Tesla driver reproduces fatal autopilot accident [video] • Hacker News Board

Re: Tesla driver reproduces fatal autopilot accident [video]

#101

Post by PuffinBlue » Mon, Apr 02, 2018, 4:35 PM UTC

simion314 wrote: post #62

Shank wrote: post #30

Part of me wonders whether or not we're going to see mutual improvement of infrastructure with self driving cars in the future. This seems like something that you could obviously fix by adding in some kind of passive or active flag that, to a self driving car, says "woah, you're /really/ not supposed to be here." I know we're supposed to be focusing on building systems that are as capable as humans are as drivers. Ho…

We should also not allow software that hits a static object, the car should be able to detect solid object, should be able to detect it's speed relative to the car and to the ground, and it should detect a possible collision and take action. You can put electronics on the roads but if the cars will depend on those then it will fail when those electronics are missing.

Not sure why this point isn't being more widely discussed, especially as it is a known problem[0] of almost all 'auto-pilot' systems based on radar.

To me it would seem to be self-driving 101, namely - don't hit a stationary object.

If the challenge is so great that current technology can't overcome the difficulties of detecting stationary objects in the path of the car (not just those stationary objects to ignore like overhead gantries) then it's time to change the language around 'autonomous' driving to ensure drivers understand the limitations. Obviously this isn't happening well enough right now as people are dying because of it.

From a personal perspective, until Lidar becomes commonplace I think I would eschew the 'autonomous' modes offered in the current generation cars.

[0] https://www.wired.com/story/tesla-autopilot-why-crash-radar/

Re: Tesla driver reproduces fatal autopilot accident [video]

#102

Post by Geojim » Mon, Apr 02, 2018, 7:28 PM UTC

jstanley wrote: post #7

This (video) is another instance where the lanes split apart but the part in the middle has absolutely no markings to indicate that you shouldn't drive there. This (photograph) is what it should look like: http://s0.geograph.org.uk/geophotos/02/24/67/2246762_3a6ea81... Is this (bad line painting) common in the US? Edit: on a second viewing, I think it actually does have some chevrons, but they're very faint.

Forget the paint, how does the radar miss this??

Re: Tesla driver reproduces fatal autopilot accident [video]

#103

Post by yellowapple » Mon, Apr 02, 2018, 10:17 PM UTC

naikrovek wrote: post #88

simion314 wrote: post #71

Earlier quoted context omitted.

It is even worse since they have a good product without the hyped autopilot, when I read about Musk wanting to use only cameras then I was sure Tesla self driving is not safe.

Humans drive only with cameras. Vision provides all the information needed to drive a car safely. Maybe the software isn't there, yet, but it will be.

No we don't. Humans use sound for a lot of context clues (like "oh there's a motorcycle in my blind spot") and use their sense of orientation to gauge how hard they're turning (among other things).

Hopefully self-driving cars use a lot more than just cameras, too.

Re: Tesla driver reproduces fatal autopilot accident [video]

#104

Post by princekolt » Mon, Apr 02, 2018, 11:00 PM UTC

Geojim wrote: post #102

jstanley wrote: post #7

This (video) is another instance where the lanes split apart but the part in the middle has absolutely no markings to indicate that you shouldn't drive there. This (photograph) is what it should look like: http://s0.geograph.org.uk/geophotos/02/24/67/2246762_3a6ea81... Is this (bad line painting) common in the US? Edit: on a second viewing, I think it actually does have some chevrons, but they're very faint.

Forget the paint, how does the radar miss this??

AFAIK The radar is tuned to explicitly ignore non-moving objects because it has to be able to ignore overhead road signs and other such stuff[1]. It is more tuned to changes in speed of the vehicles in front of the car.

[1]: https://www.tesla.com/blog/upgrading-autopilot-seeing-world-...

