https://techcrunch.com/2017/11/27/waymo-racks-up-4-million-s...
Nevermind that the initial rollouts almost always involve locales in which driving conditions are relatively safe (e.g. nice weather and infrastructure). Nevermind that the vast majority of autonomous testing so far has involved the presence of a human driver, and that there have been hundreds of disengagements:
https://www.theverge.com/2018/1/31/16956902/california-dmv-s...
Humans may be terrible at driving, but the stats are far from being in favor of autonomous vehicles. Which is to be expected given the early stage of the tech.
That's fine, we don't have to consider just fatal accidents:
https://www.theverge.com/2018/1/31/16956902/california-dmv-s...
> One result of the sharp increase in GM’s number of miles driven is a plethora of accidents. The auto giant’s autonomous cars were involved in 22 fender benders over the course of the reporting period (and two more in 2018). That’s one crash for every 5,985 miles of testing.
I didn't say that.
I'm not saying autonomous cars are safer, I'm saying that this accident doesn't prove autonomous cars are less safe.
> That you pulled out the IIHS stats without apparently bothering to look up the number of reported self-driving miles driven seems to suggest that you assumed that the fatality rate was obviously skewed against human driving.
I brought up the IIHS stats to show that merely reporting an accident in isolation from a full statistical model doesn't prove anything.
> That's a very bad look; the whole point of self-driving cars is that they can react to unexpected circumstances much more quickly than a human operator, such as when someone walks out into the road. Sounds like Uber's platform may not be up to that standard yet, which makes me wonder why they're on public roads.
> On the other hand, it sounds like it happened very recently; I guess we'll have to wait and see what happened.
How in the world do you see that as an assertion of proof.
The fact that human drivers aren't particularly good isn't really relevant, beyond setting a correspondingly low bar within the null hypothesis.
This all ties to regulation allowing these vehicles to drive on public roads, because that regulation was permitted due to the above hypothesis and hopeful expectations that it would be true.
Obviously, I haven't seen the entirety of the data set to know fatalities per car-mile. Which would be the relevant statistic here. I also didn't see such a number in your post, which I'm assuming means you are probably not aware either. But simply providing the numbers for the null hypothesis doesn't do anything.
I'd like to note that I am not the person you replied to, and I personally am not arguing that any program should be shut down based on this one incident. But it's certainly not encouraging that the autonomous vehicle seems to have failed a test at which everyone would have expected it to perform well at.
Also, to quote from your previous post:
> Why is it that you see one death from an autonomous car and conclude that autonomous cars aren't ready to be driving, but you see 37,461 deaths from human drivers and don't conclude that humans aren't ready to be driving?
I think we conclude, quite often in fact, that individual humans aren't fit to be driving. Death is one of those scenarios that will quickly lead to such a conclusion.
One huge difference between individual humans and autonomous vehicles is that we can reasonably argue that any Uber vehicle would have performed the same in this scenario. So this is perhaps more akin to saying that this particular driver is not fit for the task, except that this particular driver happens to be driving dozens or more vehicles all at once.
I said, "I admit that there just aren't enough autonomous cars on the road to prove conclusively that autonomous cars are safer than human-operated cars at this point." and I've vocally criticized autonomous cars elsewhere, so I'm not sure where you get the idea that I favor a blanket acceptance of autonomous vehicles.
I'm not saying that autonomous vehicles are safer, I'm saying this isn't evidence that autonomous vehicles are less safe.
> One huge difference between individual humans and autonomous vehicles is that we can reasonably argue that any Uber vehicle would have performed the same in this scenario.
I disagree: you can argue that the autonomous vehicles will behave the same given the same inputs, but they will never have exactly the same inputs even if the situation were identical to a human observer, so that's a fairly moot point. If you step back to a larger description of the situation (car crossing bike lane to get into a turn lane, car doesn't identify and avoid bicyclist in bike lane) then you are going to be looking at a percentage of the time where autonomous car will make a mistake. There's also a percentage of the time where a human driver will make the same mistake. The only way you can compare the safety autonomous cars to human drivers in this situation is to compare those percentages. And that's ignoring the fact that there are thousands of other situations in driving--even if autonomous cars fail 100% of the time in this situation, there may be enough other situations where they perform better enough than human drivers that they're safer. Simply saying that a car made a mistake in this situation doesn't give us any information at all.
> And that's ignoring the fact that there are thousands of other situations in driving--even if autonomous cars fail 100% of the time in this situation, there may be enough other situations where they perform better enough than human drivers that they're safer. Simply saying that a car made a mistake in this situation doesn't give us any information at all.
