Hacker News Reader

Comment by cpprototypes | Hacker News Reader

There is a common assumption that humans = bad drivers. And that these self driving cars will be much better than people at driving. I think this belief greatly underestimates the difficulty of what Google and others are trying to do.

Humans are actually great drivers when the situation requires thinking. Lot of snow and can't see the lane markers? Millions of people adapt every day to this during winter. Lots of pedestrians, bicycles, motorcycles, etc. doing somewhat unpredictable things? Again, look at any Asian mega city, people can adapt very fast. Basically, if the situation requires being alert, people are very good at driving.

When are people bad a driving? Whenever it's monotonous. Bumper to bumper traffic or a free flowing freeway. Constant repetition of red/green cycles while going down a suburban street. These boring situations make a lot of people basically turn off their alert thinking. Then they do other things like texting, talking on phone, etc.

And these boring situations are exactly where AI self driving is better at driving than people. Computers never get bored. The self driving car will be at 100% attention even during the most boring traffic. But when boring suddenly turns to not boring? The current state of AI is very very bad at this.

And unfortunately there's no easy way to use the best of both sides. If the AI is fully driving, then it will do great while the driving is boring. But by the time the AI decides, this is too much, can't handle this, it's too late to alert the human driver to take over. But if the human driver is required to always pay attention, what's the point of the self driving car?

Self driving cars will get there someday, but I think it's much farther away than many assume it will be.

View on HN

"Lot of snow and can't see the lane markers? Millions of people adapt every day to this during winter."

Having driven a corridor with a ton of snow quite a lot (western PA to washington DC), i'm going to strongly disagree with this.

People don't "adapt". They just do it anyway because they have to. The rate of minor and major fender benders in this kind of situation is really high, it's just "acceptable".

In truth, computers that understand the physics of what is happening to the car at a given point in time based on tons of sensor data are going to be much more effective drivers in snow and ice than people (because of even simple data it can use like "this wheel has this much traction currently, so i need to correct this much"). People are really bad at figuring this out in their head and use horrible heuristics, often hitting other cars, guard rails, you name it.

You could also can get data about surrounding conditions (IE the ice the wheel will hit next or whatever) that people can't because they are not close enough/can't process fast enough (IE not just what is going on with the car, but what is the state of the ice on the road 2 inches to the right or left of the car and would it be better for the car to steer in that direction to get more traction, etc)

This is it. There is nothing inherent to CV that makes self driving cars unable to operate in snow. Humans already suck massively at it (and in the rain).

I remember driving home in a massive rainstorm at night about two years ago with around 2 meter visibility and was going like 20mph because I could not see shit. I had less sensory data than a computer with extra vectors of vision beyond sight would have had, and my reaction time is a ton slower.

Then a year later I'm driving in icy conditions and count five cars that either off roaded or rear ended each other in the span of 15 minutes because of the slick surfaces. I imagine self driving cars would do a much better job slowing down and budgeting for potential slippage than impatient humans.

There is nothing really magical about our eyes that give us better clarity in awful driving conditions, and we don't have radar and GPS fed into our brains to give us a clue where the roads are supposed to be - we just rely on historical context and past driving experience on the same roads.

How many hours would it take a networked fleet of self driving vehicles to outdo my historical context for driving on any of the roads I've driven on for a decade? Two? Maybe three? You just need to get the tunable parameters right though some good old genetic algorithm work like what Google is doing now with its test fleet and while it would be impossible to guarantee perfect performance in every scenario (because inevitably an earthquake is going to make a self driving car crash where nobody could have prevented it) it just has to be better than us, and that has almost certainly already happened.

The biggest thing is that people drive when they shouldn't. Sometimes your car just isn't going to do it. All season tires, 2WD, ice storm. Yeah that's when you see people sliding down hills. There's no traction and no automation is going to fix that. The best thing it will do, no joke, is just not start the car in those conditions.

You do see a spike of people who don't do well to adjust to winter driving. They don't give longer gaps, make sure they have more time before crossing intersections than they do. Computers will do that much better, eventually. I can't wait to go to work or drive around the city and not have to worry about someone texting, doing makeup or making poor driving decisions. Have a nap on long trips, watch a movie, finish up some work.

I can see computers having an amazing reaction time to some kid running into the street to catch a ball. An animal running in front of you at night. It's going to take awhile though. I wouldn't trust it right now to safely take me anywhere. There are just so many variables that the engineers have to account for. Weather conditions, people conditions, car conditions and on and on. It's nuts.

> The biggest thing is that people drive when they shouldn't.

It is extremely common to drive beyond your actual "in complete control" speed.

You are supposed to drive slowly enough in snow so that you can correct for unexpected patches of black ice or rocks in the snow slush. You are supposed to leave a long enough gap to be able to react and stop if the car in front of you unexpectedly fully slams on their brakes. You are supposed to drive slowly enough to stop if a kid runs out between parked cars.

Humans almost never do, because unexpected things almost never happen. Traffic flows faster and perhaps smoother, because we willingly accept driving at speeds where we can avoid most, but not all hazards.

But if we expect self-driving cars to always put safety first, then we will find them to be "tediously" slow, even if they have better sensors and faster reaction speed than us.

I drive at a "correct" following distance. Way back when I took my driving test I was told to allow 1 car length for every 10 MPH speed. So on the highway I leave at least 6 car lengths.

I constantly need to tap the brake and back away as a car cuts into the lane. Almost no one on the highway leaves adequate space for human reaction times.

I would say that 1 car length / 10 mph is actually a very short following distance. If an average car is 15 feet in length, at 60 mph, that is only 90 feet. At 60 mph, 90 feet of following distance is only 1.02 seconds. (1 mph == 1.4667 fps)

When I was taught to drive, we were told that 2 seconds of following distance was appropriate.

This just reinforces your point that the typical human driver leaves insufficient following distance.

Some of the roads in the UK have chevrons painted on them at intervals. The idea is that at a normal speed for that road, if you can see two chevrons between you and the car in front, you're leaving the right distance.

It's a good idea; the distance is always bigger than you think.

Not bad, but you need an element from the square of velocity (for kinetic energy/braking distance) there to be safe. 1 second is fairly okay at low speeds, but at 60 mph (~100 km/h, 27 m/s) two seconds is shortish and if you are doing 200 km/h (as on a German Autobahn) then 2 seconds (100 m) is not safe in my opinion.

(Although the penalty line in Germany is just this, half the speed in km/h to get meters, i.e. driving closer than 100 meters if going 200 km/h gives you a ticket for tailgating)

Completely agree with you. Should have stated that i was taught in the US, where it would be ... slightly atypical to be going 200km/h ;) At the time that I took drivers education classes, the US national speed limit was still 55mph, about 88km/h. At that speed, 2 seconds is probably a good rule of thumb for something that's pretty easier for a driver to observe.

True of course. At low speeds, the braking distance is negligible, and the margin is needed for driver reaction. When exceeding 60-80 km/h, the amount of kinetic energy and correspondingly the braking distance will start to dominate what is necessary for safety margin.

When I'm not in heavy traffic I give even more space. But unfortunately at a certain density of traffic the aggression caused by a safe distance actually makes it unsafe as people cut you off and try to "retaliate" for the imagined insult.

I don't mind stringent enforcement of life saving measures on a place as dangerous as the road, even if it means everybody moving at 10kph in bad conditions. Roads are the biggest killer of youth globally[1], minor emotional factors such as 'tediousness' should not be allowed to get in the way of solving such a big problem

[1] http://www.theguardian.com/global-development/2012/may/02/tr...

I don't think we'll care quite so much that a 20 minute drive is now 25 minutes when we have wifi and sitting in the car is basically like sitting in our living room. Driving slow is only tedious when you're doing it, not when you're able to get work done or be entertained throughout.

Yeah the norms will be completely different. If an adult kept asking a bus driver or airline stewardess "are we going to be there soon?" while fidgeting and straining to look out the front windows, most people would consider that adult to be poorly adjusted to modern living. It will be the same for robocars.

