This was inevitable given Tesla calling the system an 'Autopilot', when in fact the driver must have their hands on the wheel and monitoring full time.
No, you don't, unless you want an exhausted pilot trying to land in a crosswind at the end of a journey.
Aviation autopilot is very much "set waypoints, sit back", as if anything goes awry it lets you know, loudly, and then you have time to deal with whatever needs dealing with. The whole reason autopilot was developed was to combat fatigue related accidents - and it works.
"and maintain situational awareness." Some failures can happen quickly, and the plane can change modes unexpectedly. Plus the warnings you receive may not be accurate.. see Air France 447.
You can sit back, but your role is to constantly check your instruments, your performance, your route, your radios and your environment.
Yes, but you don't need to sit there with your hands on the yoke, feet on the pedals, wired to the nines waiting for disaster to strike. I used to take a book with me for longer flights (when you're sat alone in a plane for hours flying over cloud things get tedious), as short of having a wing fall off there isn't all that much that can go very suddenly wrong - and most of the "very suddenly wrong" situations will kill you regardless. Keep one eye on instruments, mutter intermittent obscenities at the oil pressure gauge - but mostly relax.
Needless to say there's a huge difference between spectator and active participant. We're not at the point that a person can just sit back and read a book in a self-driving car, yet sitting there, hands on wheel, and doing nothing but watching - that can get tedious.
Using a soccer analogy, I can go to sleep watching certain Euro Cup 2016 matches, but if I were playing on the field, obviously that changes everything. With self-driving cars, I'm somewhere in the middle (basically, a benched player).
Does anyone else find it fascinating that for computers to be considered "safe" they must be as close to 100% as you can get, when the equivalent human drivers have fatal accidents every single day?
I don't really have any,apart from that the adoption of autonomous vehicles is going to be slower and take longer than people anticipate. Human traffic systems are complicated and not easily modelled, and humans are not easily reassured when events like this happen. While statistically people may be safer with an agent driving, most people don't crash most of the time, so they're more confident when they're in control than when someone else is.
If you actually replace all human drivers with all computer drivers, I'm pretty sure we'd hit the 100% safety mark.
I can't even imagine the resistance to self-driving cars once they're actually viable. Who even enjoys driving on city roads anymore? Or on highways? I drove on American freeways and it was mind-numbingly dull. Just acres of concrete all along the road with nothing interesting to see or do.
There are many other vehicles on the road also needing a human to drive them, tractors, heavy machinery, buses (that are not part of a regular, pre-defined route) ++.
In the future though, it will always be your fault, because you are one person trying to prove an accident was not your fault against a big-money company. Good luck.
"In the future though, it will always be your fault, because you are one person trying to prove an accident was not your fault against a big-money company."
That's why I always wear videoregistrator on my helmet. It paid off for itself many times already.
It will be decades before farm machinery is replaced by modern tech. Decades. I would like HN users to very much stop dismissing any discussion on the topic of automated vehicles with glib reactions devoid of thought.
Self-driving is a common feature on modern tractors. It's been available for almost 15 years.[1] Even when there's an operator in the cab, it's common to have automatic steering to keep the rows precise. The usual setup involves a pseudolite on the farm as a local reference for GPS correction.
Perhaps on modern farm equipment, but tractors last for decades and many of the tractors in rural Europe aren't even fuel injected. It will be many years until tractors and combine harvesters and the like are swapped out for newer equipment.
To add an actual data point, in New Zealand bike rider fatalities are not the motor bike riders fault about 30% of the time and in non fatal crashes the figure is about 40%. www.transport.govt.nz/assets/Uploads/Research/Documents/Motorcycles-15.pdf
My point was not that motorcycle drivers are worse drivers than car drivers. It's probably the other way around. But with computer drivers very soon much safer than human drivers, I'm very eager to get all trucks and cars driven by computers because it significantly reduce the risk of me or someone I care about being killed in traffic. But if someone wants to continue to steer their motorcycle themselves, I'm fine with that because it is unlikely that I'll be seriously hurt in a collision between my car and a motorcycle.
Anthony Levandowski, one of the key people behind Google's self-driving car effort, built a self-driving motorcycle for the 2005 DARPA Grand Challenge. It's now in the Smithsonian.[1]
Nice, but robot-on-two-wheels is not quite the same as motorcycle autopiloting with human on top, who doesn't sit still and doesn't handle sudden movements well.
Though to be fair I absolutely don't see the point in sitting on a motorcycle unless you're driving it. It's not like you can occupy your time with anything else than leaning over to hold the handlebars.
> Who even enjoys driving on city roads anymore? Or on highways? I drove on American freeways and it was mind-numbingly dull. Just acres of concrete all along the road with nothing interesting to see or do.
Who even enjoys writing code any more? It's just line after line of characters on a screen, with nothing interesting to see or do.
Yet another american claim that "I don't like something, therefore no one else must like it either", with the usual view that because something is terrible in America, it must be terrible everywhere.
