Hacker News Reader

Comment by ars | Hacker News Reader

Direct link: https://twitter.com/TempePolice/status/976585098542833664

The interior view tells you everything you need to know about why anything except perfection in self driving cars will lead to more accidents.

The driver specially selected and trained for exactly this job, does not look at the road! If someone specially trained for that won't do it, the general public certainly will not.

Also, in the video you can not see the bicycle till it's almost too late. But I have better night vision than this camera, I feel as a human I would have seen the bike. (For example the sides are pitch black in the video, but in similar lighting I can see stuff to the side.)

But the real issue is that self driving cars should have better vision than a human, not worse!

View on HN

"The driver specially selected and trained for exactly this job, does not look at the road!"

Humans gonna human. Our brains are trained by millions of years to not pay attention to things not directly important. It's hard enough sometimes to keep our attention on the road when we're doing the driving and the driving is too easy, like being alone on a highway at night. Expecting humans to passively observe driving while making no decisions and taking no actions... well... could you watch eight straight hours of video footage from a car, tomorrow? I know I couldn't. I couldn't even pre-HN/Reddit/instant entertainment constantly in my pocket, let alone now.

The idea that sitting a human in an otherwise-automated car will do anything after the first few hours is absurd. It's a feel-good measure and nothing more. I wouldn't have expected any other result.

Although this video will no doubt throw that driver under the, um, bus it wouldn’t matter if this were some rock star Uber engineer. You could see the exact same result.

Yeah, there's that videogame about driving a bus constantly, on an empty road, for hours and hours. It occasionally veers off, so you have to pay attention, but that's all. That game should be a minimum "driving" test for safety drivers. (Not really, but something like that)

Desert Bus. That's not a terrible idea.

Desert Bus. Not quite Twitch stream material..

I agree that it's an apt comparison, except Uber seems a worse experience. In Desert Bus you have to correct the bus.. Stupid work, but something necessary - boring - to do or you fail.

There seems to be nothing at all to do in that Uber?

If you don't think that's twitch stream material, you haven't seen nothin yet.

ahem https://desertbus.org/

I .. was aware of that, but you literally pay for Schadenfreude and charity there. I sincerely believe that no streamer can make money or get a following with a schedule of Desert Bus runs.

Nor do I want to believe that there are people out there, bored out of their mind to click on anything, anything(!) on Twitch, would pick Desert Bus over literally any other thing for more than three minutes.

It's painful. People like seeing others suffer, I guess.. But these streams do serve a purpose in this thread: The poor players do look exhausted and ready to die after a short while already. Which brings me back to my original point: These Desert Bus gamers have something to do (correct the bus repeatedly) and in your example a whole community who engages them. The Uber driver ... has nothing at all?

This is why I've always criticized Tesla over its defense of Autopilot accidents, even when most people here were defending them. Then Waymo and Ford came out and confirmed what I and several others here were saying - that humans can't react to "take over the wheel" if something shows up in front of the car.

It's either Level 5, or nothing.

You could at least use the other levels for cases where the driver has a heart attack, etc.

With autopilot now, you never know when it will start flashing to wiggle the steering wheel, so it is not a hands off, get distracted with your phone type of thing anymore.

> anything except perfection in self driving cars will lead to more accidents.

Only if perfection is defined as "statistically better than humans".

I do agree that these "assistive" technologies have the potential to do more harm than good, but the bar is actually far lower than "perfection".

I wonder what the "deaths per vehicle miles" stats for autonomous vs human-piloted cars look like right now? Humans kill, what, 30,000 other humans a year or so in The US? But there's a _lot_ of vehicle miles driven. My guess is this one incident has likely pushed the autonomous car death rate per vehicle mile way beyond the current human piloted death rate per vehicle mile...

(Having said that, there's not much doubt in my mind that stuff like antilock brakes and autonomous emergency braking have reduced the road toll (and will reduce it further as these recent-ish technology become more common i the aged vehicle fleet) - I think the difference there is then those technologies are not nearly so inclined to make driver pay less attention to what's going on around them. I _hope_ people don't think "It's OK, I've got emergency braking - I'll just watch this tv show as I drive to work!")

