Hacker News Reader

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.

"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/

> 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.

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 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.

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 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.

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.

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).

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."

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.

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.

  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 :)

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."

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

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...

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

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.

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.

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.