Waymo will beat Tesla FSD because of LIDAR

[Gaximus]

The radar returns are horrible. Sure improve the radar. Then one is presented with two data sets that fight with each other.

Some may dislike vision. Yet for me it sees the things I can't. Many a time I have had my car do what I thought was phantom braking. Only to see a deer on the other side of the road. A deer I never saw. In the dark I might add.

Can get lost in the weeds on this stuff. I like real world examples from people that drive 50k+ a year on AP/FSD.

No one is saying get rid of vision. You combine them, and get more data. For things that require precision, like following distance it uses radar, and things that require vision use vision.

Sponsored

 

[REM]

I'm not confusing it. I'm saying just that. The system can poll interpret and action data faster than a human can.

The "programmer" is AI that's supposed to be faster and better at manipulating data.

Every single derivative doesn't need to be human interpretable. It needs to be the opposite.

I can't tell where studs are, but a stud finder can tell me it within cms.

And in your example, a human is still needed in order to take action on where those studs are lol. I don't think we are on the same page here.

Ai systems aren't magic, they are just complex and high level programming. By humans.

 

[REM]

No one is saying get rid of vision. You combine them, and get more data. For things that require precision, like following distance it uses radar, and things that require vision use vision.

More data doesn't automatically equate to better outcomes. Telsa specifically got rid of radar because it's simply too much noise in the system for the programmers to properly gain a high order of safe response pathing.

Just because you can bake a turkey in the oven at 325° for 4 hours doesn't mean you can also get away with baking it at 1,300° for 1 hour.

 

[Deleted member 17810]

And in your example, a human is still needed in order to take action on where those studs are lol. I don't think we are on the same page here.

Ai systems aren't magic, they are just complex and high level programming. By humans.

Here's the thing, AI is being posited as something that humans don't need to program.

And that's where all these little shortcomings come in.

 

[kappaknight]

Sight, isnt' the most important thing humans have.

Put that aside, it feels like I'm taking crazy pills, when it's well documented that blind people have heightened senses, specifically around sound.

rhymes with cadar.

Whether it's most important or not isn't the issue. It's the most common and most human, which is what we're comfortable with. The AI model is being trained on sight because if the result is not jiving with our understanding of the world, we can understand why and fix it. We also understand what we're capable of with sight, as a baseline.

One can argue legs isn't the best way to move around vs. wheels or tank treads, or having 6 arms is better than having 2 arms; but people wouldn't adopt to Optimus if it looked like a kill-bot; they want something human-ish they can relate to and if it messes up, they can understand why.

 

[Crissa]

Yes, lidar gives accurate, mathematical, distance data, that fsd optical system has proven to not been able to deduct.

It's very expensive to do so.

The 'proof' you're using is a fallacy. Just because it didn't work in a specific combined single-stack machine-learned algorithm doesn't mean it will always not work.

That's exactly like saying onscreen keyboards are impossible or VR headsets can't track without mechanical interlinks. These were all impossible until they weren't.

This isn't about material strength or physics. It's not 'impossible' since the technology does, in fact, measure distances very accurately (tho not as accurate as lidar).

You want to pay tens of thousands of dollars for additional sensors?

Here's the thing, AI is being posited as something that humans don't need to program.

And that's where all these little shortcomings come in.

That's also not true.

First, it's not posited as something we don't need to program, it's a new way of programming. Second, these shortcomings exist is all the current implementations of robots - because making it pay attention to all of these things at once is not easy. But that doesn't mean it's impossible.

-Crissa

 

[Outdoors]

Talk to me AFTER you’ve used the technology for longer than just a sufficient amount of time to confirm your biases. ?

I would hazard a guess that I have about 300k more AP miles than anyone on this forum. So I think that is long enough. So add yours up and add another 300k.

I drive a Tesla all over the US and Canada to the tune of 50-70k a year. All in AP/FSD. So let me know what is sufficient.

Actually don't care to talk to you. That's why I am typing. I only care about my experience. I drive it. That's is the only experience that matters. Love the bias. This isn't statistics class. Let me know if you need me to teach you anything there. I studied it for many years. Love to school youngsters. Or those whose minds are young and dumb.

 

[Deleted member 17810]

It's very expensive to do so.

The 'proof' you're using is a fallacy. Just because it didn't work in a specific combined single-stack machine-learned algorithm doesn't mean it will always not work.

That's exactly like saying onscreen keyboards are impossible or VR headsets can't track without mechanical interlinks. These were all impossible until they weren't.

This isn't about material strength or physics. It's not 'impossible' since the technology does, in fact, measure distances very accurately (tho not as accurate as lidar).

