Comment by tristanj | Hacker News Reader

Comment by tristanj | Hacker News Reader

Parent Full thread

Back when this originally came out, Uber claimed "this incident was due to human error" [1]. Well technically, if you really twist the meaning of words, Uber could interpret this as "human error", considering:

- It was "human error" that the programmers who designed the self-driving AI failed to properly implement red-light detection and braking.

- It was also "human error" that the human driver in the front seat failed to notice the red lights and stop the car.

Uber's statement is effectively true, if you hideously twist the meaning of words. I fully believe that's what they did in their statement. There's similar wordplay for the word "natural", i.e. claiming "All pollution is natural", because humans are part of nature and everything we do is natural, so all consequences of our actions are also "natural". Deceiving yet ultimately, effective.

[1] Uber says self-driving car ran red light due to “human error” https://techcrunch.com/2016/12/14/uber-looking-into-incident...

The full article from the NYT[0] which is the source for the information presented in the Verge post includes a quote indicating that Uber blames the operator for not manually stopping the vehicle. Their story is essentially that the car isn't ready to handle this scenario on its own and the operator should have known to intervene.

[0] https://mobile.nytimes.com/2017/02/24/technology/anthony-lev...

Which could (and will) be said for any autonomous accident in the past and future.

The NTSB takes a similar view: autopilot failures that result in aircraft crashes are still human error, since the human should have corrected. Still not very encouraging, since the driver in this case would have had seconds at most to identify and fix the error with a panic stop.

I don't think that's really a fair comparison if we're taking Uber at their word since the car is still a dev unit.

It seems like a known problem, cars would not stop at any red light period and the driver just forgot to stop because he wasn't paying attention.

They also track disengages as a metric for progress right? So it could be that these drivers will if at all possible avoid taking over.

> who designed the self-driving AI failed to properly implement red-light detection and braking.

From the article it sort-of seems like that error is actually in the mapping software that didn't mark that intersection as having traffic lights, so the car didn't look for them.

Shouldn't a self-driving car always be passively looking for red lights?

I mean, take a school bus for instance. They don't have permanent red-light fixtures, but if you're behind them they could enable the red light sign at any moment. If the self-driving car isn't always looking wouldn't that be an issue?

Even normal stoplights can be added or removed. If a car needs an accurate internal map of every single stoplight, it's doomed.

What about the case the stoplight stops functioning? This happens many times. Recently late at night in Flushing NY crew was working on replacing switches on the stoplight and that caused outage so the Department of Transportation actually have asked the police department to send out officers to help direct traffic. Yeah, good luck with that if we are running self-driving car.

Why? Two things could happen here: The self-driving car could recognize hand gestures. Or the police could carry a simple radio transmitter device which would transmit directions to the self-driving cars (while also making gestures for human-driven cars). This is hardly some insurmountable problem.

Self-driving car needs to adapt to the world; if police need a special device to communicate with the car, the car is unsafe and should be banned from public roads.

Fortunately, recognizing hand gestures can be done.

So the enormous benefits of self-driving cars (many saved lives, mobility for people who cannot drive, reduced costs) should be prevented because we cannot equip police with a few radio transmitters? -- Assuming that is even needed, because Waymo's cars can recognize gestures made by cyclists already, so it's hardly a stretch to imagine them recognizing gestures made by policemen.

It's not just a matter of equipping police. Police are not the only people who occasionally need to make gestures that need to be recognized by cars: think construction workers in a crowded city, parking lot attendants, or anyone at the scene of an accident.

The car needs to be able to communicate with people outside of it, or it needs to have a human inside who can drive it.

Because I don't want my car stopping when a panhandler/hitchhiker/etc is trying to wave me down, this is actually a much more complicated problem than simply recognizing gestures.

It think it's more pragmatic than that. If Kansas doesn't want to equip their police with transmitters, then transmitters simply won't happen. It's a political problem which is much thornier than technical problems. I'm sure you could already apply that rubric today : If only government agency X would implement simple technology Y, everyone's life would be easier. In the real world, that fails to happen all the time.

Self driving cars are probably just going to have to recognize hand signals. That's simply the reality.

I don't think the transmitter option is viable until there's a critical mass of self-driving cars that all agree on the same communications protocol. And, as we all know, the great thing about standards is that there are so many to choose from.

I remember reading that Google's car do recognize hand gestures.

That is really interesting if true. Do you have any idea where you read it? I am genuinely interested; this is not some backhanded way of saying "citation needed."

A bit of web searching goes a long way:

http://search.lycos.com/web/?q=google+car+recognize+hand+ges...

http://www.webcrawler.com/serp?q=google+cars+recognize+hand+...

https://duckduckgo.com/?q=Google%27s+car+do+recognize+hand+g...

https://encrypted.google.com/search?hl=en&q=Google%27s%20car...

I don't remember where I read it, but this is similar http://fortune.com/2016/07/06/google-self-driving-cars-cycli...

That's the current status of self driving cars.

And yes, I think they are doomed. At least for city streets - I expect them on specially instrumented freeways though.

> That's the current status of self driving cars.

Do you work in the field? I'm a bit skeptical about that. Waymo's self-driving car seems to do pretty well with things that aren't encoded in maps like road works and unexpected obstacles. I don't see why red lights would be exceptionally challenging to detect.