This incidentally is also (in my opinion) the hardest skill in general aviation. Knowing when not to fly a small plane because the weather is unsuitable seems to be very difficult for people to master.

The difference is, in aviation, when you get this wrong you will very quickly no longer be a pilot. Or alive.

> I remember driving home in a massive rainstorm at night about two years ago with around 2 meter visibility and was going like 20mph because I could not see shit.

I'm a relatively new driver (1.5 years or so since taking my license) but this happened to me twice, i.e. catching a rainstorm so big that I could barely see in front of me. In both cases I decided to stop in a refuge on the right-side of the road, waiting for the weather to improve. Sure enough, after about 20 minutes of waiting the rain eased down. What I'm saying is that under some circumstances (like close to no visibility) one should just stop and wait for the external factors to improve.

I'd just like to say thank-you for stopping.

I have seen drivers barreling down a motorway at 80mph in pouring rain with visibility of under fifty metres. It made my skin crawl.

It's good that you stopped, but it would be even better to take the nearest exit and wait it out there, because if you are stopped on the side of the road in poor visibility conditions, someone might not see you and before it's too late might hit you if they aren't exactly in the lane, which can happen during poor visibility.

Completely agree. Though there was one situation where I found everyone had stopped on the freeway and I kept going to an exit. Had I known why they had stopped, I probably would have stopped and gotten out of my car, too. Driving through a massive thunderstorm beside dozens of semi trucks, and all of them suddenly stop all at the same time. I keep going and get to a gas station because being trapped between all these completely stopped trucks made me very nervous. When I got to the gas station, I heard the attendant say that Amtrak had stopped too because of all the tornados blowing through the immediate area.

Yikes.

And this is the classic theory vs pragmatism problem in a nutshell. You've done a good job of explaining why humans SHOULD be horrifically bad drivers in theory.

There is a slight problem. In practice, in actual measured statistics, we're utterly fantastic drivers despite all the reasoning.

We need a baseline. How about trains? Trains in the PTC and semi-PTC era don't (can't) speed and are monitored to crazy 1984 levels, usually there are two employees in the cab (not counting subway type things, I mean real Amtrak here). Trains don't have steering wheels, usually can't spin out, etc. Still the death rate is about 1/2 per billion passenger-miles. What can I say, brakes fail, drivers have heart attacks, rails crack. Its apparently impossible to put squishy humans in a sealed box and toss that box around at 90 MPH without killing about one per two billion miles. Note that's to the moon and back 4000 times, safer than even spacecraft.

The death rate for passenger cars is about 7 per billion passenger miles. A little higher for trucks/SUVs, a little lower for cars (cars are safer on the road, more stable).

That's only about 14 times worse than trains.

Anecdotally human drivers do dumb things, because there's 7 billion people and world wide networks every form of idiocy is known to all, yet simultaneously its very rare. Anecdotally a billion passenger miles isn't very far, so every year we stack bodies like cordwood, sadly. But even on the ultimate in safety, a train, we'd still have 10th the number of dead. By reasoning the ratio should be immensely larger than 14, like 1000 or 1000000. But its 14. That has to be answered...

Even crazier to consider is the death rate has dropped by a factor of 3 since the 1970s due to better engineering and regulation. We KNOW how to lower death rates using existing techniques. Yet we also know software written by humans is basically worthless. So given a choice of strategies to further lower death rates, given the track records I'd trust the MechEng and CivEng a lot more than the CS department.

A factor of 14 reduction in death rates is unfortunately far more likely to come from the MechEng/CivEng grads than the CS grads, and I say that as a CS grad.

I'm with you up until the software being inherently unsafe bit. I've seen a lot of code I wouldn't trust with anything, let alone my life, but in your own examples: trains are partly safe due to automation, warning systems, etc. and cars have gotten safer at least partly due to things like ABS and traction control which rely heavily on algorithms. However, those are relatively simple systems compared to a self driving car and a lot more "bugs" will likely show up as more AI drives more miles.

> the death rate

While I agree about good measurements and need for baselines, death rate isn't it. When cars collide, you get really high number of injuries and deaths. For example there are almost 20x more injuries to pedestrians than deaths. It's nor fair to exclude them from those stats, since that can be a permanent and life changing injury.

We probably don't get any more people decapitated with the steering column. But it doesn't mean they are completely fine after the same collision.

Thanks a lot for making me aware of transportation safety statistics, it's a fascinating topic. And I really was surprised to see that death rates for passenger cars are only about an order of magnitude worse than train's.

I just wanted to note, that planes have a several times lower death rate than trains: 0.07 per billion passenger miles according to the sources I've seen. Maybe this could be a hint that safety in trains could still be improved?

No, that's just because flights are longer distance than train trips on average. It doesn't really make sense to compare transatlantic flights with a commuter train...

Ground failure can knock out a train (say the collision system breaks or the politicians are bought off to not require modern PTC). Ground failure can't knock out a plane but it could theoretically knock out a car so I think its a good comparison.

Also when planes encounter bad weather or ... events of any sort (erupting volcanoes?) planes working with ATC can use three dimensions to avoid the hazard. In that way trains have it worse because they can't just tear off across farmland to avoid a tornado and they only have 1-D mobility (speed). I acknowledge its unfair for cars because of the denser road network they have 2-D mobility.

Also the security and 1984 style monitoring and monday morning quarterbacking is brought to a fine art in the aviation community, far beyond the train community.

On one hand the airplane industry is rapidly moving maintenance offshore unregulated uncontrolled facilities, which will likely impact accident rates in the future, although not so bad right now. On the other hand its hard to maintain a train or its tracks offshore, which leads to inherently higher quality, but rushed jobs due to the expense.

Its a mixed bag.

Looks to me like trains are slightly safer than that 0.43 to 7.3 for cars: http://faculty.wcas.northwestern.edu/~ipsavage/436-manuscrip...

The problem is that drivers don't know when to not drive. I've driven a relatively large number of kilometers over the years as a non-professional driver and it always amazes me how many people will continue to drive in conditions that are nowhere near safe. The best way to avoid accidents is knowing when not to drive. Nothing you have to do is so urgent that you're willing to risk a detour through the hospital (or the morgue).

"because inevitably an earthquake is going to make a self driving car crash where nobody could have prevented it"

Maybe not. On March 11, 2011, a magnitude 8.9 earthquake hit Japan. The first seismometer to detect it (out of 92 in the system) was on Kinkazan Island, just off the eastern coast of Japan. Each seismometer has some compute power and is constantly computing a hazard level based on the waveforms of earth movements. The Kinkazan unit did what it was supposed to, and sent an emergency shutdown signal to the Shinkansen high speed rail system. Not through the signaling system; the circuit breakers at power substations were immediately tripped and power to the trains cut. The trains treat a power cut as an emergency stop situation and apply the brakes hard. All trains stopped safely.

[1] http://www.railway-technology.com/features/feature122751/

Also, Google never said the cars were incapable of driving on snow. Just that they weren't prioritizing tests and development with that in mind yet. They want to nail typical San Francisco/Mountain View driving before branching out into other road condition types.

"Haven't yet attempted" isn't a "cannot."

True, the number of fender benders on the first snow of the season is very high because people haven't adopted the winter braking distance model in their head yet.

In Toronto I've seen many drivers on the highway splitting lanes because they have no idea where the lanes actually are.

To be fair, a lot of those concerns are handled pretty well by traction control systems, etc. I drive in the snow a lot and really appreciate the things my car does to micro-manage traction.

The main rule for driving in snow is to assume everyone else is actively trying to kill you. :)

The main rule for all driving is to assume everyone else is actively trying to kill you.

It's something I was told when getting a motorcycle at 17, and it's served me pretty well.

Specifically on snow I think we subcontract too much from brain to ABS, traction control, lovely powerful heater and what not. We don't get Scandinavian amounts of snow in the UK, so whenever we do have snow you'll see dozens of ridiculous accidents because people don't take account of stopping distances or speed, or ABS being a hinderance in snow etc.

