Start-up to bring cheap, compact LIDAR to production cars by 2022!!

[plumazul]

You can actually make the same argument about lidar.

Lidar is being used only to get to some kind of geofenced L4 quickly to get VC money.

Vision only FSD is not wild speculation. It is just "first principles". If humans can do it with vision only, theoretically we should be able to do it with NN as well.

Except that eventually the cameras will FAR exceed the natural abilities of any animal on Earth. (e.g. extreme night vision, additional spectrums, and a NN that can decide which is best in RT) LiDAR was a stopgap. Progress in FSD is a software issue now.

 

[CarlK]

It you watch the Lex Fridman's self driving state of art lecture he put clearly pros and cons of the two types of technology. The Lidar + HD map everyone uses could get you to perhaps 99.9% to do demos to impress investors or boards but will never improve from there. Vision + deep learning only Tesla is using (other than perhaps Levandowski's new company although I have no idea how he could obtain the necessary data) will continue to improve until it gets there. It will only take time but there is a clear path there.

People who say look at how it could do now should also know you can't by watching how a toddler walk to predict if he will grow up to become a NBA player. The ability to learn and improve makes all the differences.

 

[acoste]

Except that eventually the cameras will FAR exceed the natural abilities of any animal on Earth. (e.g. extreme night vision, additional spectrums, and a NN that can decide which is best in RT) LiDAR was a stopgap. Progress in FSD is a software issue now.

If those future systems will rely on laser and distance measuring, will you call them lidar or camera?

 

[diplomat33]

It you watch the Lex Fridman's self driving state of art lecture he put clearly pros and cons of the two types of technology. The Lidar + HD map everyone uses could get you to perhaps 99.9% to do demos to impress investors or boards but will never improve from there. Vision + deep learning only Tesla is using (other than perhaps Levandowski's new company although I have no idea how he could obtain the necessary data) will continue to improve until it gets there. It will only take time but there is a clear path there.

The only caveat I would add though is that companies like Waymo are not using just lidar + hd maps. They use cameras too.

 

[CarlK]

The only caveat I would add though is that companies like Waymo are not using just lidar + hd maps. They use cameras too.

That's not the same though. You can use as many cameras as you want but without deep learning you still can't do what Tesla is doing.

 

[EVNow]

The only caveat I would add though is that companies like Waymo are not using just lidar + hd maps. They use cameras too.

Waymo's problem is not lidar - it is getting real world data. They rely on simulations - and you can watch all the issues round that in the talk given by the Waymo engr for Lex's class.

 

[diplomat33]

Waymo's problem is not lidar - it is getting real world data. They rely on simulations - and you can watch all the issues round that in the talk given by the Waymo engr for Lex's class.

I know. I was merely refuting the idea that Waymo is just using a lidar + hd maps approach.

But to your point, I can't help but wonder if Waymo will struggle a bit to deploy their cars to new areas because they lack the real world data. So while they have a great "feature complete prototype" that can probably work in any city, Waymo still needs to test their cars all over again in each new city to make sure it works reliably.

Tesla has an advantage with real world data thanks to the fleet. As we've seen with past features, once Tesla has trained the network, they can release features that generally work everywhere.

 

[Eno Deb]

Tesla has an advantage with real world data thanks to the fleet.

Have we seen any evidence of that so far? Can Tesla do anything that Waymo can't? I think it is not so obvious that a given number of vehicles driven around randomly by laymen and having only limited uplink capacity will provide more useful data than a smaller fleet driven in a targeted way by trained drivers and with unlimited uplink capacity.

As we've seen with past features, once Tesla has trained the network, they can release features that generally work everywhere.

The fact that they use ADAS maps to control certain features shows that they don't just work anywhere.

 

[diplomat33]

Have we seen any evidence of that so far? Can Tesla do anything that Waymo can't? I think it is not so obvious that a given number of vehicles driven around randomly by laymen and having only limited uplink capacity will provide more useful data than a smaller fleet driven in a targeted way by trained drivers and with unlimited uplink capacity.
The fact that they use ADAS maps to control certain features shows that they don't just work anywhere.