Re: Tesla driver reproduces fatal autopilot accident [video]

#105

Post by princekolt » Mon, Apr 02, 2018, 11:04 PM UTC

trimbo wrote: post #45

gknoy wrote: post #11

That kind of lane marking (or lack thereof) seems very unsafe. I realize that often this is due to construction / etc but the immediately-to-the-left lane's marker was Very Clear, and didn't have a visible split in it. I had to watch it a second time to make sure that I didn't miss the indications that the lane was splitting away (other than the highway signs). This kind of thing (lack of clear markings) can be reall…

> the road engineers should have made sure the lines were painted correctly If this is the intersection, they were painted correctly: https://goo.gl/maps/QuuRiixk6Y42 But in this video, the left solid white line was more visible than the right one. For whatever reason -- a trick of light, snowplows tore it up this winter, who knows. That kind of thing is going to happen and autonomous cars need to deal with it in ord…

On a tangible subject: Why do so many US highways use such a light gray colored asphalt mixture? This certainly doesn't help on the contrast with the road markings.

Re: Tesla driver reproduces fatal autopilot accident [video]

#106

Post by dreamcompiler » Tue, Apr 03, 2018, 3:16 AM UTC

simion314 wrote: post #71

TAForObvReasons wrote: post #58

Earlier quoted context omitted.

Musk is following the tried and true tradition "Move fast and break things". Only in this case "things" include people's lives :/

It is even worse since they have a good product without the hyped autopilot, when I read about Musk wanting to use only cameras then I was sure Tesla self driving is not safe.

Agreed. I rented a Model S a few weeks ago and drove it around southern California. The autopilot was a nice gimmick, but I really only used it in the stop-and-go traffic jams SoCal is famous for. There it was a godsend.

The rest of the time, the pleasure of driving such a responsive machine made the autopilot superfluous.

Re: Tesla driver reproduces fatal autopilot accident [video]

#107

Post by manicdee » Tue, Apr 03, 2018, 3:57 AM UTC

princekolt wrote: post #104

Geojim wrote: post #102

Earlier quoted context omitted.

Forget the paint, how does the radar miss this??

AFAIK The radar is tuned to explicitly ignore non-moving objects because it has to be able to ignore overhead road signs and other such stuff[1]. It is more tuned to changes in speed of the vehicles in front of the car. [1]: https://www.tesla.com/blog/upgrading-autopilot-seeing-world-...

Tesla was working with their supplier (Bosch?) to get the radar to work as a six-field array, providing more resolution than “the nearest return is 100m away”. The discussion at the time was tht higher resolution radar would remove the last advantages that lidar supposedly has over cameras.

Re: Tesla driver reproduces fatal autopilot accident [video]

#108

Post by manicdee » Tue, Apr 03, 2018, 4:00 AM UTC

ImaCake wrote: post #55

Johnny555 wrote: post #5

Earlier quoted context omitted.

Does it matter? Do cars act differently depending on which state they are in? Seems like a similar situation to the crash in California.

The parent post doesn't ask that. It just states a fact. A useful one at that!

It is an irrelevant fact. Not useful at all, other than to asnwer the question that was not asked: “how common is this poor maintenance on US roads?”

Re: Tesla driver reproduces fatal autopilot accident [video]

#109

Post by manicdee » Tue, Apr 03, 2018, 4:08 AM UTC

Too wrote: post #67

Is it the driver or the car doing the final braking in the video? Even if autopilot followed the wrong line it should perform emergency braking or evasive maneuver when the barrier is right in front of it. This is not just one error, it's two.

This is the driver braking, you can hear the AP disengagement tone as the driver presses the brake pedal, and there is a distinct lack of AEB alert which should sound like this: https://youtu.be/FCegUNtgfuo

Re: Tesla driver reproduces fatal autopilot accident [video]

#110

Post by manicdee » Tue, Apr 03, 2018, 4:10 AM UTC

sudhirj wrote: post #42

friedman23 wrote: post #36

Earlier quoted context omitted.

> This puts the lie to Tesla and Mobileye's claims that camera-based computer vision is enough for autonomy. How are these things related? Are you suggesting that cameras are not sufficient for self driving?

The companies are stating that they won't be using LIDAR or Radar, and that cameras are enough. Cameras tend to rely on markers of some sort - and in cases like the markers on the road were very misleading. Even a short-sighted beginner human driver who was told to carefully follow the lanes would have made the same mistake. And cars running purely on cameras are comparably to short sighted baby human drivers.

Tesla uses radar, and are working on improving their radar to return point clouds just like lidar does.

Even then, cameras are enough for accurately mapping the world around the car, it’s just that Tesla’s state of the art has not matched the academic proofs of concept that have been demonstrated over the last few years.