Tesla's demo video of their fully self driving software looks pretty good too.

A talk about this topic: https://www.youtube.com/watch?v=k5E28fp4oc0

Delphi actively looks for traffic lights at every intersection with their self-driving technology, or at least they used to.

There are also temporary traffic lights added and removed during road works.

There are also pranksters that turn traffic lights by 90° to make them face the wrong direction.

Those are hard to map.

To anyone with a moderate driving experience, this should be obvious

This seems to be a decision on par with "depend on left-pad". reliability wise

Forget if you're behind a school bus as I'd hope it would stop for any vehicle stopping in front of it. If you're on a normal rural highway (i.e. one with two but oppositely travelling lanes) and a school bus which is coming towards you makes a stop, you must stop so that any exiting kids can safely cross the street to their homes. What happens when a school bus makes a stop and the car decides to keep rolling through the bus's displayed and flashing red stop sign?

Fairly sure (though it likely depends on geographic location laws) that no matter your lane you need to stop if a bus displays its stop sign and lights, because, like you said, kids might try crossing the street at any time/place. So a self-driving car might be a lane or two away from the bus (not directly behind it) and still need to know to stop.

The exact rules do depend on the area. Where I am you always have to stop behind a bus on your side in any lane, but traffic on the other side does not have to stop if there are 4+ lanes.

Much of the material I studied for my learners permit written test has since become just part of how I drive (I hope), but the rules regarding stopping for school-busses have always stuck with me for some reason.

In Florida, the rule is that all traffic going in either direction must stop, except when the road is "a divided highway with an unpaved space of at least 5 feet, a raised median, or a physical barrier"[0], in which case the opposing traffic does not have to stop. It's interesting to me to see how little things like this change from place to place, and how these localized rules will be handled by self-driving vehicles.

[0]http://www.leg.state.fl.us/Statutes/index.cfm?App_mode=Displ...

Uber's statement is effectively true, if you hideously twist the meaning of words.

"Who are you going to believe, me or your lying eyes?"

The test driver failed to intervene, that's all. This is a non-issue, obviously Uber's autonomous vehicles are a work in progress.

Yeah don't beta test your BETA on real drivers, even after the DMV has asked you to stop.

Screw Uber. I hope the CA government comes down hard on them. But we all know Uber will find yet another slimy way to keep going.

For the record, I'm the biggest proponent of autonomous cars and can't wait for that day. But companies that just put people in danger like this are not going to get us there faster. They're going to slow everything down by their carelessness and greed.

> But we all know Uber will find yet another slimy way to keep going.

Like moving the program to Arizona...

And Nevada, with stolen Google engineer!

I don't like Uber either, and I don't know what they were thinking testing in California without a permit, but they were otherwise doing the same thing all of the 23 companies with testing permits in California are doing. (Any fly-by-night operation can get an av testing permit in California so long as they meet some basic requirements. It costs $140 and the approval process takes 3 business days).

The AI is expected to fail occasionally, so a human driver is onboard ready to intervene the instant that happens. The test driver wasn't paying attention, and the car ran a red light. It was a human error. Uber did not lie, they didn't mince words or twist the definition of things, the human driver just didn't do his job.

Irrespective of all the other bullshit going on with Uber, this is a mistake that any test driver with any other company testing autonomous prototypes on public roads could have made.

Just like with Tesla's Autopilot, any autonomous test vehicle is technically a level 2 vehicle, meaning the human driver is legally responsible for any traffic violation the vehicle makes.

Irrespective of all the other bullshit going on with Uber, this is a mistake that any test driver with any other company testing autonomous prototypes on public roads could have made.

They could have done, but they haven't. At best that means Uber are just unlucky; at worst it means that Uber are far behind every other self-driving car company and Uber's software just doesn't work for really simple, basic traffic laws. That should be very worrying for any investors.

Seriously. A red light? That is the most basic and most dangerous thing to ignore.

Except that this car is out there driving alongside people. Running red lights could get somebody killed. That's why google took forever to actually take their prototype to the streets.

What? Google was on California's roads in 2009 before anybody, least of all the DMV, knew what they were doing.

No, Google had a permit from the DMV when they started testing. Can't find the reference now, but I remember how they'd had to take DMV people in the car to convince them that they were serious.

'The Google researchers said they had carefully examined California’s motor vehicle regulations and determined that because a human driver can override any error, the experimental cars are legal' [1]

That's from a 2010 NYT article by John Markoff. They ultimately hired Ron Medford, former NHTSA director to lead a big Johnny Appleseed awareness raising campaign with regulators, but initially they just went and started testing.

[1] https://mobile.nytimes.com/2010/10/10/science/10google.html?...

If it's a non-issue, why'd they lie about it?

It's Uber. It's in their DNA.

It's fine that the car is a work in progress and it's also fine that they'll be tested on the streets before being fully baked.

What's not fine is insisting you don't need a permit (the primary purpose of which appears to be transparency regarding safety), having a safety incident, and then implying in a public statement that it wasn't related to a software failure when it was.

I don't think the issue is whether it stopped or not, but the possible discrepancy in how it was communicated.

Reply on news.ycombinator.com