I think the advantage is in how Tesla uses it. Tesla is working to build their own neural nets from scratch. To do this, they need to feed the computer with a lot of data in order to train it. The fleet provides that large amount of data. Karpathy has given us some specific examples. When Tesla wanted to train the vision NN to recognize cars cutting in, they got lots of short video clips of cars merging and fed those clips into the machine until the NN reached a high accuracy of being able to predict if a car is going to cut in. Tesla can get clips of anything they need to teach the NN from traffic lights to debris on the road etc... So no, I don't think it gives Tesla some secret ability that Waymo can't do, but it does help Tesla build their NN's much faster.

The other advantage is that the large fleet can provide feedback to Tesla to help them tweak their software. Tesla can run the software in shadow mode on a lot of cars in various environments and see how it performs. Furthermore, when Tesla does release "City NOA" with driver supervision to the general public, Tesla collects driver intervention data and can further tweak the software to handle edge cases. We saw this with how Tesla released NOA with and then without confirmation.

Waymo is different. Presumably, Waymo already has their own vision neural nets so they don't need to build them. Waymo uses a combination of HD maps, lidar data, cameras and radars to do their self-driving. Waymo's smaller fleet with trained safety drivers operating in a small geofenced area, does have the advantage of being able to perfect the FSD software to work extremely well in that geofenced area. Which is why we see Waymo cars operating at far better FSD in their areas that what Tesla has. When you test your cars over and over again in a small area, you can really focus on making them great in that area.

My question about Waymo's approach is that I am not sure it translates super well to other areas. Just because you get your cars to self-drive super well in say Phoenix, does not mean that they can automatically self-drive super well in San Francisco. So Waymo has to test their cars extensively in each new area.

I do think that once Tesla gets to "feature complete FSD" that the large fleet will be able to provide feedback that Tesla can use to perfect their software. For example, maybe Tesla cars handle intersections really well in one city but not so well in another city. Tesla will get that disengagement data and tweak the software to work in that city. So Tesla will essentially be able to test their FSD in every area that has Tesla cars simultaneously.

Think of it this way: Waymo's testing will expand out. Waymo started in one city and then expands to 2 more, then expands to 3 more, etc... Tesla's testing is more like a rising sea level. It's starts low everywhere and rises slowly everywhere.

 

[Eno Deb]

I think the advantage is in how Tesla uses it. Tesla is working to build their own neural nets from scratch. To do this, they need to feed the computer with a lot of data in order to train it. The fleet provides that large amount of data.

For one, they can't store or upload everything (which Waymo, BTW, can, because their cars are less cost sensitive and can e.g. have as much onboard storage as they want). They have to set specific triggers, and due to cost reasons it will be only small clips and limited to a subset of the cars. Also, based on my own experiences with neural nets (in a different industry) the bottleneck is usually not the collection, but the labeling of the data. I bet a lot of the training data they use is actually from other sources.

Tesla can get clips of anything they need to teach the NN from traffic lights to debris on the road etc...

Sure, but they have to wait around until the cars where they have installed triggers actually encounter the situation they are interested in. And that still doesn't mean what they get is useful as training data (e.g. getting thousands of clips of the same traffic light isn't necessarily helpful). Waymo can specifically target things.

The other advantage is that the large fleet can provide feedback to Tesla to help them tweak their software. Tesla can run the software in shadow mode on a lot of cars in various environments and see how it performs.

The so-called shadow mode is much more limited than people think. It's not a magic bullet. That was pretty clear after Karpathy's presentation.

Waymo is different. Presumably, Waymo already has their own vision neural nets so they don't need to build them.

I bet they are constantly refining them further. By all accounts they are also further along in using neural nets for drive policy, which also requires training.

Waymo uses a combination of HD maps, lidar data, cameras and radars to do their self-driving. Waymo's smaller fleet with trained safety drivers operating in a small geofenced area, does have the advantage of being able to perfect the FSD software to work extremely well in that geofenced area. Which is why we see Waymo cars operating at far better FSD in their areas that what Tesla has. When you test your cars over and over again in a small area, you can really focus on making them great in that area.

In practice, Tesla's fleet is also clustered in a few metro areas in the US. I doubt they receive much data from, say, Indiana or Poland.