Keep in mind, you're just below the world average (17) at that value. So while Im sure Indian roads are as bad as their reputation, there are worse.
I live in Thailand, on the same list (when sorted by deaths by 100K inhabitants), we're #2 (right behind Libya), with 36.2; the deaths per year for here was 24K, based on a population of just 65 million.
My point is, I know what it's like to drive on dangerous roads, and I still dispute your claim that driving is either boring or dangerous.
Even driving across the Eyre highway from Adelaide to Perth (~2700km), part of which is the '90 mile straight' - a section of road with 0 bends for 145km - is something I can look back and be glad I've done it.
If you personally don't enjoy driving, or feel unsafe driving, you personally can take actions based on that, such as taking a bus or taxi, or car pooling etc.
But don't somehow think that because your homeland has a reputation for bad roads, and America has apparently boring roads, that no one else on the planet enjoys driving.
I wouldn't say "fascinating" so much as "irritating", as this is more about journalists wanting to get a juicy story where there isn't one, rather than a lesson on human nature.
Statistically, autopilot is orders of magnitude safer than humans.
If there's a story here, it's about how the US doesn't require underrun protection on trucks, thus making them death traps, but that wouldn't be anywhere near as good for circulation as a "rise of the machines, they want to kill you" story.
150,000,000 miles on autopilot, one fatality. Humans, it's 66,000,000 miles per fatality, in the US. Admittedly the sample is small, but still, it suggests an advantage over humans.
Apples and oranges. The statistic you cite is all cars and roads, as far as I'm aware. High end vehicles with comparable safety features on roads where autopilot is likely to be engaged is more comparable.
I wouldn't be surprised if autopilot is safer, but I think it is overstating things to say this is a statistical fact.
One problem of automation is that when an algorithm fails it reverts to 100% stupid mode, without any fallback. In this particular case the program didn't understand there was a truck, so the car drove full speed into the truck. As a result, the outcome of accidents with AI driving, while rare, will probably tend to be rather gruesome and this might lead to some really bad reaction.
What's really interesting is how, once accepted, the computer becomes 100% accurate by default and anything conflicting with it is automatically wrong.
Likewise, human drivers manage to avoid accidents every single day by paying attention, reading traffic signals and reacting to a situation.
Computers have to be 100% safe because self-driving cars take control away from the user. And where does the blame lie in an accident? With the owner of the vehicle, the 'operator' at the time of the accident, the 3rd party developers who were contracted by Tesla/Nissan/BMW, or the car company themselves? Getting customer service and a dealership to admit responsibility for parts failure as it is today is near impossible.
It's fine for people to be be pro self-driving cars but it is wrong to dismiss any and every criticism of an emerging technology on that basis alone.
That quote's taken from the tl;dr-equivalent at the start of the article. It seems to be summarising the following, from further down, which I think is perfectly reasonable:
> Karl Brauer, an analyst with the auto research firm Kelley Blue Book, said the accident served as a signal that the technology might not be as advanced and ready for the market as some proponents have suggested.
> "This is a bit of a wake-up call," Mr. Brauer said. "People who were maybe too aggressive in taking the position that we're almost there, this technology is going to be in the market very soon, maybe need to reassess that."
I think that's perfectly reasonable, e.g. it may take 5 or 10 years to reach some kind of approved minimum safety level, especially since such safety approval standards/processes are still being developed. Just because Tesla's cars already have an "auto-pilot mode" doesn't mean the technology's ready for widespread deployment.
Regarding "as close to 100% as you can get", that's not stated in the article. Instead, it says:
> The federal traffic safety agency is nearing the release of a new set of guidelines and regulations regarding the testing of self-driving vehicles on public roads. They are expected to be released in July.
> At a recent technology conference in Novi, Mich., the agency's leader, Mark Rosekind, said self-driving cars should at least be twice as safe as human drivers to result in a significant reduction in roadway deaths. "We need to start with two times better," Mr. Rosekind said. "We need to set a higher bar if we expect safety to actually be a benefit here."
Again, I think that's reasonable. While a fully-automated road system would probably be safer than human drivers, we're not going to get an overnight switch; the harder problem is the transition, with manually- and autonomously-driven cars coexisting.
It's not just a case of swapping out more and more cars, since self-driving cars will behave and interact in different ways (as an obvious example, looking at the occupants won't give you any information about intentions/gestures/etc.); hence it's worth waiting until a threshold higher than break-even is reached before beginning the transition.
I love Tesla, but make Tesla liable for this. I don't care that "the driver was supposed to be careful - it's in the license agreement!" or whatever. If you offer any kind of "self-driving" technology, even if it's only for changing lanes, then the auto-maker should be liable for the accidents that happen because of that. Don't put the blame on the driver.