But it does give us information. This incident, along with all the other incidents and non-incidents that do or do not occur, as measured by incidents per car-mile. One cannot simply wish away this one incident and make it disappear. It is now forever part of the statistics which will either prove or disprove the hypothesis of autonomous cars being safer.
> I disagree: you can argue that the autonomous vehicles will behave the same given the same inputs, but they will never have exactly the same inputs even if the situation were identical to a human observer, so that's a fairly moot point.
It's not a moot point. Yes, this exact scenario with these exact parameters only occurred once. However, we can still reasonably argue that if we reversed time and replaced that exact Uber autonomous driver with another instance of the autonomous Uber driver and turned time back on, it would have reacted exactly the same. The same way that I expect the same version of Notepad to open my text file exactly the same on this computer as on another computer. The alternative being that there is some undeterministic behavior in the driver that is not tied to input... In which case, good luck with that in court.
However, we cannot make that same argument by replacing human drivers. Because each human is, in fact, different.
This is only important in the context that a fanciful revocation of Uber's "autonomous driver's license" would apply to all instances of the autonomous driver, since they would all have been reasonably expected to perform the same.
Can you point out the part of the article or the post that I was responding to which mentions how many car-miles were traveled?
This is exactly what I'm pointing out.
> The same way that I expect the same version of Notepad to open my text file exactly the same on this computer as on another computer.
Notepad doesn't have to read your text file through a lens with slightly different focus, viewing area, and patterns of dust on it each time.
> The alternative being that there is some undeterministic behavior in the driver that is not tied to input...
The undeterministic behavior is that the hardware which collects the input will never be the same. I don't know whether the software is non-deterministic (it wouldn't surprise me) but I know the hardware is never going to be identical--hardware is always made to tolerances and always has some degree of variability.
Your claim is tantamount to saying that if we put the same person in the same situation but with two different sets of eyes, the eyes would have no effect on the results.
> However, we cannot make that same argument by replacing human drivers. Because each human is, in fact, different.
Autonomous cars are, in fact, different. Just because they're running the same software doesn't mean they're the same; even if the software is completely deterministic, software is only a component of the autonomous driver.
> This is exactly what I'm pointing out.
I think, if your intent is to show that this is incomplete information, that...
1) No one is arguing that.
2) You have not done a great job of attempting to relay that, given phrases like, "Simply saying that a car made a mistake in this situation doesn't give us any information at all."
3) Sometimes that doesn't matter. For instance, Florida law is a mandatory 6 months to one year license revocation on DUI, regardless of the circumstances or information.
> Notepad doesn't have to read your text file through a lens with slightly different focus, viewing area, and patterns of dust on it each time.
Difficulty of the task is unrelated to expected outcomes of the task given the same inputs. And we already tread the topic of duplicating the exact situation... Not sure what you're trying to gain through this line of argument.
> Your claim is tantamount to saying that if we put the same person in the same situation but with two different sets of eyes, the eyes would have no effect on the results.
I am making no such claim, and I cannot believe that you are so adamant about not understanding my actual claim. This is a hypothetical situation. There is no mention in this scenario about changing the car, including any of the sensor hardware. I am interested in replacing only the driver (or driver software) into the exact same circumstance.
(EDIT: OK, reading back, I did say "any Uber vehicle". While your point stands, I think it's a very uncharitable reading. If hardware sensor tolerances and specs of dust on the camera are going to determine whether a life is lost or not, either those tolerances need to be driven down or this entire idea needs to be rethought. After all, we don't allow those who are legally blind to drive unless they have corrective lenses...)
Assuming the software is deterministic [1], by definition given the same inputs it will result in the same output. Therefore, "replacing" the autonomous driver with another would have resulted in the same incident. You cannot say that with any measure of confidence for any two pairs of human drivers.
[1] Which seems like it would be a good thing to assume, since I don't think one would get much traction by arguing that we should be putting vehicles with non-deterministic behavior on the road...
Are you making accusations in question form so you don't have to back them up? I certainly didn't say that.
> Why can't they fabricate some tests that demonstrate their safety in a controlled environment first?
Are you assuming they haven't done this?
> I'm not saying that autonomous vehicles are safer, I'm saying this isn't evidence that autonomous vehicles are less safe.
It's true that we need to wait for more information about this particular incident -- which is exactly the caveat that the commenter you initially responded to had said [0]. But assuming the facts aren't abnormally different than what they seem to be -- an Uber AV hit and killed a jaywalker, how is that not evidence toward the argument that AVs are less safe? It obviously isn't conclusive evidence. But if Uber goes on to kill a pedestrian for every million miles driven, this first data point would surely be part of the empirical evidence, no?