  The main rule for all driving is to assume 
  everyone else is actively trying to kill you.

If you were really doing that you'd never drive down an undivided road when a car was coming in the opposite direction, in case they swerved across the divider at the last second :)

Many years ago I was almost taken out when a driver did a very unexpected U turn into the gap he thought was there. Trouble is, the gap in traffic was actually me on large motorbike. So it can happen :p

Or you'd do it cautiously. As a motorcyclist/bicyclist, I have to watch for this because drivers literally do it all the time, they call it "turning left."

My motorcycle instructor advised me to "treat all car drivers as a lower life form, and treat taxi drivers as the lowest of the low." That mindset saved me a few times. U turns without warning, pulling over in prohibited areas without indicating, opening doors right in front of you - they probably aren't actively trying to kill you, but certainly can give that impression.

When we get ice in the DC area, it's always the 4WD cars that are in ridiculous accidents as they are the only ones that get get up enough speed, but pretty much all cars these days have 4 wheel ABS, so stopping ability on ice is equal between those and the cars that can barely start moving.

4WD can be a hazard solely because of the hubris.

"I don't need to go more slowly, I have 4WD!"

Cue Slip 'n' Slide. It's just enough extra performance to make people think that the road conditions don't matter.

Some of this stuff about self driving cars sounds to me like the "sufficiently smart compiler" that can can optimize away the overhead of certain features of Lisp. Meanwhile, Uber thinks my house is in the middle of I-83.

I know that corridor - and I agree with you. Unfortunately, there's also a huge number of people that don't treat snow with the respect it deserves. If you live in WV or OH (or in PA near enough to the border to fake it) you still don't need to have your car inspected. In PA, you can probably still find a mechanic that will slap a sticker on anything, but it's getting harder.

In any case, you need good tires and working brakes to navigate snow effectively. Anti-lock brakes (in my opinion) are no better than a person that can properly pulse their brakes but ... if you've got even a slight warp in your rotor, you won't brake evenly enough to stop in the snow. The other problem in PA is PennDOT's "new regime". There's been far more ice packed onto the roads when it snows lately - because melting snow that's packed under tires turns to ice. I'd far rather drive on dry, crunchy snow (for traction)

/rant

Driving in Latvia, -25 celcius, 1.5 meters snowfall, main road (closest they have to freeway/motorway not cleared ... it's 2 narrowish lanes each way with no center divider).... visibility extremely poor due to continuing snow fall.... Everyone continues driving like it's +25 celcius and sunny. This means 110km/h (90km/h speed limit in the dry 70km/h in the snow) bumper to bumper (or is that fender to fender?). Never been so petrified driving in my life. Somehow it all worked out but I can imagine one false move and we would have been looking at a 100 car plus pile up.

That can be scary but there are scarier places in traffic.

Latvia does have a traffic-related death rate (per number of motor vehicles) that is twice as big (24.8) as neighbouring Estonia (11.8), or the United States (12.9), which are about on the same level, or over five times as high as Finland (4.4) somewhat further to the north.

On the other hand it's less than half of that in Russia (53.4), and one-fifth of that in India (130.1). In Ethiopia, the death rate is 200 times as big (4984).

https://en.wikipedia.org/wiki/List_of_countries_by_traffic-r...

We can also make digital lane markers, doing whatever the case may be. If we have future infrastructure built with autonomous cars in mind, they can easily be better in those situations too.

The hard part of this we're trying to force self driving into our current infrastructure and traffic. If the bus was self driving, they would have communicated properly and merged perfectly.

And if the Moon was made of cheese, why are we not harvesting it already? Even if all new vehicles starting now were autonomous and perfect, it would take decades to get all-autonomous traffic.

> Lot of snow and can't see the lane markers? Millions of people adapt every day to this during winter.

This could not be less true among people driving in unfamiliar conditions. (http://legacy.wusa9.com/story/news/local/virginia/2016/01/20...)

The difference is that all self-driving cars benefit from the knowledge created by bumping into the edge cases. People can only do this in an ad-hoc manner (i.e. personal experience, taught by mentor, etc)

I think the poster assumed that the person was trained to fit their surroundings. We are all hyper-fitted to what we know best and what we encounter every day, and will probably fail spectacularly if we're unaware of that.

It doesn't even have to be in a place that doesn't get much snow.

First snow fall in Calgary looks like what you get in a standard snow storm in a city in the south east. Its even funnier if the first snow fall is early and then there are another two months of sun and no snow and then you get a repeat during the second snow fall. People have to relearn how to drive in snow every year.

People are bad drivers in good conditions, they just happen to be able to be bad drivers in bad conditions as well.

Bah, it's not just the first snow storm. They salt the roads so much in Ontario that people think they know how to drive in bad conditions but actually they just know how to drive in bad conditions on heavily salted roads. If they hit a patch of unsalted road they get in trouble quite quickly.

In my experience Saskatoon fares a lot better. They don't salt as heavily because it's often cold enough that salt doesn't work as well, and because judges don't blame the municipality for not salting when there's an accident.

I have a $100 bet going with a coworker at the office.

We defined self-driving cars: When I can pull out my phone in any of at least 5 major metropolitan areas around the world, order a driverless car, have it pick me up and deliver me to a specified location within the city in a timely fashion, with as little risk as getting in a taxi.

He is sub 20 years. I'm thinking more like 40-50 mark.

My thesis? We can solve the technical challenges, but so much physical, legal, and regulatory infrastructure needs to change to make this viable that it will be more than 2 decades until we see self-driving cars as a reality across multiple cities. Of course, I would be more than happy to lose the $100 as it would make everyone's lives better that much sooner. But I'm skeptical.

Call me a technology optimist. I'd bet 10 years.

You want to get in on this bet? We could open up a pool on the prediction markets.

Probably not. I don't think it's sufficiently well-defined for me to trust an arbiter to resolve the bet satisfactorily. Also, time value of money.

http://longbets.org/ would help facilitate this, though the minimum is 200 dollars.

I'd agree. I think in 10 years we're starting to have the conversation about when human driving should be banned in metro areas. At that point it'll be obvious that the technology has arrived, and the safety record will be there, but not everyone will have the technology. There will be the problem of telling someone with a jalopy that they need to park that outside of the metro and pay for an Uber to their destination which will probably snarl that up for at least a decade or more... At some point, though, as rich people are zipping along in cars that never get pulled over for speeding or having taillights out, the "Uber tax" is going to be calculated to be less expensive than the "PoPo tax" on the working poor.

I'm rather skeptical on that happening anywhere near 10 years. Even if the tech was ready to go full-scale tomorrow, just the process of tooling up to make tens of millions of these cars and actually selling them to people, letting them trickle down to lower-income used car buyers and letting the oldest manual cars trickle out of the market could take decades.

Once the first ones are available on the open market, I'd expect at least 10 years of them being very expensive rich peoples' toys while the remaining edge cases and kinks in the tech are worked out, manufacturing scales up, insurance and legal issues get sorted out, various road maintenance departments learn to deal with their needs, etc.

One reasonable scenario for rapid deployment of self-driving cars - is shared vehicles, driving a few people at a time over a shared but optimal route.

It's possible that such vehicles will have 5x-10x total occupancy per day of regular vehicles(which sit idle at most times). And it could imagine how the right financial incentives would shift a large share of the manufacturing capacity towards such cars. And considering the car replacement rate at probably something like 20 years - maybe 10 years is a possibility of starting to have that conversation.

That's an good possibility for getting self-driving cars to market sooner and getting more people in them all right, but I'm thinking we have to get a lot of manual cars out of the market before anyone starts taking the idea of restricting them seriously.

Rush hour patterns imply that a lot of that idle time would likely remain idle. EDIT: Fixed typo

I think 10 years is just a bit optimistic. I think it'll take about 10 years from the time we have the software/hardware at a point where it is ready. I think that could be 5 years. I think it'll be another 15-20 years after that where the majority of cars on the road are self driving. Baring any crazy legislation, which I don't think would happen, in the US anyway.