Also something I've said before, I wouldn't be a "beta tester" of self-driving technology for at least the first 10 years. I think that's stupid, but hey I'm glad others are willing to die to prove the technology for me, I guess. Even if you trust companies like Google, Tesla, or Apple to do it right, the hype could spill over to other companies like say Fiat, who are clueless about software, but will also offer "self-driving" tech, and people will believe it's just as good, when it may be far from it.
Make them pay for every accident that happens because of their technology, and you'll see how quickly they improve it and how carefully they release the technology. No other regulation may be needed. Other regulations may not be sufficient on their own anyway, and the auto-makers will just do the minimum necessary to comply with them, but that could still mean people could end up dead under those regulations.
This is such a slippery slope, especially in lawsuit crazy USA.
> I wouldn't be a "beta tester" of self-driving technology for at least the first 10 years.
Yes, I totally agree.
> Make them pay for every accident that happens because of their technology, and you'll see how quickly they improve it and how carefully they release the technology.
I don't agree that the civil legal system is the right way to solve this. I also don't believe software developers for Tesla should be criminally charged for the death of drivers if it's shown that the software developers were not negligent.
What I really want to see happen, is that any autonomous cars have their software independently and fully audited before they're shipped in production vehicles.
This way, there is a neutral third party, certified by the NTSB, that certifies whether a self-driving car is road legal.
We hold drivers responsible if they kill someone while behind the wheel. I think that we're setting ourselves for a dangerous precedent if we also expect to hold software developers criminally liable. This is why the software in self driving cars must be audited and certified. This means there is a legal burden on software developers to provide safe software, and it removes the criminal liability component should there be a crash.
Treat self driving car crashes less like a witch hunt, and more like an airplane accident. In airplane accidents, no one is trying to point the finger, everyone is simply trying to determine the cause, and how it can be avoided in the future.
> Other regulations may not be sufficient on their own anyway, and the auto-makers will just do the minimum necessary to comply with them
Exactly. This is why tougher regulations are needed. IIRC even ECU code in ICE cars is currently audited. This emerged during the Toyota accelerator lawsuit.[1]
> Make them pay for every accident that happens because of their technology, and you'll see how quickly they improve it and how carefully they release the technology.
...or you'll see how quickly the industry dies off instead of pursuing the technology. I am not sure I want driverless car tech to go the way nuclear tech did in the 70's (via fear, banned tech, and large introduction of costly regulations). Some kind of balance is needed. An extreme solution won't be beneficial in the long run.
>Autopilot has been used for more than 130 million miles, noting that, on average, a fatality occurs every 94 million miles in the US and every 60 million miles worldwide.
This technology, even if imperfect, serves to make driving safer for the user, not the other way around. It's your choice not to be a beta tester; certainly no one will be forcing you.
And as others have said, the one at fault in this particular accident is the (very human) semi truck driver.
I disagree. Tesla should only be liable if it made false promises (i.e. advertising "read a book while driving"), but that does not seem to be the case.
My Toyota RAV4 has adaptive cruise control and can auto steer back into lane if I'm too far off. I wished that it could auto follow a lane. It has everything that it needs to make that happen.
But I also like that they dumbed it down that I have to have the hand on the steering wheel otherwise it complains loudly.
Automotive car crashes should be investigated with the rigour of aeroplane crashes. The only way to make it safer for the next driver is to learn as much from the bug here and fix it.
I also strongly believe that cars should pass a rigorous test of edge cases and be scored on that.
Exactly. I've referred to this as the "deadly valley" - enough automation that the driver can tune out, but not enough automation to operate safely without driver attention. Google has done some studies on how fast a driver can take over in an emergency, and it's seconds, not milliseconds. Aviation safety people note that it takes about 30 seconds after an autopilot disconnect before pilots have full situational awareness.
Tesla, by intent, has built a system right in the middle of the "deadly valley". Tesla's "autopilot" is just lane-keeping and smart cruise control. All the major auto manufacturers have comparable technology, and many high-end cars ship with it now. But the bigger players have avoided promoting it as automatic driving, and restrict the technology to avoid driver over-reliance.
Google and Volvo both have better systems then Tesla's, but aren't shipping them yet. They both take the position that safety is their problem, not the driver's. Volvo's CEO has said publicly that if one of our cars gets into an accident while in auto mode, it's our fault. Tesla blames the driver.
The NHTSA is full of smart people. They know that a Tesla in self-driving mode is safer than when driven by flesh and bones. Data proves it. We have millions of hours of self-driving Teslas and only one fatal accident.
Therefore I predict that this accident (and other future ones) will not significantly slow down the development, and adoption of self-driving cars...
No mention of the fact that the human driven truck pulled out in front of the care. From the description it sounds like the car should have break breaking hard - one doesn't just hit the side of a truck - the driver didn't see the oncoming car either.
This is key. The truck was sideways in front of the car on a divided highway. That is not normal. And all the situations I can think of where that could happen involve really quick unexpected moves on the truck driver's part. And at high speed.
Combine that with the fact the truck was high enough that it was over the hood of the car (and therefor probably over the hight of many of the sensors).
