Hacker News Reader

Comment by jowday | Hacker News Reader

Here's a Tesla hitting a highway wall from last week.

Interesting, I think this is the spot: https://www.google.com/maps/@34.0357838,-118.169596,3a,75y,2...

Seems like when the Street View pic was taken this was not a lane. I'm shocked it was converted into one. Seems very dangerous, for human or AI.

I’d probably hit that wall too. I wonder how many non-teslas have hit the same spot.

One frame is that Teslas will always hit walls another is that these mean Teslas will never hit such walls again. Absence of collisions isn’t evidence the systems work, it may be, but could also be that other systems aren’t deployed enough to hit the tail of the distribution to see these scenarios. However, known incidents like this one do grant evidence to the claim the said system won’t continue make the same kinds of mistakes from the feature space of inputs most like it (which is hard to know as a human observer)

The Tesla should have caught that.

But that is the stupidest way of temporarily? (I hope) relaning a highway ever. What were they thinking?

One of those situations where Musks insistence on not using LIDAR is clearly wrong.

The car has radar which can see that wall. It’s most likely a problem of what to do, which humans face just the same - do you swerve into the right lane potentially causing a much worse accident?

I believe Teslas rely 100% on cameras for this type of maneuvers. Regardless, it didn’t need to swerve into the next lane, just move to the right 6 inches in the same lane.

Can radar see concrete? I thought radar was only reliable for metal obstacles.

It could check its side cameras first.

LIDAR would not have made any difference if the car is not coded to handle the situation, let alone how in the hell is that even legal? We can only make an assessment of what the car saw if he were lucky enough to have the screen data as seen in car if not all telemetry from its recordings.

I own a TM3 and I do not have the "FSD beta" that is in limited testing. I will say that the car display portion of the UI really does not impress me in what it shows and does not show.

As in it is damn happy to show cones and garbage cans but I have instances where it showed me cones front and back of a vehicle but did not draw the vehicle in between. Same with a floating garbage can where sometimes the loader was visible but the truck itself would wink in and out.

I really wish people would quit ascribing super natural abilities to LIDAR let alone superiority. Plus I really want to see studies done which show the impact having hundreds of LIDAR equipped vehicles would have on people, animals, and even insect, life. (then again how many vehicles in traffic does it take to cause an issue for them)

The only reason I mentioned LIDAR is because I strongly suspect the reason the Tesla didn’t avoid that is because the angle of the sunlight made that pillar difficult to see.

I fully agree with you that that spot on the road is pretty horrible, which is why I mentioned it. But dealing with crappy roads is the inherent problem any self driving car has. It’s the car’s job to not hit things.

I own a Model Y and I feel like Tesla’s autopilot gets a worse rap than it deserves. But there is an inherent disconnect between what the car is capable of and what some people presume it is capable of. Tesla doesn’t help this with their branding.

Why did the driver just let his car plow into the pillar instead of taking control? Were they even looking up at the time? Would you have let your car drive into a pillar like that? I’d like to think I would have, but perhaps not.

How can any AI make sense from such a low-quality video? Why isn't there some kind of lens hood that blocks the sun glare?

Yes, but that's not the old version. The latest versions are failing in new and interesting ways. /s

Reply on news.ycombinator.com