deep learning is for – like +Kee Hinckley's one regarding left-turns into multiple lanes against oncoming left turns – I think that's even covered in a blog post on the Waymo site.
I'm pretty sure I'll never sit behind the wheel of any car in New York though – along with Paris, Beijing, Mumbai, Singapore, Rome, Tokyo, Miami, Hong Kong… but once autonomous cars have done a million miles in each of those cities as well, then I think they will surely be expert, and quite possibly better and safer than any human. I'm quite willing to believe they are already.
So as soon as an autonomous car makes it into my price range I think I'll be lining up with my money.
One plot hole I had with the movie Logan was the scene where the self driving trucks seem to make a lane change with a car in lane, or disregard the pedestrians in the highway. The first is a clear, defined rule of the road, the second was something that could be accounted for with proximity detectors and safety features. I thought the writers assumed the I, Robot self driving car rules were still in affect and not the real life cars being developed today. However, Logan did mention later that the corporations that owned the trucks potentially reprogrammed them to cause the farmer trouble, so I had to reevaluate that criticism.
+andrew mcmillan Yeah, that whole "How far you should stay behind the leading car to stay safe" vs "How far you can stay back without someone cutting in front of you" thing will be a challenge, esp. against manual drivers who game the system.
That said, it's an argument for fully autonomous cars — if I'm reading a book, I probably don't care if others are cutting in front of me, and the true delta in elapsed time on a drive is trivial.
Having been driven around Mumbai — the mind reels. I would never personally do it (though I saw very few actual accidents while I was there). As a critical mass of autonomous vehicles start maneuvering there, though, I would actually expect the speed and safety rate to increase.
And then my feed throws this one at me, which seems much more positive…
https://www.technologyreview.com/s/607841/a-single-autonomous-car-has-a-huge-impact-on-alleviating-traffic/
No fuzzy logic. Only discrete.
They're more like a VI than a true AI.
From something I wrote elsewhere:
In terms of real-world experience, I think it's important to note some practical numbers:
Absent any reference, three million miles feels like a huge amount of testing, but in the US alone Americans drove their various cars 282 billion miles … last June.¹
So the entire history of Waymo's self-driving testing represents ~0.001% of the US fleet equivalent for a single month, or about a three quarters of a fleet-hour's worth.
Recognize and celebrate our achievements while simultaneously recognizing where they fit in the grand scheme. There is much yet to be done.
Also, in terms of safety:
There's another statistical anomaly, too: Tesla reports 222 million autopilot miles with two fatalities (assuming we trust that the Netherlands fatality did not involve Autopilot, a fox/henhouse question since only Tesla commented or has direct access to the data) for a rate of 0.90 fatalities/100M miles, compared to the US National Highway Traffic Safety Administration (NHTSA) overall rate of 1.08 (as of 2014) [2]. Based on that, Tesla looks pretty good.
Except the NHTSA numbers include all vehicle types – trucks, motorcycles, etc.; they also include all vehicle-related fatalities – pedestrians, bicyclists, people struck by crash shrapnel, etc.
If we limit the fatalities to vehicle occupants and passenger cars, as has been the case with the Tesla crashes, the US rate drops to 0.74 (again, as of 2014). That makes Teslas look comparatively unsafe, 22% more likely to cause occupant fatalities in a crash.
The Tesla rate is probably higher since I originally wrote this; I believe they've had yet another fatal accident "with autopilot engaged" in the interim.
—–
1. (Note: PDF) http://fhwa.dot.gov – http://www.fhwa.dot.gov/policyinformation/travel_monitoring/16juntvt/16juntvt.pdf
2. http://www-fars.nhtsa.dot.gov/Main/index.aspx
+Michael Verona early days, certainly, but with over 100 Waymo vehicles driving around Phoenix now (and I believe 500 more on order) they must be clocking up some excellent experience. New York will surely add more useful data into the mix.
The comparison of total miles driven by all people, vs total miles driven by one company, seems something of a false equivalence if you can only learn from your own experience. The Waymo algorithm is learning from more than that even, since they claim they also have simulations of even more unusual events in their training, although the value of simulations seems likely to be different from the value of real world experience.
I think it's generally acknowledged that Tesla's autonomous systems are as much as a year or two behind Waymo in sophistication too.
+Dave Hill I know people who grew up in India and lived there as adults for years, but refuse to drive there now when they go back 🙂
My former wife told the story of an American pilot who lived to Iran (they retire their pilots 10 years later) and has an accident t on his first day. "They cut into my lane!" He told the other pilots. "Your lane?" The other pilots said. Did you pay for it?
+Kee Hinckley Street traffic in Mumbai was of the "how many rocks fit in a bucket?" style. Lanes were largely ignored, even as suggestions. Busses took up the most space, then brightly painted trucks, then cars, then tuk-tuks, then scooters. If there was space head of you, or to either side, to fill with a vehicle the size you were driving, you did.
I can actually imagine autonomous vehicles of the future doing much the same thing. Why not? Lanes (and lights) are to keep humans from colliding. Autonomous vehicles can negotiate with each other.
Or city streets can simply go to three lanes one way, one the other when traffic demand, then change back later.
But those things rely on only autonomous vehicles being in play. And, unlike my objections to most DRM, it will be critical that people don't have access to their car's brains to screw around with them.
And that's what I meant by the transition period being the most dangerous. Ofc, once it's all automated, then watch for disruptive hacking…
Rome was like that… any available space on the streets got filled with mopeds who drove like crazed monkeys on meth with a wreck of a guardian angel on each shoulder.