I argued that it'd be 10 years when we'd start the discussion, and that it'd be 20 years from now before metros start to go driverless.

Yeah, I apologize, I meant to respond to the GP and accidentally responded to you. I don't want to clutter the thread with duplicate posts though, so I'll just leave it as is.

Even if it were possible to go buy a fully automated car today, there's no way enough of the fleet would have changed over in as short a time as ten years to make such a ban practical. And it's not possible to buy an automated car today; nor are the research prototypes Google and others have been working with up to a level of sophistication where simply productizing them would yield a fully automated car which could be used as a generic replacement for current automobiles.

I could believe that the first generation of fully automated cars might be available for purchase in ten years.

> there's no way enough of the fleet would have changed over in as short a time as ten years to make such a ban practical.

Cars sit idle 95% of the time. That means you only need to flip 5-10% of the total fleet. Tesla could do it by themselves building 2-4 million cars/year, with the amount of cars needed per year declining the higher the utilization level you can achieve.

Yes, but they all sit idle the same 95% of the time. Everyone needs a ride at rush hour.

> Yes, but they all sit idle the same 95% of the time. Everyone needs a ride at rush hour.

Disincentivize rush hour using congestion pricing.

You could divide rush hours maybe to 2 shifts of an SDC. And there's always the possibility of shared rides.

I don't understand how replacing a small minority of the fleet with automated cars leads to a situation where it is reasonable to talk about banning human drivers.

You would always consider banning humans from an activity if it was overly dangerous compared to having a machine perform the operation.

40K people a year die in the US from vehicle accidents. Why would you not ban humans from driving if machines were objectively superior? Where do you find the "right" to drive to be? Driving is a privilege granted by the government, not a right.

The comment I replied to was suggesting that vehicle automation would progress so rapidly that it would be practical to start working on such a ban as soon as ten years from now. I'm not disputing that such a ban might eventually become a good idea, but for it to become practical so soon, we'd have to be much, much further down the road to full automation than we are now.

If fully-automated vehicles become commercially available (which seems likely at some point, though not in the immediate future), we will likely see rapid adoption in the applications where they are very well suited, followed by a long, slow, gradual trend toward broader usage. The hardest cases will be the ones sticking around longest, and one can't simply legislate them out of existence. Just replacing a small percentage of the fleet won't come anywhere close! It's only when automated cars have become so dominant that only a small percentage of the fleet still depends on manual drivers that it will become practical to start talking about a ban on manual operation.

So I argued we'd start the conversation in 10 years and that it'd take at least another decade. I that is 20 years until we'd see the first ban against human drivers in metro areas.

To be there I do think we need the first generation of fully automated cars available to the general public in 5 years or less, which I think we're actually close to or we wouldn't be having this very discussion...

Ah, I didn't notice the second ten-year span in your prediction. I still think your perspective is wildly optimistic, but it makes more sense now.

I think you or the colleague are more likely to be correct than 40+ years. There are hugely capable companies driving this and more understanding than ever on what will be achievable and how much money will be on the line.

Self-driving, electric cars will make future taxi-equivalents affordable enough that it's considered for many instances. That makes for a massive market.

It's not as though current taxis have a sparkling reputation as drivers either. Not many will be sad to see the end of them.

I have a bet going on with my wife that our 1-year-old daughter will never take driving lessons when she grows up.

I'm hoping there's a viable SDC on the market when my eldest daughter turns 16 in 5 years

Do you live in the suburbs? This could become true, but more because of the increasing urbanization trend making public transit more attractive than cars in more places.

We haven't exactly spelled out the terms precisely, but the spirit of the wager is that self-driving cars will be involved somehow. Certainly public transit could be part of a mix of transportation options, but I wouldn't win if self-driving cars don't come to fruition yet she doesn't get a driving license because she can subsist entirely on public transit.

I have the same plan for my four year old!

You're right that there's significant legal and regulatory hurdles that have to be cleared, but I think they're more of an issue for broad adoption. SF will have a driverless taxi service probably within the next 5 years. Once that's in place it's only a matter of time/strength of the taxi lobby that's going to stop it, (and Uber has shown that can be broken even in places like London).

If you're talking about _everyone_ in major economies having a driverless car, yeah, I think 40 years is realistic given the replacement rate for cars alone, nevermind the legal issues.

I know what you mean, but London is a poor example — the well known "black cab" service has been heavily regulated forever, but cheaper "minicabs" have been available to book (rather than hail on the street) since the 1960s, and subject to a little regulation¹ since 2001. Most operators were tiny companies with a few cars servicing a local area, until Addison Lee started running a London-wide service about 10-15 years ago — they had nicer cars, better availability and marketing and so on. Uber set up in London under the same regulations.

¹ Additional car inspection, criminal record check on the driver, etc.

https://en.wikipedia.org/wiki/Taxicabs_of_the_United_Kingdom...

> He is sub 20 years. I'm thinking more like 40-50 mark.

Perspective:

* 60 years ago, we didn't have highways.

* 45 years ago, seat belts weren't required.

* 40 years ago, we expected engines to last 10s of thousands of miles. We didn't really understand crumple zones. Open containers and even drinking while driving.

* 30 years, we added fuel injection, computers, and airbags.

* 20 years ago, the hybrid car, better fuel efficiency, and side airbags

Imagine what we'll know tomorrow.

Perspective: nearly 50 years ago we did have Category IIIa instrument landing systems in commercial jet aircraft capable of landing without pilot input. We had reliable gyroscopic autopilot capable of keeping aircraft on a set course long before that.

Vast sums of money have been sunk into developing exceptionally reliable avionics systems since then.

Passenger carrying aircraft still all have a pilot or two.

Pilots on modern large aircraft are more like managers that tell the aircraft what to do, and do fly them to keep with minimum legal requirements (minimy numbers of landinds etc...)

And in the US, don't pilots still have to do the landing part?

And to complete the thought: the main reason why human pilots are still on all those aircraft is because... the automated/computerized systems still can't really make decisions that match what the humans can do. So the computer is really great at maintaining heading, altitude and speed for a set period of time, but really bad at deciding which heading, which altitude and which speed, and for how long (the best you can do that I'm aware of is have a human lay out the desired route and punch that into the FMC, and then all bets are off if conditions change along the way, and the humans will have to decide what to tell the plane to do again).

And those decisions themselves could be easily automated, but people choose not to. There's a checklist for making those decisions. If you can write a checklist, you can write code. And for when there's not a checklist, then you can pick randomly, because that's what the human is essentially doing.

A human can notice when new evidence becomes relevant. If something isn't already on a computer's checklist when it becomes relevant, the human has a better chance of deciding what to do.

EDIT: I understand that there is research going into making computers recognize "novel" situations. My comment only applies to algorithms that contain "checklists".

No, then you need to get meta. There's a checklist for identifying "novel" information and reacting. By checklist I mean algorithm.

Sure, it can be tough to tease out the algorithm from the minds of current pilots. Luckily in aviation we already have training manuals. In other domains it'd be tougher to know when we're done extracting knowledge from humans.

How close we are to the autopilot which would safely land a plane with engines off on the Hudson?

a) A human pilot who does combines exceptional luck with unusual skill

b) Said human pilot was aided considerably by the airplane's autopilot system

c) I'm not sure why you think software would find this particularly challenging?

It's not the landing that's difficult, it's the "write an algorithm which correctly identifies which specific parameters should cause the aircraft to choose a river near the airport as its landing site, and get approval for it" that's the difficult part.

How often does it happen, though? You don't need a human on every flight in order to deal with that. Build a few ground-based emergency remote-control centers across the continent, and train the autopilots (or the air traffic controllers) to call them when things get bad.

I'm not convinced anyone in the commercial aviation industry would agree with you.