Think of it this way: Waymo's testing will expand out. Waymo started in one city and then expands to 2 more, then expands to 3 more, etc... Tesla's testing is more like a rising sea level. It's starts low everywhere and rises slowly everywhere.

Yeah, it seems they are trying to boil the ocean. But as far as true "full self-driving" is concerned, they haven't even gotten a tea kettle to whistle yet.

 

[EVNow]

In practice, Tesla's fleet is also clustered in a few metro areas in the US. I doubt they receive much data from, say, Indiana or Poland.

Much more than Waymo, for sure

Infact probably more than all of waymo fleet !

For one, they can't store or upload everything (which Waymo, BTW, can, because their cars are less cost sensitive and can e.g. have as much onboard storage as they want).

Nothing stops Tesla from running their own test fleet. I'm sure they do.

 

[ZeApelido]

That's not the same though. You can use as many cameras as you want but without deep learning you still can't do what Tesla is doing.

Waymo is definitely doing camera based deep learning. Just not with as much fleet data.

 

[Eno Deb]

Waymo is definitely doing camera based deep learning. Just not with as much fleet data.

Google has plenty of training data, and you help them label it every time you solve one of those damn image captchas ("click all images showing traffic lights").

 

[CarlK]

Waymo is definitely doing camera based deep learning. Just not with as much fleet data.

Your last sentence pretty much invalidated the first one. Deep learning neural net is inaccurate unless you have a huge amount of data which Waymo would not be able to obtain. Even with Tesla sized fleet it will still take months or years more data collection to cover most super edge cases. Yes Waymo does do simulated training but like Elon pointed out that just would not work. If you can simulate all edge cases you have already solved the self driving problem.

As for the Waymo program there was an excellent post describing the history which unfortunately I could not find anymore. When Google started the self drive program in 08' as a moonshot project cameras weren't as good and more importantly modern deep learning technique was not yet available. It just had to do with what it best could. After years of development it's hard to start over again with the new tool. They could add some pseudo-deep learning bandaids but they could never have the deep learning NN unless if they re-write the entire SW from scratch. Waymo could be doing development work secretly but even with that lack of fleet data would still be the show stopper.

Most people have the (false) impression that Google should be ahead in AI and deep learning of everyone else. Some of that impression came from publicity Google got from AlphaGo developed by its Deep Mind subsidiary. But do you know Google only bought Deep Mind Technology in 2014 and the company was founded in 2010? Guess who was one of the major investors of Deep Mind in the beginning? Hope you guessed it right that's a guy named Elon Musk. He of course was also the one who started OpenAI in 2015 and Neuralink in 2016. It is safe to say when Tesla started its self driving program around or before 14' Elon was the most, if not the only, knowledgeable executive in the AI area among all of those in those auto or tech companies that are now working on this. It's no dumb luck that Tesla is the only company that got it right. Elon knew exactly what he needed to do, like hiring those AI chip guys in 15', with his first principle of engineering. We just did not know that at the time (and most probably still don't even now).

 

[ZeApelido]

I didn't say Google was more advanced, or had enough data, just they are doing deep learning with camera data as well.

Google certainly has the most machine learning talent by sheer quantity, but how much works in the area of video learning is unclear and maybe not any advantage over Tesla

 

[CarlK]

Google sure has a lot more resources but it seems Tesla set the right direction and used its resource wisely. The same as GM or VW have a lot more auto engineers but they still could not make cars as good as Tesla is making.

 

[Eno Deb]

I didn't say Google was more advanced, or had enough data, just they are doing deep learning with camera data as well.

Of course they do. And they likely have far more usable training data to train vision networks for autonomous vehicles than anybody else. They have many other sources besides the Waymo test vehicles (such as all those street mapping cars that have driven pretty much everywhere).

 

[acoste]

Waymo could be doing development work secretly but even with that lack of fleet data would still be the show stopper.

No secrets there, it's called ChauffeurNet.

I have an issue with deep learning however. It needs a bunch of accidents before it learns how to avoid them.

 

[ZeApelido]

No secrets there, it's called ChauffeurNet.

I have an issue with deep learning however. It needs a bunch of accidents before it learns how to avoid them.

Tesla will have no problem getting that data. Seriously that's their benefit in lots of fleet data edge cases.