[olabyne]

Come on, it's time to ditch this "vision alone" thing for self-driving cars. A computer will never be safe with vision only. Things like reflection and false positives will always wreck the statistics of even the perfect AI, there is no way around it. And I'm talking about good weather conditions.

The numbers show it, the few that Tesla releases about FSD are not up to the level with Google. If it was good, Tesla would be transparent with their results.

[TOMDM]

I think vision some will likely eventually be sufficient for full self driving, but by the time it is vision and LIDAR will be cheap and ubiquitous and downgrading to vision alone will be seen as a cost cutting measure that unnecessarily degrades safety compared to systems with LIDAR as well.

[ArtTimeInvestor]

Then how do humans do it?

[olabyne]

Human have excellent 3d vision, even with one eye. Some ability that only a vision-lidar fusion would match, not vision only. Human vision isn't just a flat image like a camera render. Distance from an object is not only estimated from using a stereo camera (2 eyes), it is made from constant movemnt from the head, and the ability to rotate the orbit on two axis.

Vision might be viable in an unknown future (which tesla is very good at seliing), but right now it is very dumb to forbid yourself extra data.

[ArtTimeInvestor]

    ability that only a vision-lidar fusion
    would match, not vision only

What would be arguments that back this theory up? Why couldn't an image stream from fixed cameras be enough?

After all, we made machines perform on a superhuman level in many areas already. You could have said "A machine will never be able to move as fast as a human unless it has two legs and feet". But that is not how it turned out.

I remember people saying computers will never beat humans in chess because they don't have intuition.

[jodleif]

I don’t see any reason to argue possibility- it’s more interesting to know what’s optimal. Would you agree that redundancy in sensors is better?

[ArtTimeInvestor]

Optimal towards which metric?

[olabyne]

What i meant is that if you multiply the number of cameras, probably make them movable, and with amazing vision software you might approach the performance of 2 humans eyes, but really it is just extra step to simulate lidar data at this point.

> After all, we made machines perform on a superhuman level in many areas already. You could have said "A machine will never be able to move as fast as a human unless it has two legs and feet". But that is not how it turned out.

Yes, thanks to wheels. Robots with legs do exist; but none of them is close the speed of an average human running. That's my point. We are not gonna do better than human vision by simply taking 2 cameras thinking it is equivalent to a human eye.

[ArtTimeInvestor]

How do you know that taking 2 cameras and superhuman intelligence won't be enough?

[olabyne]

Because it's physically not the same. Again, 1 fixed camera does not provide as much data as a single eye. It's not about the resolution, it's about the 3d perception that even a single eye provides. You can try yourself : close one eye, and you still have a good 3d perception. Now try to make your orbit fixed, no head or eye movement, and then everything feels flatter.

[buildbot]

By being the greatest general intelligences we know of? Even then, it’s still the most lethal thing most people do.

If we get a human level AGI with vision as high dynamic range as ours then maybe we can match human driving performance.

Seems easier to use radar and lidar at that point!

[MrHamburger]

How birds fly? Well definitely not like planes.

Sometimes it is easier, far simpler and even superior not to mimic nature. Consider if it is even possible to have supersonic flight by mimicking flapping of wings.

[grecy]

They don’t do it well. 100 people will be killed TODAY on roads in the US. (And everyday)

[Zigurd]

Humans do it badly. Humans do it by killing 50,000 other humans per year in the us alone. A million worldwide.

[codedokode]

Their hardware has more memory, compute capability and throughput.

[codedokode]

Also the image from the retina is higher quality compared to primitive cameras.