You want to pay tens of thousands of dollars for additional sensors?


That's also not true.

First, it's not posited as something we don't need to program, it's a new way of programming. Second, these shortcomings exist is all the current implementations of robots - because making it pay attention to all of these things at once is not easy. But that doesn't mean it's impossible.

-Crissa

No, you are completely wrong.

Self driving needs to be based on accurate and real world dimensions. Telsa has decided, and said that their vision based system can make up the difference in accuracy. That's not up for debate.

They haven't proven that yet, and it continues to be a problem.

Exact dimensions of the real world is not hinderance. Any system that deems those things too "complex" and "confusing" is a terrible system.

Visual systems cannot right now, match mathematical sensor systems on accuracy. That's not up for debate.


It's why tesla uses them for validation.



It's expensive (monetarily and computationally) and it's why the don't want to do it, and they're trying their hardest, and it's not happening.

 

[Gaximus]

More data doesn't automatically equate to better outcomes. Telsa specifically got rid of radar because it's simply too much noise in the system for the programmers to properly gain a high order of safe response pathing.

Just because you can bake a turkey in the oven at 325° for 4 hours doesn't mean you can also get away with baking it at 1,300° for 1 hour.

That's not what is happening here, though. When training a model for machine learning, there can be millions of parameters, thousands per image, x amount of cameras, x amount of frames a second, etc. Adding one more parameter like radar, isn't going to cause any noise. In fact it can negate the necessity for dual front cameras(or at least all the calculations need to determine all the calculations needed to determine distances., buy comparing all the features in both cameras, many times a second). It is also a lot more accurate for distance than 2 cameras could ever be.

The reason Tesla got rid of radar, is because they were hard to source during the pandemic. Also they said they used more power, and needed to be heated in the winter to prevent snow build up. And with vision only, they could get "close enough" data to make it work. But the argument here is that adding radar, make it better and more accurate, it takes away nothing.

 

[Bill906]

Tesla Full Self Driving (Supervised) is only Level 2. The human must be alert and actively supervising at all times.
Level 3 does not require the human to be alert at all times. The human must be there to take control IF the ADS requests the human to intervene. Otherwise the Level 3 system is in complete control and the human can, for example, read a book, or watch a movie.
I'm assuming we are talking about SAE Standard J3016.
Tesla FSD is a very high end Level 2 ADS, but the step from level 2 to 3 is HUGE.

 

[Deleted member 17810]

More data doesn't automatically equate to better outcomes. Telsa specifically got rid of radar because it's simply too much noise in the system for the programmers to properly gain a high order of safe response pathing.

Just because you can bake a turkey in the oven at 325° for 4 hours doesn't mean you can also get away with baking it at 1,300° for 1 hour.

More absolute data, absolutely equates better outcomes.

When you're talking about spatial awareness, you need to know the dimensions of the space you are occupying.

You can absolutely bake a turkey at 1300 degrees for an hour, if you break it up into pieces and fan it out.

Details matter.

 

[dalton108]

I would hazard a guess that I have about 300k more AP miles than anyone on this forum. So I think that is long enough. So add yours up and add another 300k.

I drive a Tesla all over the US and Canada to the tune of 50-70k a year. All in AP/FSD. So let me know what is sufficient.

Actually don't care to talk to you. That's why I am typing. I only care about my experience. I drive it. That's is the only experience that matters. Love the bias. This isn't statistics class. Let me know if you need me to teach you anything there. I studied it for many years. Love to school youngsters. Or those whose minds are young and dumb.

Bro. I’m agreeing with you; you’re violently agreeing with me. Please take another look at my post.

I’m talking about people who only use attack long enough to confirm their biases and then write an article or post blog or do a YouTube video.

I’ve had auto pilot since mid-2019 and FSD Since September/October 2020. That’s five years but a lot closer to a total of about 100,000 miles over four different vehicles I know it works because I’ve used it. In fact, I’ve got well over 10,000 hours which my experience makes me an expert many times and you even more than that. So that experience simply doesn’t compare to somebody who and you even more than that. So, that experience simply doesn’t compare to somebody who is has just dipped their toes in the water long enough to be able to say I tried it and I don’t like it, just like I knew I wouldn’t.

Same side, dude, same side.

 

[CybertruckWraps]

humans don't have lidar and drive perfectly well a camera based system should be able to do the same or even better since it has 360 field of view. .

I wouldn’t say perfectly well ?

 

[CyberGus]

Visual systems cannot right now, match mathematical sensor systems on accuracy.

If you need millimeter-level accuracy for your driving, you're doing it wrong

Sponsored