This is the exact reason why we need Nationwide disclosures of automous miles driven. California requires this stat, so it's data is the most telling.

California .92 deaths per 100 million (human)vehicle miles traveled. There were only 500k autonomous miles driven in California last year.

I'd bet your guess about death rate per mile is right considering waymo which has largest and most advanced fleet only drove 2 million miles nationwide last year.

Does Tesla's autopilot count as autonomous? There has been one confirmed death in what is likely over half a billion miles of autopilot usage. That would seem to tip the rate back in favor of the autonomous cars over humans.

Autopilot does not count. I'm assuming it's because it requires people to be actively engaged in the task of driving and it only works in very specific situations. Tesla reported 0 miles of autonomous driving in CA on public roads last year.

it's pretty amazing nobody died in this tesla crash in January http://money.cnn.com/2018/01/23/technology/tesla-fire-truck-...

slammed into a parked fire truck going 65 miles per hour

the solid construction of the vehicles may be compensating for shortcomings in the driver assist features.

I think Waymo is claiming 5 million miles in total, and Uber 2 million miles.

[1] https://medium.com/waymo/waymo-reaches-5-million-self-driven... [2] https://www.forbes.com/sites/bizcarson/2017/12/22/ubers-self...

You can’t extrapolate off of one death. The 10th time you flip a coin, it could land on its side.

If one incident can reverse your conclusion then the sample size isn't large enough to draw that conclusion.

With only one death, the only conclusions you could reasonably reach from statistics are either "we don't know yet" or "we've gone so many miles with only one death that we know it's safe enough." There's no reasonable way to argue from the stats that we know it's not safe.

There have been studies that indicate that these sorts of technology (lane keeping/departure warning, auto-brake, blind spot detection) help reduce collisions.

Personally, I don't think that it's right to dismiss these systems entirely if they provide some level of safety benefit, are cheaper and easier than level 5 autonomous vehicles to develop and / or deploy.

> http://www.iihs.org/iihs/news/desktopnews/stay-within-the-li...

I wonder if there'd be a meaningful difference in accident rates when perception of a the cars capabilities change.

"My car will jump in if I fuck up"(current driver aides) seems a lot different than "I have to jump in if my car fucks up".

There was an excellent comment by curveship on the thread the other day [0] that really stuck with me so I want to parrot it a bit:

> Our car-based transportation system is far and away the most dangerous thing any of us accept doing on a daily basis ... everyone on the jury has in the back of their mind "that could have been me"

> drivers don't get punished for doing something dangerous, ... They get punished for doing something more dangerous than the norm.

> self-driving cars don't just have to be safer than human drivers to be free of liability, they need to be safe period. In a trial, they don't benefit from the default "could have been me" defense.

You are obviously correct that a "statistically better than humans" vehicle is an improvement over what we currently have. But this is the kind of technically correct answer that doesn't actually resonate with people, especially those outside of tech.

[0] https://news.ycombinator.com/reply?id=16624910

> I have better night vision than this camera, I feel as a human I would have seen the bike.

I'm not so sure. Under conditions where there is low light everywhere, so the eye has time to adapt, human low light vision is very good. But under the conditions shown in the video, where there are street lights present and the pedestrian comes out from a gap between the street lights, your eye won't be dark adapted and you won't have much of a chance of seeing the pedestrian in the shadows.

This is literally the reason for high-beams. If the road is not illuminated enough you turn those on and drive as fast as you are comfortable being able to fully stop if a deer jumps out of left field (or in this case, a pedestrian is walking across the road).

> This is literally the reason for high-beams.

I agree. From the video it doesn't look like they were on in this incident.

Is it allowed in the US to have high beams on when there are other cars on the road?

> Is it allowed in the US to have high beams on when there are other cars on the road?

You're supposed to shut them off if it would impair the vision of oncoming traffic. But there are plenty of situations where that won't be the case.

A self driving car should be far better than a human at turning them on and off.

