Will Tesla ever do LIDAR?

[heltok]

 

[StealthP3D]

It looks like the most knowledgable people in the field of AI and self-driving might be finally coming to agree with Elon Musk about the usefulness of LIDAR in self-driving. Disagree with Musk at your own peril because he has a steady history of making fools out of those who disagree with him or say he can't do something.

The Autonomy Investor Conference today did a great job of informing about why LIDAR is pretty much useless in self-driving. It actually blew the myth that LIDAR is key to self-driving right out of the water! The 3D map created by the Tesla cameras was impressive in its detail. LIDAR doesn't have the resolution to tell the difference between a plastic bag and a discarded tire. If you watched this presentation and are still convinced that LIDAR is an essential (or even valuable) component of autonomy, then I can't help you:

I think Tesla shorts might want to turn their attention to those who are wasting billions of dollars trying to make LIDAR useful to any significant extent in an autonomous vehicle.

 

[heltok]

I like the point they made that if you are gonna use cameras and an active sensor for situation where the camera cannot see, it makes sense to use a different wavelength than the cameras for the active sensor. Because otherwise they will both fail at the same wavelength-blocking smoke.... Thus Radar makes more sense than infrared Lidar to use besides the cameras.

 

[CarlK]

One interesting thing Karpathy said in his presentation today was Tesla vision NN is detecting small movements of cars and pedestrains (body languages) to predict their intentions like changing lanes or stepping out of the curb. Only vision system has the resolution to do that but Lidar does not. And these are all trained automatically with cars in the shadow mode. The machine will make predictions and than check with what actually happened. This is what Levandowski described as the main area to work today and Tesla already has it solved.

 

[electronblue]

I like the point they made that if you are gonna use cameras and an active sensor for situation where the camera cannot see, it makes sense to use a different wavelength than the cameras for the active sensor. Because otherwise they will both fail at the same wavelength-blocking smoke.... Thus Radar makes more sense than infrared Lidar to use besides the cameras.

Then Tesla really should be like the others and actually use 360 degree radars. They talk about redundancy but at the same time they only have a single radar in front that currently is incapacitated by snow very easily and visual redundancy only up front. