Hacker News Reader

Comment by JoeAltmaier | Hacker News Reader

But must the standard for safety be higher than existing human drivers? We're already living the 'hindenburg' life - every day my freeway overhead sign shows me the fatality total this year for my state. Its almost exactly one per day. This is 'normal'.

What are the on-the-road stats for automatic driving? Aren't they better than human drivers already?

Its not a failing, just because they have different failure modes than humans. Because our human failings are so many and so frequent. There's so much low-hanging fruit. Auto-driving never gets tired, never argues with the passenger, never gets distracted by changing the playlist.

We're net ahead already with auto-driving.

View on HN

> But must the standard for safety be higher than existing human drivers?

This is my line of thought as well. Same problem I have with the anti-nuclear environmental crowd. Are there substantial drawbacks? Absolutely.

That doesn't matter. The important question is not "Does this new technology solve all previous problems?" but rather "Does this technology solve a net positive number of problems while offering more opportunities for improvement?"

With decades of empirical data that average humans are pretty bad at reliably manually piloting powerful, heavy vehicles, I can't see a reasonable reason to reject deploying autopilot technology in an aggressive manner. And that sadly means a few deaths. But those should be measured against the number that would have died were nothing done, not against zero.

Do you test the Autopilot feature? Wouldn't creating a much safer vehicle be real progress? How is it progress for a machine to decapitate you? What value is an Autopilot if it kills you?

> Do you test the Autopilot feature? Wouldn't creating a much safer vehicle be real progress?

What you don't understand is deploying a driving assistance feature is probably the fastest and most efficient way of improving the system because it is the fastest way to collect huge amounts of data. When used properly (As the machine reminds you each time you use it), it should always be a safer option than regular driving with the added benefit of providing valuable data to improve the system.

You should rephrase your question to: "What value is an Autopilot if it kills you when you use it in a way it explicitly told you not to?"

I can't think of too many recent ML breakthroughs that haven't largely come on the back of access to enormous data sets.

Interestingly, curious if that's why Tesla equipped cars with sensors before enabling autopilot as a user option. Hypothetically, can't see why they couldn't have downloaded the autopilot code to cars and been running it in ghost mode while users drove, generating faults whenever errors were encountered and sending the relevant sensor data back home.

There's no reason Tesla/MobilEye couldn't be using an off-policy reinforcement learning algorithm, no.

You assume a lot about my understanding.

Deep plowing was probably the fastest and most efficient way to farm the Great Plains because you could quickly and easily turn up huge tracts of fertile land.

Then the dust storms came.

What value is an Autopilot that knows you are not using it the way you are supposed to, but keeps accelerating you anyway?

> What value is an Autopilot that knows you are not using it the way you are supposed to, but keeps accelerating you anyway?

Good point. An autopilot, whose sole purpose is learning, should make sure the user is engaged so it's provided correcting feedback.

This kind of deeply emotional knee jerk response is why this discussion needs to be left to scientists and statisticians.

Hilarious. Asking questions (which you have not even bothered to answer) is an emotional knee jerk response...contrasted against science and statistics!

Seriously. You used the word you over and over. Yes, of course I don't want to be decapitated. That's irrelevant and an appeal to emotion. What is important is: what are the odds of getting decapitated while driving vs. autopilot.

I used the word three times. All in highly relevant questions.

>What is important is: what are the odds of getting decapitated while driving vs. autopilot.

Such odds are pretty much impossible to calculate. We know of one instance where a Tesla intentionally drove into a situation where decapitation was inevitable. It happens to manual drivers as well, but not typically without additional circumstances (high speeds, prior crash/accident). Usually if your head is about to impact something, you will react quickly enough to save it.

> Usually if your head is about to impact something, you will react quickly enough to save it.

That's is patently false since 30K people die in car collisions every year in the US alone. Stop focusing on the anecdote, the one death, and look at the numbers.

> I used the word three times. All in highly relevant questions.

They are all irrelevant because they focus on the one, the individual, the act, the emotional.

That's not how good policy is made.

I don't know where you are coming from, but the individual is the most important consideration of any policy.

