-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Will Tesla ever do LIDAR?
- Thread starter diplomat33
- Start date
Uh?!?! Elon was born in 1971. He was not even born yet in '61.
Daniel in SD
(supervised)
He's talking about all the things they use cameras for throughout the presentation. What are they using LIDAR for that you think cameras would be a superior solution?
Daniel in SD
(supervised)
It's an analogy. I guess I need to explain it.
I'm arguing that usually technology proceeds in a linear fashion. First you put an object in space, then an animal, then a man, then you go to the moon, then you go to mars (hopefully!).
I believe the same thing will happen with autonomous vehicles. First they'll be super expensive and very limited in operational domain and then... hopefully you get the idea.
@StealthP3D
The thing we disgree mostly on — and I’m pretty much thinking we’re approaching the polite agree to disagree stage — is where machine learning can and will be applied. For you it seems to be only vision and vision-related driving policy.
I see machine learning being applied more widely.
First, machine learning is of course used to make sense of Lidar inputs (including weeding out false positives), just like in vision.
But more importantly I see it being used in sensor redundancy as well. The thesis goes like this: Every sensor-type has its inherent limitations that can not be overcome, they are laws of physics. No amount of AI will overcome the laws of physics. However, when you have different sensor-types feeding the neural networks with data, with inherently different physical limitations, suddenly that AI can overcome those physical limitations by merging those inputs.
Machine learning is used to make use of that data. Basically a neural net decides whether or not to trust the visual, or the radar or the Lidar (or whatever sensors may be availbale now or in the future) in a particular scenario.
The thing we disgree mostly on — and I’m pretty much thinking we’re approaching the polite agree to disagree stage — is where machine learning can and will be applied. For you it seems to be only vision and vision-related driving policy.
I see machine learning being applied more widely.
First, machine learning is of course used to make sense of Lidar inputs (including weeding out false positives), just like in vision.
But more importantly I see it being used in sensor redundancy as well. The thesis goes like this: Every sensor-type has its inherent limitations that can not be overcome, they are laws of physics. No amount of AI will overcome the laws of physics. However, when you have different sensor-types feeding the neural networks with data, with inherently different physical limitations, suddenly that AI can overcome those physical limitations by merging those inputs.
Machine learning is used to make use of that data. Basically a neural net decides whether or not to trust the visual, or the radar or the Lidar (or whatever sensors may be availbale now or in the future) in a particular scenario.
Daniel in SD
(supervised)
So what is phantom braking if not a false positive? All systems can have false positives.
Yeah. And Lidar’s greatness isn’t lack of false positives — unfortunately all sensors can have them and need to be weeded out through machine learning and the like.
The great thing about Lidar is virtual lack of false negatives. If Lidar says there is nothing there, and your vision also says there is nothing there, you can be pretty darn sure there is nothing there even if your radar is pinging like crazy on the overhead bridge.
I think you are exaggerating. Tesla has a lot more than a powerpoint. They got hundreds of thousands of cars on the road right now that are getting better and better with each software update. And they got a working, almost "feature complete" self-driving system that is running on a computer that Tesla designed in house.
Ok, I get it. I guess I like that Elon shoots for the Mars of FSD, instead of going to the Moon first. But I do agree with you about tech starting off expensive and limited and then getting better. I am sure 30 years from now, level 5 autonomous cars will be common place and we will look back at how "primitive" 2019 self-driving tech was.
Of course, but only to make a point — not to be disingenious.
If we are talking about naysaying, then calling Waymo and its ilk DOOMED certainly is naysaying.
I guess the comparison would be if Musk had entered the conversation in 1961 and said we need to go to Mars instead and going to Moon is DOOMED.
I mean, it makes sense. If someone had come out and said in 1961 we should go to Mars by 1965 instead of Moon by the end of the decade and then explained how it can be done at a tenth of the cost, that would have been a similar scenario.
I’m sure it would have been compelling too. Too bad it could also be totally false and we wouldn’t really know for sure. If that Mars by 1965 failed spectacularily, it is possible no human would have stepped on another rock yet.
I’m open to it working. I’m just not saying either way for sure — analyzing the pros and cons of each instead.
I’m sure it would have been compelling too. Too bad it could also be totally false and we wouldn’t really know for sure. If that Mars by 1965 failed spectacularily, it is possible no human would have stepped on another rock yet.
I’m open to it working. I’m just not saying either way for sure — analyzing the pros and cons of each instead.
CarlK
Active Member
He's talking about it but it's only academic. Cruise is not using it.
The biggest disadvantage of Lidar, other than it's expensive, is you can't equip a large fleet to collect data for deep learning. Just get rid of Lidar if you want to go to vision deep learning system. There is no benefit of it being there other than to tie your hands. That said even if Cruise or others decides to go vision deep learning they still have to in some way get a large fleet going and to start train the neural net. That will take a long time to implement. Again Elon had the whole plan laid out years ago and integrated to the entire Tesla business plan. That's why Tesla could have what it has today. People can ridicule him all they want but in the end the vision (as in visionary) and first principle practice of his still won the game.
Last edited:
Well, I agree that calling Waymo doomed is ridiculous. But we need to remember that Tesla and Waymo have very different business perspectives and business models. To use your space travel analogy, Waymo is indeed going to the Moon and Tesla is going to Mars. Waymo's goal is having a small fleet of robotaxis that can take customers for rides in a small geofenced area. Tesla's goal is to turn every Tesla car into a robotaxi that works everywhere. Completely different! Tesla's goal is much more ambitious but it makes sense based on Tesla's business model. Tesla already has hundreds of thousands of cars on the road. If Tesla can turn all those cars into robotaxis, then it gives their customers huge value, as well as income potential. It also needs to work everywhere because Tesla's customers live everywhere. If Tesla did achieve self-driving in say California, that would be worthless to all the Tesla cars driving outside California.
You don’t need a large deployment fleet for deep learning. It may or may not be an advantage, but you definitely don’t need it. There are alternative ways of pursuing that.
I don’t disagree it makes sense for Tesla, though it is questionable of course if going ”all in” back in 2016 (and now doubling down) was necessary (or wise) for that goal.
Anyway, the question is: will it work in the given timeframe, to the given extent (Level 5 no geofence) and with the selected sensor suite? I mean the story changes if instead of landing on Moon in 1969, you would have killed, say, three consecutive sets of astronauts launched towards Mars in 1965 and perhaps lost the manned space program for a few decades as a result.
I guess that’s pretty much what I enjoy analyzing here about all these players. I’m trying to understand what and why they do and how well they do it — and how well they can be expected to do in the future.
I do think that everybody has a piece of the FSD puzzle, but not the whole.
No, I think Tesla will achieve FSD before SpaceX goes to Mars.
StealthP3D
Well-Known Member
Ummm, no, that's not the question. The question is "Will Tesla ever do LIDAR?", read the thread title!
And the answer is an unqualified "No". Musk has not wavered on this and he is a bonafide LIDAR expert. When you understand how neural nets work with machine vision, you will understand why. It's more productive to continually improve the machine vision than it is to try to interleave LIDAR data into the mix.
That’s not what I said though.
I said I find SpaceX’s expectation setting on Mars more realistic than Tesla’s on FSD.
Similar threads
- Article
- Article
- Article