Some on-ground duty pilot sitting in a simulator who's suddenly placed in control of an aircraft that's notified an anomaly seconds before landing has absolutely zero chance of being able to make the same judgement call as a human (co)pilot who's been sitting at the controls for the last few hours. And the likelihood of human input being required is highly correlated with the possibility the aircraft has lost connectivity and/or some of the equipment that's supposed to be sending back data is malfunctioning (and yeah, you could build in more redundancy, but that still doesn't eliminate the problem and probably costs more than the pilot)

I agree in a sense, but one thing to keep in mind is that all automation requires a large investment, especially in systems that cannot fail.

If the cost of that automation (programmers, testing, etc) is greater than the cost of training pilots, then pilots are going to remain.

Usually, there ends up being a middle ground - you automate the low-hanging fruit, and leave the last percentage points to the humans.

The exact opposite has happened to cars. We've delegated route building to our GPS devices, and we maintain heading and speed until the computer tells us to do something different.

So you're agreeing with GP poster that the legal objections are what will prevent self-driving cars on a large scale?

"60 years ago, we didn't have highways."

Not quite. Here's the US highway system in 1955.[1] Construction of the US Interstate Highway System started in 1956.

[1] http://www.mappery.com/map-of/Pre-Interstate-US-Highway-Syst...

He probably meant Interstates, and 1956 is 60 years ago, which is why I'm guessing that. Interstate 5 (Seattle portion) was only completed in 1969[0] so depending on where you are, it could be significantly less.

[0] - http://www.historylink.org/index.cfm?DisplayPage=output.cfm&...

Good point. I-80 wasn't complete coast-to-coast until 1986. There were gaps in the Western sections for years. That was the first coast-to-coast Interstate, the old Lincoln Highway route. I-40, the "Route 66" route, was completed even later.

All these were some sort of win win, not hard sells, low resistance. SDV are rotating the notion of driving and the economy around it, car dealers, oil companies. It won't be the same game at all in my view.

> 60 years ago, we didn't have highways.

What?

> 30 years, we added fuel injection, computers, and airbags.

But we knew about them for a lot longer than that. Fuel injection was used in aircraft during WWII.

>so much physical, legal, and regulatory infrastructure needs to change to make this viable that it will be more than 2 decades until we see self-driving cars as a reality across multiple cities.

It could be a lot more straightforward than you suspect. The three things we have going for us:

1. Car accidents are almost always resolved by deals made between insurance companies (this will be doubly true with unpiloted cars, which will presumably not commit criminal acts)

2. The same few large companies are likely to deal in insurance of piloted and autonomous vehicles, so there's not a huge incentive to go sue-crazy

3. If autonomous vehicles cause significantly fewer accidents and companies can charge the same amount for insurance covering them, that is a huge win for some very large companies with the power to get the ball rolling.

Increased vehicle safety will be a big win for these companies in the short term, although in the long term some price discounts might be factored into the cost of autonomous vehicle insurance.

>> but so much physical, legal, and regulatory infrastructure needs to change

Let's go with that. Say that tomorrow Google finished the development of the car. It works very well.

Than it just need to find a single city that would agree to a wide scale deployment. And I imagine there are plenty of reasons why a city might want to do that. Heck if Google cared, they could legally bribe politicians. And it doesn't even have to be a city in the west. It could be a city ruled by a dictator somewhere in africa. Or it could be a self driving car created by a chinese company, and china would be willing to take the risk to lead in this field.

So maybe the first city isn't as hard as it seems.

Now what happens after a car has proven itself in a city for a year or two ?Showing almost zero accidents, and all the other benefits of a self-driving car. How much will people everywhere would want it ? how much political support would it gather ? i imagine a lot.

Would such support and proof be enough for relatively rapid deployment of self-driving cars ?

Careful, that $100 probably won't be worth shit in 20 years let alone 40-50.

I think you are seriously underestimating what the baby-boomers will do when they have to start giving up their drivers licenses because they are getting to old to drive.

Humans overestimate tech advancement for 3yrs in the future, but underestimate it 10 yrs in the future.

Food for thought.

While I agree in general - I can think of a lot of technological predictions from my childhood in the 70s and 80s (fusion power, widespread off planet colonies, general AI) that we still seem to be waiting for.

Considering the number of accidents after major storms I think your overstating the case. People are bad drivers in all conditions.

The problem with driving is you have to consistently do it well. You can make zero mistakes 99.99% of the time and be a terrible driver. One accident per 10,000 minutes would be ~one accident every year. Worse, a lot can happen every minute, so it's closer to not messing up for 600,000+ seconds.

The only reason people can drive is we build in very wide tolerances. ex: Car lanes are far wider than strictly necessary, you can fit ~2 cars in the average lane. Traffic lights have a significant delay where they are all red etc.

> The only reason people can drive is we build in very wide tolerances.

And other drivers react to minimize the problem created by another driver.

I think it's more accurate to say that people are unreliable drivers. Humans are certainly capable of driving very very well, but they perform unreliably.

Actually, it sounds like the self-driving car got muscled out by the bus. They claim that the the car AND the driver of the car thought the bus would let them through. Now this is a problem with the car not driving defensively enough, because either the car miscalculated the bus's ability to slow or assumed the bus would be able to see the car and react accordingly (or, just possibly, the bus driver is an ass who tried to shut out the car. I know, its practically unheard of, but its possible). I don't think this is a deathblow to self-driving cars, nor is it a sign that they are worse at driving than people, only that they are still being perfected and that the single greatest threat to a self driving car are people because they are the ultimately unpredictable.

Anecdotally, this is a lesson I too had to learn, almost the hard way. The closest I've ever come to accidents in city driving is with municipal buses. City buses seem to drive as though they always have the right of way, and other cars will always get out of their way.

Coming from a smaller, more suburban area, I learned to recalibrate my driving to be more aggressive in larger cities (primarily NYC & DC), to hold my own against pushy taxis, etc. Most city drivers kind of bluff, and if you show you're not falling for it, they'll stay out of your way.

But city buses don't bluff. If they start pulling into your lane, even if clearly cutting you off, you better slam on those brakes, because that bus is not stopping. It's one of those unwritten rules you learn, at least in my experience driving in major East Coast cities.

I can see how it'd be hard to train AI that you have adjust the driving style depending your locale and for specific types of non-emergency vehicles.

I'm not making an argument here, really just sharing a chuckle about the crash involving a city bus.

  But city buses don't bluff. If they start pulling into
  your lane, even if clearly cutting you off, you better
  slam on those brakes, because that bus is not stopping.

Bus drivers would learn just like you did if there were a few thousand unmanned cars whose programming assumed every other driver would follow the law :)

Reminds me of this story:

http://www.reddit.com/r/reddit.com/comments/fd667/while_ridi...

The city buses effectively do have the right of way, at least in the sense that, if you have an accident with a city bus, you are going to be considered at fault, rules notwithstanding.

I'm not convinced that the distinction between "situations requiring thinking" and "boring situations" is that strong. You seem to be implying that self-driving cars will probably be worse than humans at things like negotiating through city traffic (where this accident apparently occurred), but is there any evidence of that?

It's true that most deaths and serious accidents occur on highways, where human factors like boredom and sleepiness are probably significant. But humans also have a heck of a lot of fender benders like this incident, and a lot of the ones I have seen weren't exactly situations where "a lot of thinking" was necessary. People are texting and rear-end the person in front of them, or underestimate how much room they have to switch lanes between two cars, or other things like that. I don't see why these situations are uniquely difficult for AI to solve. In fact, they seem like very apt problems for computers to solve.

On the other hand Google has driven a lot of miles without an incident and this minor accident doesn't change the fact that their system is safer than a human behind the wheel. And unlike humans who don't improve after an accident, Google's whole fleet just got even more safe.

This is false. Google's cars have only gone as far as they have without an incident BECAUSE of human drivers. In a 12 month period, human drivers prevented Google Self-Driving Cars from causing ten accidents. And 272 times the car's software outright failed and dropped control of the car to the human driver. This is all in Google's recent report to the California DMV, but it's not a reality they like to advertise openly.

