Hacker News Reader

Comment by fabian2k | Hacker News Reader

> Waymo, formerly the self-driving car project of Google, said that in tests on roads in California last year, its cars went an average of nearly 5,600 miles before the driver had to take control from the computer to steer out of trouble. As of March, Uber was struggling to meet its target of 13 miles per “intervention” in Arizona

There can certainly be arguments about whether the kind of miles driven are equivalent here, but with more than two order of magnitude difference it hardly matters. I expected Uber to trail behind the top a bit, but I didn't expect that kind of huge difference.

This looked bad from the start for Uber, and it's only getting worse. I can understand that tests on public roads are necessary to actually achieve the goal of autonomous driving. But that step should come after sufficient tests in controlled environments that indicate that you got the basics down.

View on HN

Uber's competitive advantage has always been its willingness to engage in unethical behavior. No surprise here.

They were forced out if California because their system didn't work so they just put it on roads in Arizona instead and killed someone.

Edit: grammar

They weren't forced out, they just didn't want to comply because they would have to report incidents which would show how far behind competitors they are, which would kill their valuation

It's a semantic difference and exactly my point. They were effectively forced out of California because their technology doesn't work.

https://www.theguardian.com/technology/video/2016/dec/15/ube...

It's interesting how so many techies initially had the knee-jerk reaction of "well, despite this accident, I bet these Uber autonomous cars are still safer than humans".

It seems like all information available to us suggests the polar opposite.

After nearly getting hit by a car while on a crosswalk in broad daylight today, I don't think "polar opposite" is correct.

I actually made eye contact with the driver, and expected her to slow down (I was right in the middle of the street). Instead she sped up to get through the crosswalk ahead of me.

Instead of cars driven by assholes, we have to worry about cars programmed by assholes.

And given that a single asshole can drive only one car at a time, but can program a million cars, and given this is Uber we're talking about, I'm worried too.

I've had this happen to me so many times. It's insane that driving laws aren't strict in proportion to how much damage an irresponsible driver can inflict.

I don't know if that's better than what happened to me: she just screamed and threw her hands up while I jumped out of the way.

Neither is exactly confidence inducing, but I'm pretty sure the one who deliberately endangered another person's life to save a few seconds is worse.

I believe it about google. I don’t trust uber at all. Or Tesla for that matter.

Why do you believe that about Waymo?

We have no information at all about the progress Waymo is making. All we've seen so far are a few videos produced by for PR under highly controlled conditions.

There's basically zero transparency here.

If they are struggling to meet 13 miles per "intervention", how was the "safety driver" not paying full attention on the road at night? That's baffling.

They were averaging I think 200 miles per safety critical intervention. So, every mile the driver would have to take over to do something like go around a truck blocking the lane or construction or whatever, but the car would come to a stop safely.

Whoa. The car couldn't even pass a stopped vehicle on its own? That's terrifying. It's glorified cruise control.

I think uber is basically pretending to develop a self driving car.

> As of March, Uber was struggling to meet its target of 13 miles per “intervention” in Arizona

To be fair, Uber is not necessarily acting badly -- especially since AZ doesn't appear to care either way -- having the safety driver is ostensibly supposed to make the AV at least as safe as if it were a regular human-driven car. I hope during the investigation we find out whether Uber made it clear to its human drivers that they had to be at least as alert as if they were driving a normal car.

As a human driver, when do you know to take control of the vehicle and start stopping? At 100 feet? 200 feet? 300 feet (a football field?)

A quick Google[1] suggests the average stopping distance of a human traveling 35mph is 95 feet (including reaction time.) The promise of autonomous is 60 feet. If the human decides to stop at 60 feet, it's too late. If they stop before that, how do they know if the vehicle is working properly?

1. http://www.brakingdistances.com/35Mph

You don't need a complete stop to turn a fatality to a minor injury

If you see something and the software does not. It probably doesn't provide that sort of information to the driver though. Even then the driver probably isn't paying attention.

I suspect this is simply asking too much of human driver. The human brain is not wired to be constantly alert without any ongoing stimulus. If someone else is doing the cognitive load, it is extremely difficult not too get bored and 'drift off'.

The article says that Uber instructed their drivers to hover their hands over the wheel in order to be able to take over at a moments notice. But they also supplied an iPad app to report issues, that some safety drivers used while in motion (others were cited as using it at red lights)

The hovering hand thing weirded me the hell out. You'd get terrible RSI if you actually did that all day. Try it.

It is pretty much impossible to be "at least as alert as if they were driving..." When you aren't actually driving. And yes Uber is acting badly. Just because Arizona let them do it, didn't mean they have to do it. Of course, I believe Arizona also acted poorly.

I mean, in terms of "conditions", it doesn't really get much better than Arizona. That's why most cars drive there. By most metrics, it's actually easier there than in California.

Well, except for the other drivers on the road.

Horribly horrible drivers here.

It's a cliche that everyone thinks this about their own particular area, but it's true that people in different areas (sometimes within the same state) drive badly in very different ways. Just another factor that SDVs will have to test.

Yep, the rub is that this time of year we get people from all over bringing their "drive badly in very different ways" with them which makes for some interesting interactions on the roadways.

Summertime, not so bad -- nobody willingly comes to Phoenix when it's 117° except the conventioneers who don't have a lot of money to burn and get a discounted rate on hotels and the convention centers.

They seem about middle of the pack to above average. They're definitely not New Jersey:

https://fivethirtyeight.com/features/which-state-has-the-wor...

Also note: The number of miles per intervention isn't a metric of AI driving skill, but of how safe the company is being to other road users.

In science, when there's a two orders of magnitude difference, it's more indicative of distribution mismatch between the test sets. The logarithmic property of the learning curve suggests that, unless one approach is a completely random walk, two learning algorithms after a period of training must converge to their maximum capacity. As a scientist, I'm refused to be clouded by my prior judgement of the companies behind the approaches and must question the nature of the metrics and the various definitions that were used.

What are you smoking?

This supports my thesis waymo is building a self driving car but uber is only building a pretend self driving car (to fleece investors, who realize uber will go bankrupt if snd when anyoneother than uber achieves total success with self driving)

Google is playing chess, uber is playing checkers. Dont be fooled just because the board looks the same.

Reply on news.ycombinator.com