>They are all irrelevant because they focus on the one, the individual, the act, the emotional.

I can parse this sentence, but I don't know what it means. The act, the emotional?

> That doesn't matter. The important question is not "Does this new technology solve all previous problems?" but rather "Does this technology solve a net positive number of problems while offering more opportunities for improvement?"

To play advocatus diaboli: Shouldn't it be sufficient for the autopilot to be as good as the worst human drivers that we still let on the road? Or to sharpen the point: As good as the worst licensed drivers under human supervision (compare learner's permit)?

After all the technology is still "learning", under human supervision. And similarly a teenager having a parent as co-pilot does not only endanger their own life but also those of others too.

So if those are allowed on the road with the expectation that they'll get better and that giving them a licence will eventually have net benefits, shouldn't that apply to self-driving cars too?

If there are more Teslas than worst human drivers, that would be a net negative. The threshold should be that a Tesla must be at least as good as the Xth percentile, based on the number of drivers replaced. Or higher. If autopilot is instead replacing the best drivers, that's also a net negative.

Long haul trucking on routes with less than 2000 feet at each end off of a limited access highway could be automated very soon. The reduction in sleepy human drivers on the road would probably increase highway safety while saving companies money.

How do you figure? I don't believe there's been any interstate testing of automated driving on freeways with human drivers, let alone a standardized legal framework that would enable end-to-end automation.

Have you seen logistics centers in Texas? They are often out in the middle of nowhere, often with dead-easy access to the highway. Automating driving in those conditions would be much easier.

If nothing else, one could have "pilots" guide the trucks by remote control for the last mile.

My point was that it's not technology that will be the hard part, it's legality and unions.

If there are no employees, what say will the unions have?

There are far from no employees with about 3.5 million truck drivers in the US. A company cannot operate an automated commercial service today because there is no licensing system for driverless trucks. The unions aren't going to allow legalization of automated truck driving with that many jobs on the line.

You're not providing any argument for the ease of legalization here.

The unions aren't going to allow legalization of automated truck driving with that many jobs on the line.

The unions can only have direct leverage with companies that currently employ truckers and those that work closely with those companies. All it takes is for one tech company that doesn't fit into the above category to develop the technology, then the market will make the licensing of such devices inevitable.

You thought you were giving a clever rejoinder. However, I was specifically talking about companies that don't directly employ large numbers of truck drivers. Then it comes down to the economic power of the companies that operate the trucks versus the economic power of the companies that have to pay the trucking companies. The latter is going to win the lobbying contest by sheer size.

Sounds like someone should reinvent the railroad...

With the advantage that the Interstate can be used for the long-haul infrastructure, without having to build it yourself.

This.

Many people are unwilling to cede control of their automobiles even if the statistics support the case of their car being safer when driven autonomously.

The "infinite trolley" problem puts this into better perspective, IMO: https://backchannel.com/reinventing-the-trolley-problem-85f3...

> What are the on-the-road stats for automatic driving? Aren't they better than human drivers already?

Probably not. The fatality rate for urban interstates in 2007 was 0.54 per 100 million miles. That's the type of environment--safe, predictable--where autopilot is most likely to be activated.

With one fatality, we don't have a really accurate picture of Tesla's fatality rate yet, but our best estimate at this point (0.76 per 100 million miles) is higher than that.

http://freakonomics.com/2010/01/29/the-irony-of-road-fear/

> Probably not.

Based on what?

> With one fatality, we don't have a really accurate picture of Tesla's fatality rate yet

Correct. You'd need at least 32 deaths. Until then, there is no "probably" one way or the other.

Deaths are also a poor metric, since car safety comes into the picture. You're better off using pure reported accidents instead.

The point of the parent was that the statistics are not better for auto pilot - they are either worse OR ambiguous/unknown, neither of which qualifies as the "better" of the gp.

> The point of the parent was that the statistics are not better for auto pilot

No, that was not his point. He said "Probably not."

That's wrong. It's not probably anything. Both the parent and gp are wrong.