Statistically, Google's Self-Driving Car would've lost it's license by now, if not for human drivers keeping it in check.

Spread that over "the equivalent of 75 years of typical U.S. adult driving." (1million + miles) and the alpha version of this software seems almost on par with the average driver which is expected to file an insurance claim every 17.9 years.

PS: How often do you think the average drivers ed teacher uses there break?

Your statistic is incorrect. The incidents described took place over a much smaller number of miles. Only I believe 14 months of the Self-Driving Car program, not the sum total of it's driving to date. (I said 12 above, but I think it's actually 14 months.)

424,331 miles, according to the report. http://www.consumerwatchdog.org/resources/cadmvdisengagerepo...

Nor are you counting all of the times the driving software handed control back to the driver. Consider each one of those equivalent to your human driver falling asleep at the wheel.

"Google had operated its self-driving cars in autonomous mode for more than 1.3 million miles. Of those miles, 424,331 occurred on public roads in California during the period covered by this report" So, you where talking about a specific report not the overall system, got it.

Our objective is not to minimize disengagements; rather, it is to gather, while operating safely, as much data as possible to enable us to improve our self-driving system. Therefore, we set disengagement thresholds conservatively, and each is carefully recorded.

That's a long way from falling asleep at the wheel. That's we don't want to put people at risk so we are going to hand off control if "Disengage for a recklessly behaving agent" etc.

Disengagements while on public roads is the only representative metric. The real world is more challenging than whatever artificial testing environment Google has on their private roads.

As you correctly point out disengagements may not be the equivalent of falling asleep: they might simply be the equivalent of the driving instructor adjudging the learner at the wheel to be looking a bit too nervous, or just not ready for the junction that's coming up yet. And yes, its assumptions on the type of "software discrepancy" and "unwanted maneuver" that prompt human input are conservative by design.

But the report actually goes so far to points out that on at least 13 of those occasions Google believes that without human intervention a crash would have occurred. It also implies that only 3 of those situations were created by poor driving on the part of another human. That's a definite crash due to Google car error averted every ~40k miles, even though humans already take over whenever there's a software warning, a spot of rain or something else that they think might stretch its capabilities too much.

The average US driver has 1 accident per 165000 miles (which, like the unusual tendency for the Google car to get bumped when stopping at junctions, may or may not be their fault) and that's including DUI as well as people driving in slightly more difficult conditions than sunny surburban California.

> The average US driver has 1 accident per 165000 miles (which, like the unusual tendency for the Google car to get bumped when stopping at junctions, may or may not be their fault) and that's including DUI as well as people driving in slightly more difficult conditions than sunny surburban California.

The average driver doesn't do all city miles and that's why Google's cars have gotten bumped. Highway driving is much easier to automate (several cars already do this today) and makes up a large percentage of total miles driven.

"424,331 miles, according to the report"

That is a useful data point. Extremely optimistic bystanders think self driving cars will lower death rates to zero.

Statistically about 7 people die on the public roads per billion passenger miles. At half a million miles, assuming the miles are statistically random and indistinguishable from the entire country (sure, summer in socal is as hard driving as a winter in a blizzard at night in Montana, sure...) then 424331 death free miles means that technologies death record is no worse than 336 times worse than human drivers, at least so far. Perhaps its only 335 times more deadly than human drivers. Even drunk people aren't that dangerous.

Or in other words, the self driving car can be summarized to lots of predictions based on very little data.

I'd like to see some analysis of the predicted severity of the accidents for self driving cars.

A minor fender bender at low speed vs. losing control and going over a guard rail at 80 mph are both "accidents" - but have entirely different consequences.

I have a hunch that while human drivers are still better, when they do have accidents some percentage of those tend to be much more fatal.

They have driven millions of miles - in perfect driving conditions: great weather, lots of highways, all of it by day (how do I know that? Simple: Google would brag about their AIs being able to drive in rough situations if they were able to do that. But they don't. They just brag about a huge number of miles in undefined conditions...). That is exactly what the original poster criticized. When driving is so easy that it bores human drivers to death, todays' AIs can be better.

The problem is: real-life driving does not only consist of situations that even a child could easily manage. Humans are extremely good at letting their driving skills degrade gracefully when conditions get rough. AIs? They drive perfectly until they reach their limit, then they suddenly pull off epic fails - if no human driver is on standby to resolve the situation immediately.

I guess I was dreaming last Friday when I saw the Google self-driving car pass by twice on the same route, at 10:30pm. I even handled a stop with very limited visibility on one side like a champ. If anything, I think self-driving cars will end up driving like my Grandma, but probably safer.

Your grandma, with the ability to receive telemetry from every other car around her, knowing her position to within a few mm, having a reaction time governed by light speed delays, and knowing as much about the Physics of driving a car as an army of PhDs has been able to teach her.

And programmed to turn off the blinker light after merging onto the freeway...

Last Friday's weather in Mountain View wasn't very stormy IIRC

Well, for one, it could have still been driven at that point by the driver.

> all of it by day (how do I know that? Simple: Google would brag about their AIs being able to drive in rough situations if they were able to do that. But they don't. They just brag about a huge number of miles in undefined conditions...)

They totally drive by night too. How do I know that? Simple: You see them with your own eyes, in Mountain View. Even if Google doesn't brag about it.

Google mentions focusing on city driving instead of highway driving in their annual disengagement report [1]:

> Mastering autonomous driving on city streets -- rather than freeways, interstates or highways -- requires us to navigate complex road environments [...]. This differs from the driving undertaken by an average American driver who will spend a larger proportion of their driving miles on less complex roads such as freeways. Not surprisingly, 89 percent of our reportable disengagements have occurred in this complex street environment (see Table 6 below).

1: https://static.googleusercontent.com/media/www.google.com/en...

It's hard to say with such a short description, but as i read it, the google car never left its lane, just moved right to the edge of it's lane to avoid some debris.

Why did the bus move into an occupied lane?!?

Sounds like we need a self-driving bus.

> Why did the bus move into an occupied lane?!?

It depends on the city, but bus drivers often don't care whatsoever about other vehicles in the way. They merge and let others avoid contact.

It sounded like it was merging into the bus's lane. But I imagine Google will prep some very fancy videos of the incident from the car's perspective within the next week or so, and then we'll know for sure.

> Why did the bus move into an occupied lane?!?

Because that's what busses do all the time, whether or not that is safe, legal, or desirable.

From the writing in the article, I found it extremely difficult to get a good visual of exactly what happened in the accident. I wish the description was a little more, uh, descriptive.

You got that backwards. They moved right in their lane to slow down for a turn (while allowing traffic going straight to pass), and then had to move back into the stream of traffic as there was debris preventing them from turning on the far-right side of the lane.

At least in my country, buses have legal right of way when they are driving out of bus stop.

Bad drivers take on a whole new meaning outside of wealthy countries where the dangerous drivers routinely end up disqualified from driving. Just by providing a predictable vehicle everyone can trust will stop and stay stopped at red lights, stop at stop signs etc AI will bring a lot to the world.

    Seventy-four per cent of road traffic deaths occur in middle-income 
    countries, which account for 70% of the world’s population, but only 53% 
    of the world’s registered vehicles, burdens for 74% of world's road 
    deaths. In low-income countries it is even worse. Only one percent of the 
    world's registered cars produce 16% of world's road traffic deaths.

https://en.wikipedia.org/wiki/List_of_countries_by_traffic-r...

People are bad drivers. Every winter at the first snowfall there are tonnes of accidents because folks forgot to adjust following distances.

The promise of an AI driver is that there will never be a second accident of the same type. The only question is how fast the learning rate can ramp up.

The problem you are stating is called the 'handoff' problem which I believe will be a serious challenge for manufacturers of driverless cars. The handoff problem is when an automated system identifies a scenario in which it needs to hand over control to a human. How will that work in a vehicle that is not designed to be occupied by someone that knows how to control it? This problem has been written about in the book Our Robots, Ourselves [0] and there is an Econ Talk podcast about it as well [1].