Nope, I'm just having trouble grokking your argument and I thought that might be it. It's true I added a rhetorical edge to the language that was probably unnecessary. I apologize for that--I'm not trying to put words in your mouth. What is your argument?
https://www.newscientist.com/article/2095740-tesla-driver-di...
But that's not really my point. My point isn't that self-driving cars are safer, it's that merely looking at one car accident doesn't prove that self-driving cars are less safe.
https://www.forbes.com/sites/bizcarson/2017/12/22/ubers-self...
so, 6 million miles driven and one fatality is significantly worse record than human drivers. So far.
This metric was often decried because it had poor statistical significance. It would however be nice to update this metric according to this death.
Maybe now this metric would indicate that IA are more dangerous than human, will be interesting if the perception of this flawed metric will evolve in reaction...
I'm curious to see that, do you have a link? That would actually be a valid comparison.
This is Elon musk bragging before the first autopilot accident : https://www.telegraph.co.uk/technology/2016/04/25/elon-musk-...
This is a counterpoint with some fact: http://safer-america.com/safe-self-driving-cars/
And concerning the crash, last update seemed to point driver’s fault... with autopilot... don’t know if final ruling is already out:
https://www.washingtonpost.com/news/the-switch/wp/2017/06/20...
If self-driving cars are between dumb drivers and normal drivers, and self-driving cars take dumb drivers off the road, it may be a net positive.
Let's say we have 9 normal drivers and 1 dumb driver. The 9 normal drivers cause 0 accidents per year each, and the 1 dumb driver causes 15 accidents per year. Let's say autonomous cars cause 1 accident per year. If autonomous cars replace all the drivers, there's 10 accidents per year instead of 15.
Obviously these are hypothetical numbers. All I'm saying is that if you only have one of these numbers, you don't know which kind of driver is safer.
I'm not sure I see anyone here making that conclusion. I think you're the only one who's brought it up.
For one, I personally sure as fuck do not think humans are ready to be driving.
> That's an invalid conclusion to draw from this accident.
The conclusion was not invalid at all. Other self-driving car companies have driven more miles than Uber has and they have done so safely. Uber has even taken its cars off the road, so even Uber agrees that their self-driving cars are not ready for the roads yet.
It is also important to take into account what Uber is like when it comes to safety and responsibility. They have knowingly hired convicted rapists for drivers, they have ridiculed and mocked and attempted to discredit their paying customers who have been raped by their employees/drivers, they have spied on journalists, they have completely disregarded human safety on numerous accounts. A company with a track record like Uber's probably not should be granted a license to experiment with technology like these self-driving cars on public roads.
They just aren't responsible enough.
Look at the post I'm responding to, and the section I quoted.
> Uber has even taken its cars off the road, so even Uber agrees that their self-driving cars are not ready for the roads yet.
Uber avoiding a PR nightmare should not be taken as Uber thinking their self-driving cars aren't ready.
> It is also important to take into account what Uber is like when it comes to safety and responsibility. They have knowingly hired convicted rapists for drivers, they have ridiculed and mocked and attempted to discredit their paying customers who have been raped by their employees/drivers, they have spied on journalists, they have completely disregarded human safety on numerous accounts. A company with a track record like Uber's probably not should be granted a license to experiment with technology like these self-driving cars on public roads.
I'm not defending Uber as a whole. I think in general they're a fairly typical rent-seeking middle man whose only real innovation has been figuring out a way to break the law ambiguously enough to get away with it (admittedly, a law I don't agree with). I don't even necessarily think autonomous vehicles are a good thing. I just think that if you're going to criticize autonomous vehicles, safety isn't the criticism I'd level against them, because it's not backed up by the evidence when compared to how unsafe human drivers are.
That's an invalid conclusion to draw from this accident.
There were 34,439 fatal motor vehicle crashes in the United States in 2016 in which 37,461 deaths occurred. This resulted in 11.6 deaths per 100,000 people and 1.16 deaths per 100 million miles traveled.[1] As far as I know, the instance this article is about is the first death in an autonomous vehicle accident, meaning that all the 2016 accidents were humans driving.
Why is it that you see one death from an autonomous car and conclude that autonomous cars aren't ready to be driving, but you see 37,461 deaths from human drivers and don't conclude that humans aren't ready to be driving?
I admit that there just aren't enough autonomous cars on the road to prove conclusively that autonomous cars are safer than human-operated cars at this point. But there's absolutely no statistical evidence I know of that indicates the opposite.
[1] http://www.iihs.org/iihs/topics/t/general-statistics/fatalit...
EDIT: Let's be clear here: I'm not saying autonomous cars are safer than human drivers. I'm saying that one can't simply look at one death caused by an autonomous car and conclude that autonomous cars are less safe.