And the negative tone of your comment is wrong too, since there is no evidence to suggest a problem with autopilot. Wait until the data is in. One dead dude and a couple of accidents doesn't mean a damn thing with millions of miles driven. It might be worse, it might be better. That's not an indictment, that's the case with anything new.

> Aren't they better than human drivers already?

I haven't drove the only system that calls itself autopilot, but some of the comparable systems from other manufacturers.

In my opinions the cars are significantly worse drivers than the average human. Their perception field is tiny to what a human can recognize (especially with 1 camera / 1 radar systems), they partly react later than a human can do (e.g. you see a traffic jam 700m ahead or other cars wanting to change the lane far ahead) and their lane holding is not really smooth.

They have one theoretical advantage at the moment: They don't get tired or distracted and make errors because of that. They might even have better reaction times in some situations.

However in the same way that humans can fail in 1000 ways the software can do that too. You could argue that humans won't be prone to buffer overflows, segfaults, deadlocks or other kinds of bugs. Most automotive software is written in C - which is here most often not even trusted for non-safety related jobs. And although the software is probably better reviewed and tested than the average web or consumer app there's still a lot of room for errors due to various reasons. And developing some completely new system with an insane kind of complexity (automated driving) won't make that easier.

They're nice assistants and reduce my commute fatigue considerably, but yes, when I use my Jeep's adaptive cruise control system I constantly have to be on the lookout for stopped traffic too far ahead for the system to properly recognize or things entering my lane from the side. My understanding is that this is common on all of the radar+camera systems, Tesla's included.

Life would have been a little better for them had they not bundled relatively common ACC tech with bleeding edge lane-management tech, called it all one thing (autopilot) and then said that thing was in beta. Basically they hyped almost-industry-standard ACC off the charts and are now getting special scrutiny for it.

On another note, saying your driver safety system is in beta testing is its own fail, just with the word "beta" that screams "still has bugs." They really need to watch out with the crossing over of tech culture idioms to vehicles. There were probably other ways to position that tech.

There are not enough Teslas on the road to know if their autopilot is safer or not. But if you want to make a premature conclusion, it would not be that they are safer.

Beyond that, tesla is not an autonomous vehicle. From what I can tell, it is not outfitted with the sensors to make it a viable autonomous vehicle. The danger is marketing it as an autonomous vehicle while having fine print saying that it is not.

Isn't it just enhanced cruise control?

It is, and this is a problem that's plauged auto safety advances. When ABS was introduced people started following so much closer that they more than made up for the stopping distance advantage. Now we have advanced cruise control systems, and some people more than make up for the safety advance by not watching where they are going.

It is a little more than that. It is enhanced cruise control, lane departure warning system, and auto parking system all rolled into one. The new part of the system is that instead of just warning the driver about lane departure it will automatically steer back into the lane, and if the driver puts on the turn signal it will automatically change lanes.

Honestly it seems like the big deal is that it is Tesla and how they marketed the feature.

> But must the standard for safety be higher than existing human drivers?

Well, if you want people to use the tech you need to be able to tell them that the car is at least as good as they are and in reality better than they think they are.

That's no really the problem though. The problem is the current generation of automatic driving is not automatic at all. It is assisted driving - i.e. an evolved cruise control - and the problem with it is that it requires inhuman circumstance to work alright ( driver being alert despite not having to do anything ).

There is case on what amount of assistance is actually too much to avoid that "death valley".

Automatic systems are another problem entirely. If a human intervention is necessary, automatic system need to be designed for a human not ready.

The currently assisted driving system in the Tesla is used as an automatic driving system by misguided drivers. The risk is that this mistake is going to hurt public goodwill in adopting real automatic system when they are available.

> But must the standard for safety be higher than existing human drivers?

> We're net ahead already with auto-driving.

We don't really have enough data to know that. Really, we haven't even scratched the surface of figuring out if that is true. E.g. we have no idea what kind of emergent phenomena will arise when roads full of self-driving cars interact with each other.

History is littered with examples of engineers thinking they were 90% of the way there when they were actually like 10% of the way there.

Would be really interesting to have some kind of race track filled with self driving cars of different manufacturers running 24/7.

I believe that could give them alot of data to analyze.