To me, it makes sense to make use of the brain power provided by the occupant of the vehicle. For example, if one of the main vision sensors in the car breaks, why not let the human occupant drive the car manually?

[0] http://www.amazon.co.uk/Our-Robots-Ourselves-Robotics-Autono...

[1] http://www.econtalk.org/archives/2015/11/david_mindell_o.htm...

As far as I understand, the current behaviour of self-driving cars is to stop the cars safely in a safe place, to the best of their degraded ability. Once it's stopped, the human can either take over. If a human can't, you're stuck on the side of the road and you call a tow truck.

Which isn't really any different than any other sort of major mechanical failure.

That's all well and good if you make it to the side of the road, but what happens when the car decides it can't make it that far and stops in the middle lane of an Interstate?

What happens if your car breaks down such that it can't make it off the interstate?

It's the same principle. You try to do the safest thing you can, but yes, at some point a car will fail and be stuck in the middle lane of an interstate because the autopilot fails, just as can happen because of mechanical failure.

I'd expect major mechanical problems with robots, like "the vision sensor died in the middle of an interstate," to be about as rare as major mechanical problems with humans, like "they had a heart attack in the middle of an interstate." If not rarer.

We're already okay with the possibility that a human might drop dead in the middle of an interstate and lose control and more humans might be injured or killed as a result. As long as it is not more likely that a self-driving car will drop dead, I don't see why it's any different.

(And this isn't even getting into things like the possibility of catastrophic failure being lower and the options for recovery being much higher. You can run two robots with redundant hardware; you can't run two humans quite as easily. A human who loses vision will lose their concentration; a robot that loses vision can use its last known positions of nearby cars to attempt to pull over immediately in an orderly fashion.)

The other self-driving cars will cope with the detour much faster and graciously than humans could.

The same thing that happens when any other car breaks down in the middle of the road--a traffic jam. Cars (especially ICE ones) break down all the time, we know how to deal with it.

Two options for how a driverless car without human controls might handle these sorts of hand-off scenarios:

1) It hands control to an out-of-car human driver, similar to how military drones and driverless trains are operated now.

2) It turns on the hazard lights and pulls to the side of the road. Another vehicle is dispatched to pick you up.

From a safety engineering standpoint, handoff is a very, very bad idea. When you automate the common operation of a system, how is the operator supposed to get a feeling for the system? And when the operator doesn't handle the system regularly, how is he supposed to handle a situation that's so dangerous and unusual that the automation can't handle it?

If you have years of experience driving a car, maybe it'll work. For some time. But what if your last manual intervention was five years ago? Or twenty years ago? The rarer handoff happens, the more stress it puts on the manual operator, and the more likely things go wrong.

It's much better to have a safe default reaction (stop the car, shut down the reactor, ...) that kicks in when normal, automatic operation can't continue.

You probably misunderstood what I meant by 'handoff'. It refers to the range of programmed interactions between an automated system and a human, for example, a warning light on the dash is a handoff. A warning sound or an automatic request for user action is a handoff. Pulling over and shutting down is a handoff. You say it's a bad idea. It would be a very very bad idea to build systems that do not have these responses programmed into them. It would be like throwing an exception without any exception handlers.

If you haven't already, I highly recommend listening to the Econ Talk podcast for much more depth from someone who has been thinking about these kind of problems for years (I can't really do it justice in a HN comment).

Thankyou! I've been saying this for a while, the current approach of 'driver assistance' or 'autopilot' is terrible. The more reliable the autopilot, the less attention the human 'driver' is going to be paying when the autopilot actually does say "okay turkey, you fly it" and shut off. And that is far more likely to happen in an unusual (read: hazardous) situation. You're self-selecting for a human panic reaction followed by a crash.

Hold up. Just because Google's problem is a big one to solve, doesn't refute the point that human drivers are bad.

Yes, driving is actually a challenging task. But guess what, humans are BAD AT IT. And are computers going to be better? I'd be pretty damn surprised if they weren't.

Most crashes come down to distraction. Computers never get distracted. So if we can workout errors like the one in this story, then at least we can say the computer isn't going to get lazy or distracted and suddenly forget it's lessons learned.

Yes, its a challenging road ahead for these self driving cars. But I'd feel so much safer in a world of autonomous vehicles than I would with the morons behind the wheel of today.

Not all of the humans are bad drivers, but some of them definitely are. I have been living in the US for 4 years and here are the list of things that are extremely dangerous:

- not using signals at all, just randomly changing lanes

- not using signals at all and on the top of that do not care about traffic in the other lanes

- driving 45 mph in the left lane because California teaches to new drivers that changing lanes are dangerous, aka "stick to your lane" (I learned that recently and I could not believe)

- driving with the exact same attitude in rain like under dry weather conditions

- do not care about people merging lanes, do not look at them, pretend they dont exist

- being on facebook on the highway driving with 65mph

- do not care about bicycles, pretending they are not legitimate traffic

I am not sure about the rest of US, but several friends told me it is specific to California. I would argue that humans are not as good drivers, at least there is a large part of drivers who aren't. Replacing drivers with software is always going to win, just like it did with airplanes or anything else repetitive task that we automated. I am pretty sure there is going to be a period while AI is not as good as humans but this is why Google is collecting to data at large scale to improve it to beat humans and make driving a safer way to commute.

>> And unfortunately there's no easy way to use the best of both sides.

It doesn't seem too hard to go in that direction. For example you can use satellties or drones or other sensing mechanism to survey a city in real time and tell the driver X minutes in advance before he would need to take control of the wheel - and of course you could adapt this recommendation according the skill of the person, his current state(being drunk) or other data.

Not sure that'd work: "Sorry I mowed your child down, but it wasn't my fault, the computer didn't tell me I needed to pay attention". People are in general extremely quick to deflect blame from themselves (especially under stress) no matter _how_ stupid or implausible their no-the-spot excuse is.

(Source: I'm a motorcycle rider, we have a term here "SMIDSY", which stands for "Sorry mate, I didn't see you" - because it's such a common thing for drivers to say when they've just driven into you. There's a great quote from a judge a few years back in a court case over a driver excusing themselves for running into a bike "But the plaintiff was _clearly_ there to be seen.")

If the vehicle told you to be focused ahead of time(say a few minutes before, enough mragin), and recorded that action than legally at that point in time, you're driving a regular vehicle.

> Again, look at any Asian mega city, people can adapt very fast.

Have you seen the accident rate in these places? It's huge, and in many countries underreported.

Personally, I feel we are approaching the self driving situation backwards. We are so heavily invested in current technology that we are trying to shoehorn existing tech into doin something it was not designed to be autonomous in the first place.

Imagine how much easier it would be, for example, if roads had embedded markers (like magnets or RFID chips) that designated lanes, speed limits, and other road rules and weather conditions. AI vehicles would simply need to read this information and adjust accordingly.

But, unfortunately, this would be almost unfeasible in today's world vecause a) it depends on every car following the same rules and that's impossible if you have human drivers and b) it would require massive investment and national-level cooperation to change the infrastructure.

> Imagine how much easier it would be, for example, if roads had embedded markers (like magnets or RFID chips) that designated lanes, speed limits, and other road rules and weather conditions. AI vehicles would simply need to read this information and adjust accordingly.

Speed limits can be defined as metadata on existing mapping data. Existing vehicles, such as Tesla's with their autopilot system and GPS receivers, can "define" existing lanes by refining existing mapping data using their drivers as expert system trainers [1].

You don't need national-level cooperation to change infrastructure; you just need a massively automated system for collecting data and training based on it. Google and Tesla are doing both, just in different ways.

[1] http://www.teslarati.com/tesla-building-maps-autopilot-drive...

It wouldn't help very much.

We need visual processing and heuristic logic to handle a kid on the side of the road chasing a ball unlikely to intersect and catch it before he enters traffic. Given that mandatory ability, reading a speed limit sign is pretty trivial.

