Self-Driving Cars Have a Problem: Safer Human-Driven Ones
wsj.com
wsj.com
There's a "trust but verify" element of airplane automation that is ingrained in you as a pilot. The ML aspect of self-driving cars generally make it a lot less easy to reason about and predict. That's problem 1. Problem 2 is then speed; Even in the airline world, there are published minimums at which pilots need to be absolute certain that the AP is doing the right thing before they commit to land, even in cases where the autopilot is responsible for the entire landing (Autoland is a capability that a lot of airliners have had for many decades).
What we're talking about here is several seconds heads-up with visual AND instrument confirmation, that the computers are doing what they're supposed to before we allow the plane to land. Not only that, but we bend over backwards ensuring that there is all kinds of redundancy and procedures in place to ensure the accuracy of the system: Multiple instruments using different kinds of technologies confirming one another, greater separation between aircraft to guarantee ILS radio beacon accuracy, etc. Conversely, in the self-driving world, we're having these ML algorithms make split-second decisions with no real way of informing the driver what it is basing it on in an environment that is far less predictable and constantly-changing with someone who is not a professional behind the wheel.
A better approach would be to require the human driver to actively control the vehicle at all times, but have the computer step in to brake or steer when it detects a high probability of a dangerous situation. Like how auto GCAS works on modern aircraft.
A few months ago, I lightly rear-ended a vehicle ahead of me when they stopped at a green light while turning.
Legally, I was in the wrong, and it was a stupid mistake. I had looked left to make sure no traffic was coming at the exact moment the person ahead of me stopped. Had either event (my head turning or his breaking) been different by about a second, no accident would have happened.
A simple anti-collision system would have prevented this, and saved me some money (and him from some scratches on his bumper).
As I have an older car - 2012 model - so this feature wasn't prevalent when I bought it. But my next car will have this capability, if for no other reason than I hate stupid mistakes.
This one feature would make my drive to/from work much less stressful (there are a ton of people who will switch lanes in front of you at the last minute and then hit the brakes here in Houston). Add in lane-change warnings, and I think a significant number of accidents could be avoided. They may be small accidents, but a bumper replacement or painting (times 2 - one for each car involved) is easily $1,000, if not more.
I got a car that automatically closes the door so you don't have to slam it. Now I keep forgetting to close the door on a normal car properly. It's eliminated my muscle memory in just a few months.
Another instance was when I got a car without a clutch interlock. I never would have realized how dependent I was on that until I started it up in gear.
1 - All cars add this feature either through regulation or other pressure.
2 - Most cars add this feature, and those that don't see higher accident rates, and therefore higher insurance bills. That should result in those cars being less desirable (and therefore less profitable for the car companies), OR the car companies figure it out and add the missing feature to the rest of their cars.
Either way, I think we'll end up there, just not sure of what path we'll go down.
If you own a car, you will get used to it, so it won't be more dangerous.
With self driving cars we use some fancy ML/AI where we don’t exactly know why the machine chose to do what it did.
He's also leaving out checklist-driven procedure, which is a huge component of airplane safety that is entirely absent from car driving.
Similar to how if on an instrument approach the altitude doesn't agree with the vertical guidance. You can check other systems to see which instrument has failed and whether it's the vertical guidance or altimeter.
Of course, this method of driving and having to pay more attention doesn't jive with the fantasy of getting wasted and having a computer take you home.
Which is why the system you're describing will never come to cars: American drivers would never use it correctly. If it isn't something that complete morons can jump in and use with zero training, it's just not going to work here. What you're describing requires a competent pilot with training, and that's a completely unrealistic expectation for a car in America.
Anything that moves the needle noticeably closer to "plane levels" is a solid improvement in terms of reducing unnecessary deaths.
Or a citation that Tesla's auto-pilot is safer than driving? I have no idea whether they're safer than driving. My argument was about what's a sufficient goal for self-driving cars and what's needed to reach that goal. Not about whether Tesla had reached that goal.
I have never met a pilot that didn't religiously follow their checklists. If there is ever any doubt as to whether something is safe, there is no doubt: it is not. Weight limits, weather reports, etc. all are carefully attended to.
Meanwhile, 30% of adults sitting in the back seat of a car forget to buckle their seatbelts. About 5% admitted to driving drunk within the last month. I see people on their cell phones all their time. I've ridden with a young woman who was driving with her knees while doing her hair in the rearview mirror (yes, I was freaked out, and no, I never rode with her again).
Pilots are professionals, and they act professionally, and they respect the possibility of death every time they get into their vehicle. Drivers are not. There's a long history of every safety innovation in cars actually making accident rates go up, because drivers start taking them for granted and then behaving in ways that compensate for the increased safety and maintain the same level of subjective risk (like driving faster, driving more, or becoming more distracted). The only way to break that loop is to take the human out of driver's seat entirely.
Someone forgot to follow the second page of the checklist of a 737 in Argentina in 1999 https://en.wikipedia.org/wiki/LAPA_Flight_3142
The part about the checklist is more clear in the Spanish version of the article, but it's not 100% clear, now I'm not sure if it was confirmed or it was only a rumor.
Something like they forget to read the second part of the checklist because they were distracted by someone, so they forget to open the flags. When the alarm sound, they ignore it because they have a similar alarm that always sound due to a malfunction. There were also some logistic problems, like a schedule with less rest time than required. As always, there is more than only one cause.
https://en.wikipedia.org/wiki/Delta_Air_Lines_Flight_1141
What we're talking about here is several seconds heads-up with visual AND instrument confirmation, that the computers are doing what they're supposed to before we allow the plane to land.
Let's say we never get to full autonomous driving on roads mixing autonomous cars and human drivers. In that case, is it ultimately going to be that different from the way we teach people to look both ways before they cross the street, and teach new drivers how to deal with highway on-ramps and intersections?
It would seem to be foolish to have a goal of having "no fatalities per 100 billion miles", or whatever the current spec level for airline travel is, as the minimum operating requirement for automobiles. Likewise requiring two expert-level operators at the controls, with re-certification every six months seems to be out of scope. So why should the AP try to be designed assuming those things are true?
Your proposal also lacks reliability. What if one of these "transponders" fails? We can build redundancy into a self-driving car quite easily (a second computer, never relying on just one sensor). Providing redundancy to road infrastructure planet-wide is ... a much larger problem.
sure, self-driving cars would be nice and we could even rationalize how this could make sense, but at the end of the day I would rather see everyone get 1) high-speed internet 2) clean water 3) reduce the dependency on fossil fuel 4) ensure access to quality food 5) move towards environmental friendly materials.
All the other things you mention are good, and could probably be a priority, I am not saying they shouldn't be. But if we are talking about a technical solution to the problem, here's the thing:
a. You can either have cheap self-driving cars and an upfront infrastructure cost + additional ongoing cost, or expensive self-driving cars capable of driving on shitty roads. My conjecture is that the latter won't 100% solve the problem anyways, so why bother? Let's not let perfect be the enemy of good.
b. Why do roads get lower priority for you than high-speed internet? And why do you prioritize high speed internet above reducing reliance on fossil fuels?
c. Maintaining roads is already a thing we do, it is already done via public funding, and again that makes your existing cars cheaper. You'd need a damn tank to roll through unpaved streets at 40 mph, let alone at 80 mph like everyone currently drives in their little Civics with 15 inch wheels.
Again, I don't know which thing should be a priority, not a problem I set out to discuss here. I am saying that IMO the cheapest and best way to get to self-driving cars is as I described. Let's upgrade the roads so we can all have nice things. This also lets us upgrade the rest of the roads to make them self-driving capable later. Or we could try to make a computer that's as good at driving in Montana blizzards as a human, and not see a single consumer self-driving car for the next 50 years.
my bigger point is that we’ve fallen in love with the idea of having self driving cars. do we need self driving cars? who should shoulder the cost?
the things I gave as examples are not ordered. truly high-speed internet everywhere would be great as i believe can actually enable telecommuting and eliminate traffic. it can also enable access to education and drive us, as a society forward.
Self driving cars are already dependent on standard infrastructure markings such as lane lines, and whenever those markings are confusing or faded it has lead to fatal crashes, such as the fatal Tesla crash on 401.
I predict self driving will not fully succeed until we build infrastructure to support it into the roads.
A Tesla is not a self-driving car and they won't become one with that approach.
Waymo vehicles have virtual maps that include the lines (along with a LOT of other data) so they are not affected by fading lines as they always know where the lines are supposed to be.
Basically Waymo is building the infrastructure you all want, but in a virtual sense instead of the extremely expensive and impossible physical one.
You only need one reality <-> virtual desync to cause a crash.
IgorPartola doesn’t talk about “planet-wide.
”Not only is the modification expensive, you have to convince the local authority to foot the bill, do it right, maintain it, etc. Never going to happen.”
Not in the USA, where government solutions are “socialism”, and voters rather pay for individualized solutions (example: not allocating money to maintain roads, but letting individuals bear the cost of repairs caused by potholes (https://www.wired.com/2015/07/heres-much-citys-crappy-roads-...)
Other countries may act differently.
”What if one of these "transponders" fails?”
You use passive transponders that rarely fail, and put them at every x meter, where cars work fine if they’re at every 2x or 5x meters, and maintain your roads.
I don’t see why that would have to be more expensive than the existing practice of adding, e.g. https://en.wikipedia.org/wiki/Cat%27s_eye_(road) and various lane markers to roads (cat’s eyes probably could be sufficient for relatively good lane keeping)
Even if it is, it may still be a net win if it means lanes can be made less wide, or distances between cars can go down, or even if it ‘just’ means fewer people getting killed and maimed in road accidents.
(Seriously, cones really ought to be dealt with by image recognition? Easiest possible case, an object specifically designed to be seen?)
Another press of the button dispenses a martini.
I used to do QA for the Waymo cars, and boy do they get worked into a tizzy if there's any kind of construction going on. It was the most common situation for remote assistance (read: people sitting in a room sending pathfinding suggestions) to get involved.
Judging from some of these comments, the problem is still very difficult.
Think about it this way: the roads themselves are very expensive. But they make each car much cheaper. Most cars don't have to be equipped to drive over rocks and roots (we have tarmac), or lakes/ponds/rivers (we have bridges). We already spend a ton of taxpayer money to make it so that a 15 inch wheel with a mostly smooth tire and 5" of suspension travel is all a car needs. Why not install transceivers to make cars not require to have a super complicated LIDAR system?
"Back in 1995 the goals for self-driving cars were more modest than they are today. They weren’t called autonomous, but self-driving. And there was no plan to have cars drive themselves on city streets, just on freeways and highways — on the Interstate. The plan was to bury cables in the pavement over which all the cars would drive and communicate with each other and with the road, itself. The goal was to fill the road with cars driving at the speed limit, spaced precisely one meter apart" [0]
[0] https://www.cringely.com/2016/08/25/self-driving-car-old-eno...
Given sufficient energy we can create anything because E = mc^2, but we have yet to actually produce enough energy to create whatever matter we want.
Edit: Downvoters, please explain yourselves. If you have evidence that ice and snow have been solved for level 5 I’d love to see it.
He was also right because, I believe the LIDAR units on say a waymo car probably cost more than an entire tesla vehicle.
But I couldn't help thinking if one of these LIDAR startups comes out with an "automotive pricing cheap" chip, it would be a great sensor fusion thing to add.
Basically, "update all roads on the planet".
And we're not going to get to a system that's so much safer as to render human attention irrelevant without a much more solid and theoretically sound approach to safety engineering.
https://en.wikipedia.org/wiki/Boeing_777#Accidents_and_incid...
Perhaps in the short-term, but if you listen to Elon Musk compare it to like having elevator operators, it makes sense that computers will eventually do a much better job and humans will only introduce problems.
https://www.zdnet.com/article/elon-musk-on-teslas-autopilot-...
Most likely computers will become far better than humans at some point. But it may be far enough in the future that it doesn’t really contradict the point made above. Contrary to popular belief, “640K should be enough” isn’t the kind of statement meant to cover all eternity.
If you listen to Elon Musk we have fully autonomous cars every 2 years since 2015 and we'll have Mars cities in 2050.
Landing a few people and picking up a few rocks and taking a few selfies cost, inflation-adjusted for 2019, a whopping $288.1 billion [1].
This money was spent to "stick it to the communists" because they were winning at the space race (first man in space, etc...). We just happened to get a bunch of benefits along the way.
The several more hundred billion that would have followed if we decided to build entire colonies - with a multitude of massive problems still not figured out today - would have been a hard sell over the decades it would have taken to construct.
I mean just look at operating cost of MIR and ISS and that's in LEO...
[1] http://www.planetary.org/blogs/casey-dreier/2019/reconstruct...
Because there are many different numbers:
http://www.thespacereview.com/article/1579/1
https://www.quora.com/How-much-did-the-Moon-Mission-cost
The point is that it’s a matter of cost, and not a technical problem. The Apollo program was a “rush job” done in 10 years. Much of the technology had to be invented.
There an interesting BBC podcast called 13 Minutes to the Moon for anyone interested in finding out more about the Apollo program:
When I was reading that, I was expecting a much higher number than the one I ended up reading.
That's < the cash Apple has on hand, approximately as much revenue as they make a year.
That's 1/2 of the annual U.S. military budget.
It's 1/3 of a small portion of the 2008 bailout.
===
If you annualize that by the 10 years it was spent in, the only discretionary funding significantly less than that is Food and Agriculture.
But the difficulty is in
a) allocating this money b) convincing the public that removing this money from, say education, is a good strategy
A good chunk of people seriously require heat warnings on their hot coffee to realize that their coffee is hot. I don't hold high hopes in making a voting majority of people understand long-term global benefits.
He's very good at PR moves, but sending a rocket to space and getting it back in one piece, which is very impressive, is nothing compared to actually going to mars, which itself is nothing compared to building a city on a planet where everything is opposed to human life.
Tesla and Space X have a very high "hype factor", and sure they're probably the best at what they're doing, but he is insanely overselling it. Look at that timeline: https://www.inverse.com/article/51291-spacex-here-s-the-time...
No. Stating something as a fact without analyzing it in the slightest is bad.
Example:
(1) Only a computer is "paying attention" performing the "assisted emergency breaking" role. The computer has an effective reaction time of 1/100 of a second. Within 2/100 of a second of a qualifying event, the vehicle will be in full antilock brake mode.
(2) A human and a computer are "paying attention" performing the "assisted emergency breaking" role (this is a little silly in that the human is already the primary brake operator...). The computer still has 1/100 s reaction time, but the human's reaction time is 25/100 s. What exactly is the human bringing to the table? It is true that the human may detect an actual threat the computer doesn't...however that slow reaction time means it likely won't make a difference in the outcome. On the other hand, the human may "detect" a false alarm, and emergency brake by mistake. That activity often leads to wrecks!
I submit that the computer-only emergency brake assist is safer than the combination of human and computer. Further, all of these "human in the loop" systems suffer from a fatal flaw - the inability of people to pay attention unless it's absolutely crucial.
The way people use Tesla Autopilot shows that full autonomy is necessary, and will in fact be safer than any "assisted" system.
The two can, and should, be complementary.
The two can, and should, be complementary."
What you're saying is true, but doesn't fall under the purview of "assisted emergency braking".
You know, not driving tired, drunk, high, angry, or stupid...
> No. Stating something as a fact without analyzing it in the slightest is bad.
> Example:
> (1) Only a computer is "paying attention" performing the "assisted emergency breaking" role. The computer has an effective reaction time of 1/100 of a second. Within 2/100 of a second of a qualifying event, the vehicle will be in full antilock brake mode.
> (2) A human and a computer are "paying attention" performing the "assisted emergency breaking" role (this is a little silly in that the human is already the primary brake operator...). The computer still has 1/100 s reaction time, but the human's reaction time is 25/100 s. What exactly is the human bringing to the table? It is true that the human may detect an actual threat the computer doesn't...however that slow reaction time means it likely won't make a difference in the outcome. On the other hand, the human may "detect" a false alarm, and emergency brake by mistake. That activity often leads to wrecks!
> I submit that the computer-only emergency brake assist is safer than the combination of human and computer. Further, all of these "human in the loop" systems suffer from a fatal flaw - the inability of people to pay attention unless it's absolutely crucial.
> The way people use Tesla Autopilot shows that full autonomy is necessary, and will in fact be safer than any "assisted" system.
I think your example is uncharitable because you allow for a human to react to a false alarm in the second but don't account for a computer to react similarly in the first. Personally speaking, I have seen very few occasions of people exercising extreme braking wrongly. I feel computers would be less apt at discerning from a false positive than a human operator with even a little experience behind the wheel.
Given that autonomous cars can be misled by painting roadways certain ways [0], it's clear to me that the number of unknown unknowns that must be handled by a car are large in number, and I have less faith in the skillful handling of those events by a computer than a human.
He can see the big truck 200 yards in front of him, and recognize that it is indeed a truck, and not an overhead traffic sign. Apparently, Teslas don't bring that to the table.
Alternatively, the human could recognize that the truck is on fire and smoking, and that it would be a good idea to stay away from it.
You're choosing to ignore the fact that computers still make mistakes.
My friend's car (an Acura) recently slammed on the brakes @ 70mph because it mistakenly thought a car in an adjacent lane was in his lane. Had there been someone following close behind him, it probably would've caused an accident.
This is not an uncommon occurrence, especially with Honda/Acura's system. You can find tons of complaints online about AEB systems reacting to false positives.
No, but I'll unequivocally assert that the human false alarm rate is much higher than that of computers.
"My friend's car (an Acura) recently slammed on the brakes @ 70mph because it mistakenly thought a car in an adjacent lane was in his lane. Had there been someone following close behind him, it probably would've caused an accident."
Does it have a readout indicating what it "thought"?
It seems possible that the car in the other lane drifted, and the computer thought it was changing lanes into the Acura. Braking was appropriate in that circumstance.
"This is not an uncommon occurrence, especially with Honda/Acura's system. You can find tons of complaints online about AEB systems reacting to false positives."
That may be, and I suspect a fix will be forthcoming which will address every Honda/Acura on the road (or new models, worst case).
Meanwhile, humans will continue to look down at a text, hamburger, or whatever, and then panic when a few seconds later they look up and think they're in trouble. Not to mention driving tired, drunk, high, angry, or stupid.
"Had there been someone following close behind him, it probably would've caused an accident."
And of course the person running into the braking vehicle would have been at fault. In most cases, only assisted emergency braking would help, since most people follow way too close on the freeway - human reaction time isn't fast enough.
.. to objects that it can identify as a hazard. We've already seen incidents where it fails to spot hazards such as cyclists and crash barriers, due to limitations such as having to filter out stationary objects from radar returns.
Really? I've seen no such reports with regard to assisted emergency braking systems. Please cite a reference.
https://news.ycombinator.com/item?id=19802284
https://news.ycombinator.com/item?id=17257239 "These vehicles will run into stationary obstacles at full speed with no warning or emergency braking at all."
Anyway, the point was a general one - the computer hazard identification is very far from perfect at the moment, and benefits from a human spotting hazards that the automated system does not.
https://en.wikipedia.org/wiki/Emergency_brake_assist
I see no mention of problems on the Wikipedia page. Do you know of any?
Anyways, my VW has assisted emergency braking. I've had opposite issues - water splashes across sensor and the car brakes for no good reason. Generally going through a large shallow puddle and moderate speed. Happened 2 or 3 times and now I disable when driving near the beach after rain.
Your "issue" is essentially the system being extra safe - similar to slowing down when a bunch of water hits your windshield. Seems highly appropriate to me.
You've done nothing to support that assertion. "MUCH better at assessing the situation and acting appropriately" doesn't compensate for 10x slower reaction time. Nor does it account for the high percentage of time that humans are paying attention to something besides driving. Machine systems don't get distracted, at least.
We'll agree to disagree, but anticipate fully autonomous cars sooner rather than later. ;-)
So, the Tesla issues involved not identifying objects in the field of regard correctly, and thus not steering appropriately. Brake systems weren't involved in either crash, IIRC.
Assisted emergency braking is much more narrowly focused than full autonomous driving. It doesn't care what the object is, if it's solid, in front of the vehicle, and in it's path, the system will brake. These days it potentially has vision, sonar, lidar and radar to provide sensory input. That's a whole lot more information than a human can access.
Switching on the AutoPilot also makes me suddenly care a lot less about optimizing for that 1/10th of a second here and there. When the car’s driving, I get there when I get there. I can just relax and the traffic matters that much less.
My favorite expansion of the overloaded motorists' workload is when "firemen" holding "boots" get in traffic lanes during the morning commute in order to gather donations. Is that in the model?
in any case, i'd guess current algorithms can detect that fairly reliably in the noted conditions, since self-driving cars also trained on urban streets.
But sometimes it would be nice to have a 0 below the 1-7 following distance setting. (sort of like going to 11)
Is this a joke? The roads aren't centimeter-accurate from day to day.
"Decades away"? LOL
Fully autonomous vehicles are already on the road in large numbers. See Waymo in AZ and Voyage in Florida...
The economic and safety incentives are huge. I predict by 2030 over half of new vehicles sold will have L5 autonomy.
Uber messed up big time in multiple ways, but legally it's pretty clear that the pedestrian was at fault in this case, not Uber.
the same thing can not work with machines cause they are bad at detecting when humans are not all right and humans are bad at trusting machines.
as soon as we just accept that the save way to operate in close proximity of others, is to do it at way slower speeds. we will have very save roads with humans and robots behind the wheel.
The correct conclusion is that the safe speed directly depends on the present conditions, and depending on that, you can be going unsafely fast or unsafely slow (or both, if you’re drunk or inattentive).
https://www.google.com/amp/s/www.thelocal.de/20190201/are-ge...
The desire to minimize the speed at which crashes happen literally costs lives when applied in the real world because reducing the frequency of crashes is the superior option.
While I understand that minimizing crashes is important, I'd rather be faced with scenarios where we crash in a manner where people survive more often - any accidents that do happen at the higher speed is going to skyrocket the chances of a fatality.
It's too bad the civil engineers working as traffic engineers absolutely refuse to actually follow the recommendations of these studies. Honestly, I can't think of a more intellectually dishonest profession than civil engineering (specifically, the traffic engineering subset of it; the guys who build buildings and bridges are OK).
Do you remember where you saw the data on this?
Here's a good explanation: http://www.mikeontraffic.com/85th-percentile-speed-explained...
Basically, you set the speed limit at the speed most of those drivers are already naturally driving at, so that the people who do mind the speed limit will travel at the same speed of travel as everyone else.
And then, of course, obviously there will be outliers.
Number of accidents drop, but the number of fatalities increase when you raise the speed limit because the accidents that do happen have so much more energy in them (e=mv^2). So "safety" only improves if you have a bad definition of safety.
This situation is only a problem for those manufacturers who want to pass off partial autonomy as the real thing.
1. Lack of higher reasoning is a disadvantage that self-driving cars won't escape any time soon.
2. But SDCs can more than make up for that by making fewer dumb mistakes, giving them a better overall safety.
3. However, if you use machines to take away the dumb mistakes from humans, you change that equation, and the overall stats could go the other direction.
In reality, I don't think it's that simple. When ABS was newer, they studied it (https://crashstats.nhtsa.dot.gov/Api/Public/ViewPublication/...) and found that ABS improved overall safety, but safety paradoxically got worse in slick conditions (rain, snow, ice). The effects of safety improvements aren't always what you'd expect, so theorizing doesn't tell you much.
Nobody has actually created a level 4 system yet, not even a prototype, let alone one ready for production. So level 4, too, is still science fiction. The same goes for level 3, actually. It's science fiction. And so are claims like the following:
>> Researchers at Cleveland State University estimate that only 10 to 30 percent of all vehicles will be fully self driving by 2030.
2030 is in ten years from now. In ten years from now, we'll reach "full self-driving"? Waymo was founded ten years ago and its cars are still in level 2 (allegedly, trying to "jump over" level 3 and go straight to level 4). How are we going to be suddendly, magickally transported from level 2 to level 5 in the next ten years, when we haven't budged from level 2 in the last ten?
1. Addressing (and reducing) major causes of road accidents and fatalities such as impaired and/or distracted driving (alcohol, drugs, sleep deprivation, inattentiveness, fiddling with phones, radios, touch screens, etc.) would make a tremendous difference, and automated driving systems and sensors could potentially intervene in such situations without necessarily being online all the time.
2. Inattentiveness and distraction caused by (over)dependence on automated systems (as in jets with autopilot) could become a major hazard. This is already a problem with road designs that minimise driver engagement by removing obstacles to traffic and designing roadways entirely for cars (compare European style traffic calming to US style widening of lanes and rounding of corners).
3. As mentioned above, infrastructure and urban planning practices are major issues, both in terms of maintenance and design. Automating passenger cars may improve safety in some respects, but perpetuating the dominance of individual motor vehicles (and the vast infrastructure outlay they require) as the dominant mode of transport is probably the wrong approach. Gradually transforming urban planning and design to promote mass transit, reduce commuting, and accommodate bicycle and pedestrian traffic would generally reduce the need to use cars in the first place. One difference between traffic death statistics in countries like Germany and the US is that virtually everybody has to drive long distances on a daily basis in much of the US, with few if any alternatives, and urban sprawl encourages lengthy detours to travel trivial distances between fully separated residential/commercial/industrial zones, whereas e.g. in urban Germany, there are many alternatives, and driving is not a necessity (compare death rates per capita vs. per km driven). In Germany and much of Europe, driving is and has always been a privilege, and one that requires a degree of skill to earn, whereas in the US driving is viewed more or less as a right, something everyone does and has to do on a daily basis, i.e. it's easy to get a license and you only lose it for serious infractions, and even then penalties for driving without a license (as many people do!) are comparatively minor.
What I am getting at is that the best way to tackle road safety and environmental problems is to gradually abandon the passenger car as the dominant mode of transportation. Hybrid automated/human-operated vehicles could be a great improvement in the short term, but using bikes would be better in the long term. Freight transport might be another matter though.
https://crashstats.nhtsa.dot.gov/Api/Public/ViewPublication/...
My personal takeaways:
1) The sum of speeding + alcohol is 55% of the total number of fatalities. Self driving cars won't get drunk (we hope) and speeding is something which could certainly be limited in software. Lots of reasons to believe self driving will make a significant impact here, whenever it finally arrives.
On the other hand, you don't really need self driving to prevent these. You could force all vehicles to have a breathalizer interlock before driving (assuming these devices could be made to be very accurate and not spoofable, I believe these are both technically feasible goals). Or just enforce the living daylights out of it - driving drunk, lose your license for the rest of your life. Not nice I suppose but neither are ten thousand deaths per year. Speeding is a similar story. Why not put a GPS device in every car (I think this may already be the case) and geo-fence speed limits? Technically very feasible. Some people will claim you need to speed sometimes but I believe these arguments to be total bullshit. Ambulances may need to speed sometimes. Or again just enforce the living daylights out of it. Speeding over 5MPH above the speed limit, license gone for the rest of your life. Not nice I suppose but neither are ten thousand deaths per year.