You could even make some kind of tournament out of it...

And a cool movie! Add betting and deathraces.

Maybe not exactly what you had in mind, but this is going to happen. It looks like it hasn't started this year as they planned.

* http://fiaformulae.com/en/news/2015/november/formula-e-kinet...

* http://www.roborace.com/

Formula E isn't well regarded in racing fan circles - I personally try to catch the races when I can but I don't have cable and hardly ever watch so I don't know the state of the racing this year.

In practicality, everything you said is correct. Politically, however, the burden of proof is on the new technology as if the old technology doesn't exist or is completely safe. The old technology has the political support of anyone who makes a living driving, despite the dangers and drain on public finances. That's the hurdle we need to get over and it might just need to take until this protected class is replaced with the new technology via attrition.

"But must the standard for safety be higher than existing human drivers?"

Well, being a different sort of entity, an auto-pilot will never be the same as a human driver. So "as good as" has to essentially imply "better than".

"What are the on-the-road stats for automatic driving? Aren't they better than human drivers already?"

Apparently not as enoch_r points out.

Further, there's miles-driven and quality of miles. Suppose, just suppose, the average driver starts out barely trusting Tesla and using auto-pilot on short runs in clearly safe conditions and this same average driver over-time trusts Tesla more and more until they reach the point that they trust in situations they shouldn't - and then you get a series of accidents.

>But must the standard for safety be higher than existing human drivers

Yes. Imagine that it crashes with the same average incidence rate as humans. Humans are imperfect creatures. They break the law. Let's even say that only 1% of crashes involve someone breaking the law (intoxicated, no licence etc). If the car is only on par with humans, then the car is only able to match a law abiding citizen 99% of the time.

You could argue that the car is basically driving buzzed. We have laws that would arrest 1 out of every 100 of those humans, but what would we do to the algorithm? What sort of punishment could you do to it when it messes up?

> Let's even say that only 1% of crashes involve someone breaking the law (intoxicated, no licence etc).

Speeding? Let's then raise that 1% to like 40% or something. Looking at some random sources[0][1], all the most common causes of traffic fatalities are the ones machines already deal better with than humans. Distracted driving? Check. Failure to stay in lane? Check. Speeding? Check. People die on the roads mostly because drivers behave like idiots or hit their natural limitations.

[0] - http://www.businessinsider.com/the-cause-of-the-most-fatal-c...

[1] - http://www.sixwise.com/newsletters/05/07/20/the-6-most-commo...

Interestingly, despite the graphic nature of the fire, more than 60% of the people on the Hindenberg survived. It was not as huge a disaster as it appeared: https://en.wikipedia.org/wiki/Hindenburg_disaster#Death_toll

> But must the standard for safety be higher than existing human drivers?

Yes it must. I can readily accept responsibility if there was something on the road to which I wasn't able to respond properly and resulted in an accident but if it was autopilot that screwed up I would sue $brand's ass.

> But must the standard for safety be higher than existing human drivers?

You better believe it. If I sell you a climbing rope and it fails, saying "yeh, but the average human wouldn't have been able to hold you as you came over that ledge" is not a suitable excuse.

But as long as the strength rating of the rope was made very clear to me prior to buying it, what's the problem? If the strength rating was still higher than any other viable alternative, I'd probably be very happy to buy that rope.

All the way agreed. Plus, in the last incident, the autopilot explicitly told the driver to take the wheel, which he did not do. So the accident is at least still in part human error. I expect incidents to go down even more once humans are removed from the equation entirely.

That future can't come fast enough for me.

> can't come fast enough for me.

It'll be rather amusing to see the older generations chomping at the bit for the introduction (nay, making mandatory!) of revolutionary new tech.

The youngsters mostly seem rather apathetic towards driving, but there must be a bunch of about-middle-aged drivers who'll resent having their toys taken away (or the ability to use them on public roads, same difference).

I am quite literally one of those "about-middle-aged drivers who will resent having their toys taken away". However, I'll get over it. I'm (begrudgingly) willing to give up my enjoyable hobby for the safety of millions.

Reply on news.ycombinator.com