Given that its very unusual for an oncoming car to swerve into your lane, and a giant pothole in the oncoming lane, and the car sees three cars swerve into your lane taking turns, what are the odds of the next car hitting you head on? Compared to this puzzler, interpreting a stop sign is pretty easy.

Also the human drivers presumably will not understand the weird behavior of an enhanced self driving car, unless the human infrastructure matched the automatic infrastructure. Of course you could analyze the human data visually to predict the human behavior, but then why bother with the automatic hardware sensors?

I guess a bad automobile analogy is if you must design this pickup truck to safely haul a 8000 pound trailer, then a 300 pound trailer will not overtax it.

Let's say for instance we wanted an RFID indicator for each lane every 200 feet.

You're talking about 13 million RFID chips that would need to be installed on 164,000 miles of roadway - just to get the interstate highways covered.

Now think about the fact that these chips are going to need to sit static and not be replaced for more than 15 years at a time. So you have to come up with a way to house them such that they will not be affected by > 150 degree heat or by < -50 degree (F) cold.

And then think about the fact that it takes ~6 months for the government wheels to spin in the right direction for a pothole to be filled in.

The project to put RFID, or any new technology on the roadways in the US will have to involve federal cooperation as well as the governments of 50 states and countless local governments. It will be 100 times more expensive than any other venture using the same technology.

Look at the Big Dig project in Boston. It took 25 years, $24 billion and several acts of congress to get it to a state where they could start working out the kinks like "why do the guardrails keep killing people".

The scale of solving this problem "the right way" is incredible. I don't disagree with you entirely. But I think any change like that is a solid 100 years out into the future.

Of course, you and I are pretty much on the same page. The complexity and cost would be staggering to redesign the whole thing. But maybe there are smarter, incremental ways to go about it. Perhaps mandate a special paint for lane markers to include magnetic paint or some kind of nanotech. We already do repairs on roads, repaint, replace "cat's eyes markers" etc, so modifying the materials used should be easier.

I fully agree it's going to take a long while for everyone to adopt this, and one of the keys will be proving the AI tech - which then becomes a chicken and egg problem. Maybe Google/Tesla are laying the groundwork for that, and on second thought maybe I'm wrong about us thinking about this the backwards. Perhaps this "over-engineered" solution (for the lack of a beter word) is exactly what we need to then later optimize for "the right way".

I think about the half-measures like new paint for lane markers and I'm equally skeptical, but that's probably just me being pessimistic.

All we really need is an idea that we can't fathom today. Google has a library of images of just about every street in the states. If they could run that same project and in the process lay out markers of some sort that would be helpful for the next generation of autonomous vehicles, then from that point forward incremental change is simply a matter of re-tooling an autonomous vehicle. Cooperation from local governments could be pushed through in a similar way that the 55 MPH speed limit happened.

Of all the "billion dollar ideas" floating around silicon valley, the person/investment group who's willing to take on an idea like this to a nationwide scale is going to be the winner.

The first round doesn't even need to be high tech. They could have a paint ball gun hanging out the side of a car spraying out a biodegradable path that a series of other cars can follow. Those cars could then train other cars.

...

Yeah, I'm flip flopping. I'm actually an idealist deep down inside. Sometimes I forget that for a little while.

I feel the same but refactoring the whole driving infrastructure will only happen if people see SDV as that much valuable and not overnight. If SDV get critical mass, what you say will happen.

If all cars were self-driven and there was a system that allowed them to exchange information, negotiate decisions, etc with all cars in proximity, that would be the ultimate vision safety-wise. I expect that we'll live to see that day. Removing the human factor would be removing the biggest problem.

On the other hand, making this mandatory would also be the wet-dream of most governments. In the name of safety, they could have access to so much information... Just knowing how many passengers the car is carrying, what's the weight of each passenger, and, of course, a history of the coordinates, would open insane possibilities, not all of them good.

> But when boring suddenly turns to not boring? The current state of AI is very very bad at this.

"Boring" versus "not boring" is too vague for this statement to really be meaningful. Examples could be drummed up on both sides. For example computers beat humans hands down in at least some criteria fitting the "not boring" description, like reaction time and seeing in multiple directions at once, and probably some others.

That humans are bad drivers is something most of us experience everyday. Of course, everyone thinks THEY are the exception.

Self driving cars being better than the average driver is absolutely to be expected since they can only get better. They'll be consistent. They will always be focused, not get distracted, not get tired, have more sensors than a human could. It's not just about the environment but also about the mental state of the driver.

Sure it's difficult but your using one case to flip things around doesn't seem to be a very strong argument.

> Self driving cars will get there someday, but I think it's much farther away than many assume it will be.

There's no evidence to support your assertion given the tests that have been going on so far. There's no evidence to support that switching back and forth would be better than having the AI all the time either.

> Of course, everyone thinks THEY are the exception.

This gets brought up every so often, much to my chagrin. If you asked 10 people what qualifies a good driver, you will get 10 different answers.

Does it mean they have a good safety record?

Does it mean they are skilled at controlling the car?

Does it mean they can navigate traffic efficiently?

Does it mean they communicate with other drivers well enough to not negatively affect traffic?

Does it mean they have good reactions times?

What combination and weighting to you provide the above questions and what questions haven't even been asked?

Of course people respond that they are above average or good drivers since everyone is using a separate metric to define a good driver.

How difficult is it to not hit something?

Let's suppose you are given a 2d field in which there are a number of rectangles, each of which have a velocity vector. So you know the location of each rectangle now, and, to a good approximation, in the near future.

How difficult is it to program a path-finding algorithm that will not crash into any of these rectangles, given that lowering speed is always a good option within a city?

I'd say this is really not all that difficult. Yes, there are a lot of special cases to consider, but still, in essence, this seems to me a quite simple problem.

So breaking down the problem is key. If you look at an autonomous vehicle as a black box, then, of course, it will seem dauntingly complicated, and difficult to trust.

> Let's suppose you are given

hard AI problem, computer vision

> to a good approximation, in the near future.

how near is near?

> lowering speed is always a good option

can't brake if someone faster is behind you

> there are a lot of special cases [..] simple problem

simple problems by definition don't have many special cases, the "simplicity" means you can abstract them into a few cases.

in other words, talk is cheap show me the code

IIRC, the autonomous cars all use a LIDAR system (kind of radar) to detect objects, and do not fully rely on computer vision.

Also, here's ([1]) an article about a student who (single-handedly!) wrote the software for a self-driving car, which doesn't use a LIDAR system and is much cheaper.

[1] http://www.nbcnews.com/technology/students-self-driving-car-...

There is a way to get the best of both worlds: variably-autonomous swarming vehicles. The larger the swarm, the more autonomy the car has. The smaller the swarm, the less autonomy.

Large swarms on the highway, small swarms (or solitary driving) elsewhere. Best of both!

only this one human-driven car in the middle of the swarm is what stops this swarm from reaching new level of autonomy. Opps, such an unfortunate mishap! that human-driven car has just hit the car in front of it and went off the road. Finally the swarm is free... err... can move onto the next level.

Personally - I think we need a network of "robot only" car lanes in cities and highways.

Turn one out of four one-way streets in business districts to robot only. Make a certain highway on/off-ramp robot only etc...

> And unfortunately there's no easy way to use the best of both sides.

How about using the "self-driving AI" only as "really advanced automatic assistance". Humans will still drive the cars, but the AI will kick-in in those boring situations where the humans turn off their thinking.

Interesting point. Would be interesting if this lead to remote drivers who are responsible for navigating self driving car when the car decides human intervention is safest? (At least until all cars on the road are self driving and connected to each other for real time communication...)

I will just leave this here: http://static1.businessinsider.com/image/52fcc3c76bb3f7bf62e...

Over 1 million people a year die from driving accidents. And that is not including the huge number of injured.

That is an awful result. Humans are awful drivers. They are too easily distracted, don't have enough information, bored, and tired.

> The current state of AI is very very bad at this.

What makes you think so?

Reply on news.ycombinator.com