In Australia it is relatively common in semi rural areas with similarly sparse street lights that you would turn high beams on and off depending on oncoming traffic.

And I have literally never had to use high beams on a road with street lights. In my opinion, the camera footage vastly under-indicates the amount of visibility available to the human eye between the lighted areas.

> I have literally never had to use high beams on a road with street lights.

I have, plenty of times. Not all road streetlights are optimally placed.

> the camera footage vastly under-indicates the amount of visibility available to the human eye

I agree with this; it's pretty much always going to be true for a visible light video camera under nighttime conditions. But I don't think that means a human driver, without high beams, would necessarily have been able to see the pedestrian. (As I've commented elsewhere in this thread, I think I would have been using high beams on this street, streetlights notwithstanding; the gaps between the lights look too long for good visibility.)

It's typically illegal[0] and otherwise a bad idea to operate high beams when there is other traffic around as you can blind other drivers.

[0]: Page 46 of https://www.azdot.gov/docs/default-source/mvd-forms-pubs/99-... Other states have similar laws, such as California: http://leginfo.legislature.ca.gov/faces/codes_displaySection...

It's not a function of seeing it the dark, it's that human eyeballs have really good dynamic range, which is exactly why they'd beat cameras in those conditions.

> human eyeballs have really good dynamic range

Not if they're not given time to dark adapt. As I've posted elsewhere in this thread, human vision would still be better than that of a video camera using visible light, but that still wouldn't be very good under variable lighting conditions.

There's a range of light to dark you can see at once. Your eyes can adapt and move the whole range up and down, as can cameras. You're right that it takes time. But the width of the range at a given moment is wider on people than many cameras (sans post-processing). That's why this video is deceptive.

Our eyes can see details in the light and dark at the same time better than cameras can. This (https://i.imgur.com/3u1kN7T.png) is an example. The top picture has less detail in both the lightest and darkest portions than the bottom one. The human eye is reeeeally good at that.

> Our eyes can see details in the light and dark at the same time better than cameras can.

Yes, I agree. But that doesn't mean that a human eye would necessarily have seen the pedestrian in this situation in time to react. (Based on other posts elsewhere in this thread, I don't think a human driver would have been able to see the pedestrian in time without high beam headlights on.)

is this what it looks like when you drive at night? Can you only see about ten feet in front of you, and everything else is pitch black? No. Not even close. Our eyesight is much, much better than this.

If our eyes were as bad as this camera, it wouldn’t be legal to drive at night.

What is displayed on the center console? Is it possible that the driver was looking at a live video feed? I believe that some standard cars show their night vision feed there.

I'm skeptical of that possibility, because if that were true, you would see the driver gasp in surprise while looking down (and a lot sooner BTW, given that it's supposedly night vision), instead of roughly one human-reaction-time after looking up at the road.

> Also, in the video you can not see the bicycle till it's almost too late. But I have better night vision than this camera, I feel as a human I would have seen the bike. (For example the sides are pitch black in the video, but in similar lighting I can see stuff to the side.)

While I'm happy for you, there are a lot of human drivers who don't have good night vision, have poor reaction time due to sleep deprivation, etc., etc., and get into accidents like this all the damn time. Is it a bad idea to have self-driving cars that are about as safe as the 10th percentile human driver? If the 10th percentile human driver would have hit this person, too, then we either need to keep these self-driving Ubers on the road, or revoke 10% of drivers licenses.

The article below predicted a scenario much like this one. It made a point that putting a computer in control that works really well on average will make humans reaction worse in general and especially if it's a weird, emergency situation happening right as their attention is kicking in.

https://www.theguardian.com/technology/2016/oct/11/crash-how...

>The interior view tells you everything you need to know about why anything except perfection in self driving cars will lead to more accidents

Ah, no, the cars only need to be better than humans, not perfect. This seems self evident, so did you mean to communicate something else?

Better than humans on average or better than the best human?

Thanks for the link. I'm now fully expecting for regulators to call for airbags on the exterior of all AV's...

The research has been done. We know how to improve outcomes of collisions with pedestrians.

Reply on news.ycombinator.com