We probably won't do any of the non-self driving things I mentioned above because they are politically untenable in America. Death on the roads is such a normalized facet of modern life that most people don't really consider the alternative: limited loss of freedom with many fewer deaths and injuries.
2) The wearing of seat-belts seems to still be a big problem for some people. 44% of motor vehicle occupants who died were not wearing one. No self driving needed here. In fact, this problem is somewhat orthogonal to self-driving. You would still want to be restrained in a self driving car. It would be easy enough to create an occupant sensor and seat belt interlock to operate the car. The car already beeps at you if you're not wearing it. Why not take the beeping a step further.
3) Running into cyclists (both motorized and human-powered) is also a reasonably large problem. Self driving will almost certainly help here (uber's incident not withstanding) but there are non technological solutions to consider. What about increasing the liability for drivers who injure or kill cyclists? What about steep criminal penalties for hitting a cyclist?
4) There isn't even a category for "mechanical failure". Cars don't really crash due to problems with the car per-se. Or at least not a rates that matter compared to the others.
5) Rural driving seems pretty dangerous - 50% of fatalities but only 19% of the population lives there. I don't think they dig into the details here but I'd be interested to see the cause breakdown for rural users. Growing up in Wisconsin, driving home drunk from rural bars was basically ubiquitous. I always advise to stay off the roads around bar time if possible.
It is important to consider that our roads are unsafe mostly because of the system of policies we have set up governing them. We still aren't serious enough about drunk driving. The debate around safety and speeding is hardly even happening. Hitting and killing non-automobile road users is often just a traffic citation. The licensing system is a joke. In what other safety-critical certification system can you take one exam at age 16 and then renew your certification for the rest of your life without any kind of additional training or examination? There is a ton we _could_ be doing outside of self-driving technology to make the roads safer.
Personal preferences on the above discussion: FOR self driving cars, even ones that are only marginally safer than humans. FOR geo-fenced speed limits. FOR breathlizers in all vehicles. FOR steep penalties for drunk driving, speeding, and killing non-automobile road users. FOR much more stringent licensing requirements. Death on the roads ought not be a feature of modern life.
Speed is also an interesting topic in that speed limits are often kept artificially low in order to generate revenue. Any automated system to force cars to respect limits would likely meet resistance from municipalities that are dependent on traffic fines for revenue.
High speeds are allowed on highways in Germany, but only in certain areas, and German drivers are extremely disciplined with regards to lane changing and the 'hierarchy' of the road. However, traffic deaths are lower than the US, even when adjusted to reflect deaths per km traveled. Driver discipline and skill undoubtedly play a role here, but infrastructure quality and maintenance (and urban planning!) is also a big issue. US roads are comparatively poorly designed and maintained, and outdated engineering and planning practices are still widespread (widening streets and removing barriers instead of using traffic calming--perhaps counter-intuitively, the former causes more serious accidents by encouraging speeding and disengaged driving).
I tend to think that at least in urban areas of the US, ditching zoning and onerous anti-alcohol laws (i.e. putting pubs within easy walking distance of where people live) would drastically cut down on drunk driving. Rural areas are of course different.
Just my two cents
At least in my country we have rules that going 2x speed limit is loss of driving license.
There's also interesting traffic light system on entrances to cities, that if someone goes over speed limit - light is turning red for a short while. This is very nice, as lights are more respected than speed limits.
What I think gets overlooked in this conversation is, human drivers routinely break the law in "innocent" ways. If self-driving cars follow the rules in ways that human drivers fail to, the self-driving cars are already safer in numerous ways.
It is simple: highway speeds + opposing traffic. e.g., 55mph -> <- 55mph = bad accident