Hacker News Reader

Comment by cbames89 | Hacker News Reader

This is a (perhaps short-term) failure of first principles reasoning. Elon's known to favor thinking about things from first principles, and there's no theoretical reason that vision can't work. However, technical limitations might take a while to overcome.

View on HN

Indeed. There's also a slightly less charitable interpretation–Tesla could't ship every car with LIDAR for cost reasons, so they're strongly incentivized to claim they don't need it.

That said, I don't think Tesla are wrong. If Tesla can deploy MobilEye-level automatic visual mapping (see https://youtu.be/GQ15HWCw_Ic?t=1381) this can obviate the need for LIDAR-based localization and dramatically improve their perception systems (by having very good prior information about all static obstacles, such as lane dividers).

Dynamic obstacle detection will still be worse than it would be with a vision+LIDAR approach, but not that much worse. Having superhuman ability along two axes (map-based priors and alertness / reaction time) is likely sufficient to drive better than humans, even if several remaining axes (dynamic obstacle detection and prediction) are worse.

Additionally, Tesla is not (AFAIK) using any structure-from-motion (https://www.youtube.com/watch?v=KT2KsN7yKo0) or stereo-vision (https://www.youtube.com/watch?v=SskSDjUG8ZY) techniques, but there's a chance that these could also improve perception, especially in tough cases where a single frame is not enough for good detection (e.g. white semi in sunlight).

LIDAR is not being used in the (compared to Tesla) small-scale AV systems of Waymo, Cruise, Aurora et. al because it's essential, but rather because it's convenient, and because the companies producing those systems want to give their AVs every advantage at any cost. I fully expect Tesla (and MobilEye) to do superhuman self driving without LIDAR, but it will take longer (i.e. not by EOY).

People should strive for correct reasoning. First principles reasoning, as I interpret that term, is just an approach that helps to achieve correctness by overcoming ingrained assumptions.

If Elon really thought "in theory, camera-only full self driving is possible, therefore I should invest resources in that approach" than he's dumber than I thought. He's done it because Lidar was not an option - and the claim that FSD is coming soon was an intentional lie.

The simple fact that Humans only have eyes and don't need lidar or radar to drive, however we cause accidents when our eyes misinterpret a situation, there are bad road markings or our vision is impaired due to bad conditions a camera only car will have the same faults.

So Personally I would feel safe in a camera only car driving during the day in good conditions and with good road markings or on a controlled road where humans are excluded. I would not feel safe in any other situation..

I still want a model3 it is so good in every other way but I wont be trusting the self driving.

Humans are general intelligences requiring almost two decades of specialized training and growth before we even think about letting them drive. The human eye is also vastly superior to any camera in existence, so even being generous there is. O comparison. Maybe such a comparison will be valid decades from now, or longer, but Tesla is selling people a lie today and it’s a potentially dangerous lie at that.

> the claim that FSD is coming soon was an intentional lie.

What's your evidence of intent?

We figured out the first principles of fusion power in 1930s, but are still nowhere close to practical fusion power generation. The engineering difficulties are just too great, and simpler alternatives exist which get much more attention, funding and optimisation.

Similar might happen to vision-only self-driving. It might take a decade or two longer to develop compared to LIDAR-based approaches, and meanwhile LIDAR is only going to get cheaper.

Vision alone fails pretty frequently even with a human brain and 700 million years of evolution behind it. Count me as a skeptic that cameras + computer will produce a safe self-driving system.

> 700 million years of evolution behind it

Sure, but only about 0.000014% of that time occurred while cars existed. It's not like we evolved to control big metal objects hurtling down artificially made flat surfaces.

...with devices that we're emotionally and physically addicted to flashing and making noises that are designed to ignite a flurry of activity in our reptilian brains sending the message PAY ATTENTION TO ME IMMEDIATELY!

On the other hand, while our brains did not evolve to control cars, our cars were designed to be controlled by our brains. A system designed from scratch for autonomous vehicles would look radically different.

And even then, accidents where all parties are following the rules, e.g. being sober, undistracted by something in the car, giving enough stopping distance to the vehicle you're following, going the appropriate speed for the visual and road conditions, &c aren't the norm. Most accidents happen because somebody was irresponsible, usually blatantly.

We can't measure that evolutionary advantage.

What we can measure is fatalities

Really doesn't though. Humans are pretty good at driving, and 90%+ of accidents come from people being drunk/distracted/tired etc. We don't crash because we miss categorize a truck as a plastic bag or w/e.

This implies that the vision itself is to blame and ignores the actual causes of crashes.

A computer doesn't get distracted by a text message. A computer doesn't get drunk. A computer doesn't day dream.

Cases where a crash happens and the person at fault says something like "I didn't see them!", it's usually because they didn't even look, a mistake a computer won't make.

> there's no theoretical reason that vision can't work

Human brains, which are general intelligences, exist, and are collections of atoms.

I'll happily, for $500,000/copy, upfront, promise to sell you artificial general intelligences. I swear I'll get them built in the next two years.

There is, after all, no first principle reason for why my promise isn't worth the paper it's printed on. Human brains are collections of atoms, so we should be able to, out of atoms, build artificial general intelligences.

Reply on news.ycombinator.com