I don't think as a human I would have done any better.
Even if self-driving cars will be 10x safer than humans, I doubt if they can eliminate all deadly accidents. Though many people die or get severely injured every day, each machine failure is treated as major news.
Similarly plane accidents are perceived way more and that perception persist over decades: "A sold-out 727 jet would have to crash every day of the week, with no survivors, to equal the highway deaths per year in this (USA) country."
http://anxieties.com/flying-howsafe.php#.V3YVE5N95AY
What that means for AI? Either we create 1000x more safer AI than humans, or we self-driving cars would face harsh public/regulatory opposition?
While flying is certainly safer than driving, that statistic speaks more to the gross volume of cars driven vs. planes flown than anything else.
The perception of danger in planes and self-driving cars is perceived as such due to lack of control. If a plane is about to crash, there's not much to do besides hold and pray. If your car accelerates when it's supposed to brake, you better be paying attention and act fast to avoid a crash. With your hands on the wheel, you're always in a position to act, and it's easier to believe that there's always something you can do to avoid disaster.
These statistics are typically deaths per passenger-mile travelled or similarly normalized. Or do you mean that highway traffic is more dangerous because there are more cars in a smaller space?
I don't think you can ever eliminate all accidents, but the ones that do happen shouldn't be as simple crashing into a truck in front of you because the sensors didn't detect it properly.
This article doesn't mention it but the Tesla press release pointed out... "the vehicle was on a divided highway with Autopilot engaged when a tractor trailer drove across the highway perpendicular to the Model S"[1]
Think about that for a second. How does that happen? Sounds like the truck was jack knifing, making an absurdly abrupt lane change, or the highway was not completed divided. Either way it's not like it was simply rear-ended, the car went under the side of the truck.
I'm not a Tesla apologist by any stretch but this sounds like a bit of a freak thing. My guess is the physical sensor wouldn't pick up something that high off the ground (keep in mind this was the side of the truck not the back where there is a bumper) and any camera sensors didn't have enough contrast. To prevent this they would have to have sensors point UP and forward, which I would guess they will be doing after this.
Don't forget manufacturer liability. 100x safer self-driving cars would still cause 330 deaths and 22,000 injuries per year. Each one of those could result in a negligence lawsuit against the manufacturer.
Even in the case of "mostly" autonomous vehicles where it could be argued that the human driver was lulled into complacency.
And it's unclear if the insurance industry would shoulder that like they do human errors. Serious accidents can bankrupt responsible people after insurance is exhausted, and those would all now be concentrated on manufacturers with much deeper pockets. Plus there's all the single-vehicle accidents which currently don't result in lawsuits, but would with autonomous vehicles.
I'm kind of surprised to learn that the autopilot was relying on visual sensors (since I haven't read much about the Model S). I wonder if onboard lidar or radar could have prevented this accident.
I think the Tesla does have front facing radar, but radar isn't as precise as a visual camera; so the autopilot hardware probably uses a combination of sensors.
EDIT: From the Tesla website:
> Every single Model S now rolling out of the factory includes a forward radar, 12 long range ultrasonic sensors positioned to sense 16 feet around the car in every direction at all speeds, a forward looking camera, and a high precision, digitally controlled electric assist braking system.
I know my car has all the sensors mounted low. I'm not a radar expert. If the side of the truck was significantly above where the radar is looking would it be able to see it?
I really do not like Tesla's strategy of luring the drivers to doze off while driving. However, the Guardian article on the accident claims that Teslas drove 130 million miles in autopilot mode, while the rate of fatal accidents is closer to 100 million miles per accident. [0] It is statistically complicated to measure anything from a single event, and ars had a nice article on the problems for self driving cars [1], but I would interpret this as a Tesla autopilot together with the human driver is as good (with some indication of slightly better) than a unassisted human driver.
Apart from that, it is important to remember that fatal accident statistics are not just numbers, but are tabulations of individual tragedy. My condolences to his family, but sadly we can not expect autonomous cars to be perfect in the first try.
There is only one data point from Tesla so i won't read into it too much either way.
That said, all Model S are expensive and fairly new care with good safety features compared to the average car on the road today. Thus,I wonder if it is better to compare the 1 fatality / 130 million miles statistic to Model S (and S-class / 7-series) fatality rates while not on Autopilot mode.
A comparison also needs to include the fact that Autopilot is really only used on highway like streets, which are much safer than most other types of driving.
->The roads traveled least are the nation's deadliest roads
Two factors for that tho.
Firstly is bad road markings. Which computers will be infinately better at handling.
But more importantly is the response time of emergency services and letting someone know theres been an accident.
Even minor injuries end up fatal if no one finds you trapped in the car for a week.
And every second counts for more serious collisions.
Only so much manufacturers can do about either.
Dui and driver fatigue. As you will see are among the big killers. (And before you think "yeah but i dont drive drunk or tired" remember you arent the only person on the road).
Think its just about being carefull with what is meant by "safer" and "other types of driving".
You are more likely to have an accident in busy periods on busy roads.
But you are much more likely to die if you have an accident somewhere quiet or on high speed roads.
Bad road markings are very hard for current self driving systems to solve.
Your other points back up the claim though - self driving cars need to be compared to like-for-like mortality: same roads, same demographic, similar safety features. My guess is they're currently considerably less safe.
The obituary of the driver, Joshua Brown, is now available. Age 40. Former Navy explosive ordnance disposal specialist. Worked with Navy SEALs. Studied physics and computer science. CEO of Internet of Things startup.[2] It would be hard to find someone more qualified to operate a vehicle with Tesla's "autopilot". Placing the blame on that driver probably isn't going to fly.
Whether Tesla has a lower accidents/mile rate (or is some factor better or worse) is a red herring [1].
The key factor in whether this is a legitimate flaw in Tesla's system is: could an alert, competent driver easily avoid the accident?. An alternate formulation of this criteria is: would it be relatively easy for an engineer to hard-code a brute-force solution to the edge case, if they really wanted to?
Obviously we'll need to see the crash video to verify, but my guess is the answer is yes, this is a legitimate flaw in Tesla's system. Tesla's system should have been able to detect a car traveling perpendicular to the highway [2]. Tesla's system was unfortunately not sophisticated enough to detect a car that an alert human would be able to detect in their peripheral vision, and their system was not sophisticated enough to perform the simple action that a human would be able to perform (slamming on brakes). In fact, if they ever want to scale up to a full urban-capable system this is a situation they need to be able to handle.
Examples of accidents where I wouldn't blame Tesla's system: the other car swerves across the center divide and hits you head on, the other car is speeding and rear ends you, the car in the next lane swerves into you, the tire blows out and you hit a tree, etc. Although in the long-term we would ideally have systems that could handle these situations.
[1] We need to wait for more miles and (inevitable) accidents to be logged rather than rely on this one data point. Also my guess is highway driving is less fatality prone than urban driving, on a per-mile basis.
I don't think that autonomous cars are anywhere near a situation where a large scale deployment is feasible. In particular, I think that there needs to be a cross-vendor car-to-car or car-to-road communication protocol. Until then, autonomous cars will only work as long as there are relatively few of them or they run in controlled conditions like major highways.
When I'm in traffic, I notice myself communicating with other people a lot. Most commonly it's just brief eye contact ("I have seen you"), but it enables smooth and safe traffic for everyone.
Just blindly following rules and reacting to your immediate surroundings won't work in large scale. There are plenty of situations where that would end up in a deadlock traffic jam (e.g. 4 way intersection with no right of way).
This is by no means an impossible task, I just have not heard of an effort trying to solve it yet.
I rarely communicate with other drivers (other than with signal lights and brake lights), so I don't actually think it's necessary for self-driving cars. The only use of communication is right of way at stop signs, but this can be resolved by seeing which car decides to move forwards into the intersection.
I think at these early stages of autopilot driving, drivers should always assume they are the ones driving as there are sharing the roads with mere mortals behind wheels. There are reports that the driver may have been watching a movie on the screen. I would have thought the driver should be unable to watch a movie on the screen while driving even in autopilot mode. For example, the range rover has a dual screen view where the front passenger can watch a movie on the screen, but the driver will not be able to view nor hear whats going on even though (s)he is looking at the same screen.
94 comments
[ 3.3 ms ] story [ 148 ms ] threadAviation autopilot is very much "set waypoints, sit back", as if anything goes awry it lets you know, loudly, and then you have time to deal with whatever needs dealing with. The whole reason autopilot was developed was to combat fatigue related accidents - and it works.
"and maintain situational awareness." Some failures can happen quickly, and the plane can change modes unexpectedly. Plus the warnings you receive may not be accurate.. see Air France 447.
You can sit back, but your role is to constantly check your instruments, your performance, your route, your radios and your environment.
Autonomous cars on the other hand seem to be targeted just at that, allowing the driver to do something else like read or text or something.
Using a soccer analogy, I can go to sleep watching certain Euro Cup 2016 matches, but if I were playing on the field, obviously that changes everything. With self-driving cars, I'm somewhere in the middle (basically, a benched player).
Does anyone else find it fascinating that for computers to be considered "safe" they must be as close to 100% as you can get, when the equivalent human drivers have fatal accidents every single day?
So if you see a Tesla car crashing on days with white sky, you want to know whether this can happens to you as well.
I can't even imagine the resistance to self-driving cars once they're actually viable. Who even enjoys driving on city roads anymore? Or on highways? I drove on American freeways and it was mind-numbingly dull. Just acres of concrete all along the road with nothing interesting to see or do.
I drive motorcycle (and enjoy that), I doubt autopilot will come to those anytime soon..
In the future though, it will always be your fault, because you are one person trying to prove an accident was not your fault against a big-money company. Good luck.
That's why I always wear videoregistrator on my helmet. It paid off for itself many times already.
I would expect that it's quite easy to autopilot a tractor on the road. They already use autopilot in the field.
[1] https://www.youtube.com/watch?v=D7xc45LyhoU
[1] https://www.precisionfarmingdealer.com/articles/2211-retrofi...
In Lithuania it's even worse, around 60% (2014 data).
[1] http://americanhistory.si.edu/collections/search/object/nmah...
Who even enjoys writing code any more? It's just line after line of characters on a screen, with nothing interesting to see or do.
Yet another american claim that "I don't like something, therefore no one else must like it either", with the usual view that because something is terrible in America, it must be terrible everywhere.
Driving has now become a choice between the dull, automated driving I did in America, or the mad, chaotic rush I see on Indian roads.
Neither of these are enjoyable or even particularly safe
Keep in mind, you're just below the world average (17) at that value. So while Im sure Indian roads are as bad as their reputation, there are worse.
I live in Thailand, on the same list (when sorted by deaths by 100K inhabitants), we're #2 (right behind Libya), with 36.2; the deaths per year for here was 24K, based on a population of just 65 million.
My point is, I know what it's like to drive on dangerous roads, and I still dispute your claim that driving is either boring or dangerous.
Even driving across the Eyre highway from Adelaide to Perth (~2700km), part of which is the '90 mile straight' - a section of road with 0 bends for 145km - is something I can look back and be glad I've done it.
If you personally don't enjoy driving, or feel unsafe driving, you personally can take actions based on that, such as taking a bus or taxi, or car pooling etc.
But don't somehow think that because your homeland has a reputation for bad roads, and America has apparently boring roads, that no one else on the planet enjoys driving.
Statistically, autopilot is orders of magnitude safer than humans.
If there's a story here, it's about how the US doesn't require underrun protection on trucks, thus making them death traps, but that wouldn't be anywhere near as good for circulation as a "rise of the machines, they want to kill you" story.
I wouldn't be surprised if autopilot is safer, but I think it is overstating things to say this is a statistical fact.
Source? I hear that it is imperfect software.
Computers have to be 100% safe because self-driving cars take control away from the user. And where does the blame lie in an accident? With the owner of the vehicle, the 'operator' at the time of the accident, the 3rd party developers who were contracted by Tesla/Nissan/BMW, or the car company themselves? Getting customer service and a dealership to admit responsibility for parts failure as it is today is near impossible.
It's fine for people to be be pro self-driving cars but it is wrong to dismiss any and every criticism of an emerging technology on that basis alone.
> Karl Brauer, an analyst with the auto research firm Kelley Blue Book, said the accident served as a signal that the technology might not be as advanced and ready for the market as some proponents have suggested.
> "This is a bit of a wake-up call," Mr. Brauer said. "People who were maybe too aggressive in taking the position that we're almost there, this technology is going to be in the market very soon, maybe need to reassess that."
I think that's perfectly reasonable, e.g. it may take 5 or 10 years to reach some kind of approved minimum safety level, especially since such safety approval standards/processes are still being developed. Just because Tesla's cars already have an "auto-pilot mode" doesn't mean the technology's ready for widespread deployment.
Regarding "as close to 100% as you can get", that's not stated in the article. Instead, it says:
> The federal traffic safety agency is nearing the release of a new set of guidelines and regulations regarding the testing of self-driving vehicles on public roads. They are expected to be released in July.
> At a recent technology conference in Novi, Mich., the agency's leader, Mark Rosekind, said self-driving cars should at least be twice as safe as human drivers to result in a significant reduction in roadway deaths. "We need to start with two times better," Mr. Rosekind said. "We need to set a higher bar if we expect safety to actually be a benefit here."
Again, I think that's reasonable. While a fully-automated road system would probably be safer than human drivers, we're not going to get an overnight switch; the harder problem is the transition, with manually- and autonomously-driven cars coexisting.
It's not just a case of swapping out more and more cars, since self-driving cars will behave and interact in different ways (as an obvious example, looking at the occupants won't give you any information about intentions/gestures/etc.); hence it's worth waiting until a threshold higher than break-even is reached before beginning the transition.
Also something I've said before, I wouldn't be a "beta tester" of self-driving technology for at least the first 10 years. I think that's stupid, but hey I'm glad others are willing to die to prove the technology for me, I guess. Even if you trust companies like Google, Tesla, or Apple to do it right, the hype could spill over to other companies like say Fiat, who are clueless about software, but will also offer "self-driving" tech, and people will believe it's just as good, when it may be far from it.
Make them pay for every accident that happens because of their technology, and you'll see how quickly they improve it and how carefully they release the technology. No other regulation may be needed. Other regulations may not be sufficient on their own anyway, and the auto-makers will just do the minimum necessary to comply with them, but that could still mean people could end up dead under those regulations.
This is such a slippery slope, especially in lawsuit crazy USA.
