[theshrike79]

Tesla's camera-only driving is a gamble. It has the _potential_ to be the best in the business.

...but it also takes a ton of computing power to determine how far objects are just from a moving image. A LIDAR unit would give you range + position instantly.

[steveBK123]

Also assumes the current camera suite is high enough resolution, high enough dynamic range, sensitive enough at night, and positioned in correct locations to provide the data the compute needs.

Arguably that is not true still.

The issue with the camera-only model is it is very non-deterministic.

Personally, every accident I have gotten into is a result of some sort of inattention, distraction, or fatigue. Computers don't experience these things.

Nonetheless, Tesla's occasionally do unpredictable insane things like turn directly into a bollard, cross in front of a moving trolley (that it can see & has rendered on screen), run right into a firetruck, run over child size dummies, etc.

It's worse than supervising a teen driving because a teen will fail in somewhat predictable ways, generally in being either overconfident or over hesitant. You can verbally coach a human as a situation approaches. With FSD you either let it ride, intervening at last moment.. or you need to constantly step in proactively.

[jfengel]

"Not hitting things" isn't really the hard part of driving, though. Yeah, it does sound like the absolute minimum level of effort, but it doesn't get you anywhere. It just tells you some of the places not to go.

You need to read lane markings, locate cross streets, interpret signs, stop at traffic lights, etc. All of that requires a camera... and if you can do that, you might as well use it to not hit things.

LIDAR would be great in a belt-and-suspenders way, and it seems as if it would have saved some lives already. The other stuff is all incredibly hard; it's not ready and may never be ready. But if self-driving ever works, it would have to be able to work without LIDAR, because it has to solve the harder problem that also encompasses the things that LIDAR could help with.

[lm28469]

> "Not hitting things" isn't really the hard part of driving

Until your tesla hits a stopped firetruck at 60mph, then it's a real hard part

[jfengel]

That's just it. If a Tesla can't figure out that it shouldn't drive through the big red thing, then it can't do anything at all.

I don't have any idea why that happened. It seems odd. I'm afraid I'm always a little suspicious of such stories, since so many turn out to be people lying about the auto-driving.

It's entirely possible that Tesla failed to train the things on fire engines. I rode in a Tesla the other day and was a bit surprised that it didn't note ambulances as different from other trucks. If they've failed to train them on any emergency vehicles, that's a serious problem: there are laws like "pull over for fire trucks" that it can't follow.

And again, that's not a problem LIDAR can solve. Something is very wrong if it identified a big red object as a thing it was allowed to drive on. LIDAR would solve the accident, but not the fundamental question of whether self-driving would work at all.

[lm28469]

> And again, that's not a problem LIDAR can solve.

Tesla are scared of clouds and shadows and don't see big metal boxes

It's 100% something a lidar would help with. It wouldn't hit shadows or clouds but it definitely would hit a big metal box early enough to at least brake instead of accelerating