Hacker News Reader

Comment by Robotbeat | Hacker News Reader

I think the answer is that the path into the barrier looks just like an actual lane, and that this was enough confusion that a human driver apparently had made the same exact mistake a week earlier. https://imgur.com/a/iMY1x

https://techcrunch.com/wp-content/uploads/2018/03/screen-sho...

View on HN

If you are building an autopilot system, I’d expect 1) identify black/yellow hazard signs and 2) don’t hit them to be basic features by version 0.5. Seems like a big miss to run straight into that.

Maybe SDC already do this, but it seems like the Agent should keep track of things that look like potential lanes and then if the potential lanes turn out to be not-lanes (have barrier in them) the Agent should flag the area that looked like the start of a lane so future SDC know to avoid it.

I wonder if the problem is the uneven pavement leading up to it. The white solid line starts close to it, then drifts left. In certain lighting conditions, the division in the pavement might seem more prominent than the white line, which I could imagine tricking a human or computer to follow it instead. That wouldn't cause a head-on crash itself, but would cause the car to get very close. I could then imagine a computer trying to correct its mistake by moving RIGHT (because that area looks like a lane) instead of left and going head-on into it.

Though of course that doesn't explain why it didn't recognize the barrier as something it should avoid at all costs. Unless perhaps something to the left confused it and made it think there was an obstacle there, too, thus causing it to think it was going to crash regardless. If so, maybe it did try to brake at the last second?

This divider looks rather unforgiving. Is it common for dividers in US to be just plain concrete blocks without anything to prevent rapid deceleration?

If you look at the imgur link you'll see a impact barrier that had previously been crushed and not reset or repaired.

I dont see it

It's in the techcrunch image, not the imgur. The barrier with the yellow face is designed to collapse.

https://techcrunch.com/wp-content/uploads/2018/03/screen-sho...

That still looks pretty risky to me - I compared a similar junction here in the UK near to where I live (on the M90 in Fife) and it has about 40 impact attenuation barrels in a triangle.

Those are also present in many locations in the US, although they seem to be phasing them out so I think they are considered old tech. I don't know the performance specs, but new cars are a lot smaller and safer than old cars - the same is probably true here too.

that image also makes me wonder if tesla's system is designed to identify knocked-over cones. It's one of those things that would be easy to code for, but also easy to overlook.

It's quite hard if you take into account that machine learning needs 10000x the training data that a human needs, and Tesla is sold in 30+ countries (and used in much more).

I think 3d mapping the surrounding world correctly using multiple cameras (what humans are doing) is more generalizable.

CalTrans doesn't diagonally stripe spaces like that? It would seem that they should.

There has been a lot of repaving work on 101, so these may well not be the final markings.

I don't think that area has been repaved in a while.

The lanes aren't very well indicated, but given how visible the divider is, I don't think either an autopilot or a driver should drive off.

Reminds me of a scene in "La grande vadrouille" where a motorcyclist is killed on a mountain road because he was following the dotted line and the painter had gone to the side of the road to take a break.

Also, that's an immense pothole!

> That’s an immense pothole!

... well you certainly aren’t from southeast Michigan, that’s for sure.

You would think they would at least add pylons leading up to the divider: https://ops.fhwa.dot.gov/publications/fhwahop13007/images/fi...

The divider is from the fast lane no less.

Does seem like a problem that a LIDAR based system would have been less likely to make, compared to AI on video classification

Except those solid white lines delimit an uncrossable barrier delimiting the HOV-specific (during commute hours) flyover from the 101 lane, which is well documented for nearly a mile preceding that point.

Reasonable, lawful, absolutely. People break the law, people get drunk, people get sleepy.

There's a spot just like this in Houston on 610 East. HUUUUUUGE flyover as the HOV lane spends about a half mile merging into the left lane of 610 (nice fat fast freeway, they need the runway). A guy had thought the flyway would be a safe place to park his car and wait for a tow. Safe enough that he was sitting on the hood of his car. Sure as shit the Challenger in front of me got confused, cut into the flyaway thinking it was a lane, hit the disabled car, and sent the guy flipping up into the air a good 10, 15 fit. Fucked up accident.

Anyway point is after the accident I was talking to the driver of the Challenger, trying to calm him down (kept saying "holy shit I fucking killed that guy!") and other that shock he was sober as a duck. He got bloodtested and everything, clean. In court he said he just got confused and thought it was a lane.

LONG STORY SHORT human error mang, human error. Still not sure why a car was able to do it.

They have those same white lines leading into intersections and people cross that shit all the time. People speed.

The entire thing is two solid white lines on either side, which are pretty strong indicators that you shouldn't pass into that area in the first place, diagonal stripes or not.

It also indicates that you shouldn't pass out of that area.

The vehicle should have at the very least stopped short of the road obstruction. But we don't yet know what led up to being in that lane, I was hoping the article in the OP would have taken us through it. Instead I'm getting the feeling this was a catastrophic software failure. Without more transparency on the issue, despite what Tesla may prefer I don't think I have any choice but to feel uncomfortable.

If you step a bit further back, it does not look so much like an actual lane: https://teslamotorsclub.com/tmc/attachments/streetv1-png.288...

A piece of software that mistakes that wedge for a lane in broad daylight has no business being deployed on the road, in my opinion.

It's not that straightforward. It's two adjacent lanes, which gets separated. Left one becomes an exit, the lane next to that one continues. The space between them keeps growing. The two lanes in question have a solid line between them before they separate to indicate you can't cross between them.

I read that that driver was arrested for DUI.

Really, I did: http://abc7news.com/automotive/exclusive-i-team-investigates...

Reply on news.ycombinator.com