> I wouldn't be a "beta tester" of self-driving technology for at least the first 10 years.
Yes, I totally agree.
> Make them pay for every accident that happens because of their technology, and you'll see how quickly they improve it and how carefully they release the technology.
I don't agree that the civil legal system is the right way to solve this. I also don't believe software developers for Tesla should be criminally charged for the death of drivers if it's shown that the software developers were not negligent.
What I really want to see happen, is that any autonomous cars have their software independently and fully audited before they're shipped in production vehicles.
This way, there is a neutral third party, certified by the NTSB, that certifies whether a self-driving car is road legal.
We hold drivers responsible if they kill someone while behind the wheel. I think that we're setting ourselves for a dangerous precedent if we also expect to hold software developers criminally liable. This is why the software in self driving cars must be audited and certified. This means there is a legal burden on software developers to provide safe software, and it removes the criminal liability component should there be a crash.
Treat self driving car crashes less like a witch hunt, and more like an airplane accident. In airplane accidents, no one is trying to point the finger, everyone is simply trying to determine the cause, and how it can be avoided in the future.
> Other regulations may not be sufficient on their own anyway, and the auto-makers will just do the minimum necessary to comply with them
Exactly. This is why tougher regulations are needed. IIRC even ECU code in ICE cars is currently audited. This emerged during the Toyota accelerator lawsuit.[1]
[1] https://news.ycombinator.com/item?id=9643204
...or you'll see how quickly the industry dies off instead of pursuing the technology. I am not sure I want driverless car tech to go the way nuclear tech did in the 70's (via fear, banned tech, and large introduction of costly regulations). Some kind of balance is needed. An extreme solution won't be beneficial in the long run.
[0] http://www.salterhealy.com/blog/who-is-at-fault-in-a-left-ha...
>Autopilot has been used for more than 130 million miles, noting that, on average, a fatality occurs every 94 million miles in the US and every 60 million miles worldwide.
This technology, even if imperfect, serves to make driving safer for the user, not the other way around. It's your choice not to be a beta tester; certainly no one will be forcing you.
And as others have said, the one at fault in this particular accident is the (very human) semi truck driver.
It's not reasonable to expect that -- and this sort of thing will happen when you rely on that system working.
Edit: Also I realize that technically you are always driving. By driving here I mean in the active sense -- no automatic steering or cruise control
But I also like that they dumbed it down that I have to have the hand on the steering wheel otherwise it complains loudly.
Automotive car crashes should be investigated with the rigour of aeroplane crashes. The only way to make it safer for the next driver is to learn as much from the bug here and fix it.
I also strongly believe that cars should pass a rigorous test of edge cases and be scored on that.
Tesla, by intent, has built a system right in the middle of the "deadly valley". Tesla's "autopilot" is just lane-keeping and smart cruise control. All the major auto manufacturers have comparable technology, and many high-end cars ship with it now. But the bigger players have avoided promoting it as automatic driving, and restrict the technology to avoid driver over-reliance.
Google and Volvo both have better systems then Tesla's, but aren't shipping them yet. They both take the position that safety is their problem, not the driver's. Volvo's CEO has said publicly that if one of our cars gets into an accident while in auto mode, it's our fault. Tesla blames the driver.
Therefore I predict that this accident (and other future ones) will not significantly slow down the development, and adoption of self-driving cars...
[citation needed]
Tesla press release says this is the first accident in 130 million miles, whereas other vehicles had an accident every 94 million miles.
But that includes old cars, old drivers, drunk drivers etc.
Doesn't seem that much safer
Combine that with the fact the truck was high enough that it was over the hood of the car (and therefor probably over the hight of many of the sensors).
I don't think as a human I would have done any better.
Similarly plane accidents are perceived way more and that perception persist over decades: "A sold-out 727 jet would have to crash every day of the week, with no survivors, to equal the highway deaths per year in this (USA) country." http://anxieties.com/flying-howsafe.php#.V3YVE5N95AY
What that means for AI? Either we create 1000x more safer AI than humans, or we self-driving cars would face harsh public/regulatory opposition?
The perception of danger in planes and self-driving cars is perceived as such due to lack of control. If a plane is about to crash, there's not much to do besides hold and pray. If your car accelerates when it's supposed to brake, you better be paying attention and act fast to avoid a crash. With your hands on the wheel, you're always in a position to act, and it's easier to believe that there's always something you can do to avoid disaster.
Think about that for a second. How does that happen? Sounds like the truck was jack knifing, making an absurdly abrupt lane change, or the highway was not completed divided. Either way it's not like it was simply rear-ended, the car went under the side of the truck.
I'm not a Tesla apologist by any stretch but this sounds like a bit of a freak thing. My guess is the physical sensor wouldn't pick up something that high off the ground (keep in mind this was the side of the truck not the back where there is a bumper) and any camera sensors didn't have enough contrast. To prevent this they would have to have sensors point UP and forward, which I would guess they will be doing after this.
[1] https://www.teslamotors.com/blog/tragic-loss
http://electrek.co/2016/07/01/understanding-fatal-tesla-acci...
Even in the case of "mostly" autonomous vehicles where it could be argued that the human driver was lulled into complacency.
And it's unclear if the insurance industry would shoulder that like they do human errors. Serious accidents can bankrupt responsible people after insurance is exhausted, and those would all now be concentrated on manufacturers with much deeper pockets. Plus there's all the single-vehicle accidents which currently don't result in lawsuits, but would with autonomous vehicles.
- https://www.reddit.com/r/Showerthoughts/comments/4n9tbv/tesl...
EDIT: From the Tesla website:
> Every single Model S now rolling out of the factory includes a forward radar, 12 long range ultrasonic sensors positioned to sense 16 feet around the car in every direction at all speeds, a forward looking camera, and a high precision, digitally controlled electric assist braking system.
https://www.teslamotors.com/de_AT/blog/dual-motor-model-s-an...
I know my car has all the sensors mounted low. I'm not a radar expert. If the side of the truck was significantly above where the radar is looking would it be able to see it?
Apart from that, it is important to remember that fatal accident statistics are not just numbers, but are tabulations of individual tragedy. My condolences to his family, but sadly we can not expect autonomous cars to be perfect in the first try.
[0] https://www.theguardian.com/technology/2016/jun/30/tesla-aut...
[1] http://arstechnica.com/cars/2016/04/car-makers-cant-drive-th...
That said, all Model S are expensive and fairly new care with good safety features compared to the average car on the road today. Thus,I wonder if it is better to compare the 1 fatality / 130 million miles statistic to Model S (and S-class / 7-series) fatality rates while not on Autopilot mode.
You may be more likely to have an "accident" elsewhere.
But im reasonably sure that you are much more likely to die at highway speeds.
This is one of the great things about tesla logging everything tho.
The data is there to "prove" it one way of the other.
Two factors for that tho. Firstly is bad road markings. Which computers will be infinately better at handling.
But more importantly is the response time of emergency services and letting someone know theres been an accident.
Even minor injuries end up fatal if no one finds you trapped in the car for a week. And every second counts for more serious collisions.
Only so much manufacturers can do about either.
Dui and driver fatigue. As you will see are among the big killers. (And before you think "yeah but i dont drive drunk or tired" remember you arent the only person on the road).
Think its just about being carefull with what is meant by "safer" and "other types of driving".
You are more likely to have an accident in busy periods on busy roads.
But you are much more likely to die if you have an accident somewhere quiet or on high speed roads.
Your other points back up the claim though - self driving cars need to be compared to like-for-like mortality: same roads, same demographic, similar safety features. My guess is they're currently considerably less safe.
The obituary of the driver, Joshua Brown, is now available. Age 40. Former Navy explosive ordnance disposal specialist. Worked with Navy SEALs. Studied physics and computer science. CEO of Internet of Things startup.[2] It would be hard to find someone more qualified to operate a vehicle with Tesla's "autopilot". Placing the blame on that driver probably isn't going to fly.
The NHTSA is investigating.[3]
[1] https://news.ycombinator.com/item?id=12011419 [2] http://www.detroitnews.com/story/business/autos/2016/06/30/t... [3] http://www-odi.nhtsa.dot.gov/owners/RecentInvestigations
The key factor in whether this is a legitimate flaw in Tesla's system is: could an alert, competent driver easily avoid the accident?. An alternate formulation of this criteria is: would it be relatively easy for an engineer to hard-code a brute-force solution to the edge case, if they really wanted to?
Obviously we'll need to see the crash video to verify, but my guess is the answer is yes, this is a legitimate flaw in Tesla's system. Tesla's system should have been able to detect a car traveling perpendicular to the highway [2]. Tesla's system was unfortunately not sophisticated enough to detect a car that an alert human would be able to detect in their peripheral vision, and their system was not sophisticated enough to perform the simple action that a human would be able to perform (slamming on brakes). In fact, if they ever want to scale up to a full urban-capable system this is a situation they need to be able to handle.
Examples of accidents where I wouldn't blame Tesla's system: the other car swerves across the center divide and hits you head on, the other car is speeding and rear ends you, the car in the next lane swerves into you, the tire blows out and you hit a tree, etc. Although in the long-term we would ideally have systems that could handle these situations.
[1] We need to wait for more miles and (inevitable) accidents to be logged rather than rely on this one data point. Also my guess is highway driving is less fatality prone than urban driving, on a per-mile basis.
When I'm in traffic, I notice myself communicating with other people a lot. Most commonly it's just brief eye contact ("I have seen you"), but it enables smooth and safe traffic for everyone.
Just blindly following rules and reacting to your immediate surroundings won't work in large scale. There are plenty of situations where that would end up in a deadlock traffic jam (e.g. 4 way intersection with no right of way).
This is by no means an impossible task, I just have not heard of an effort trying to solve it yet.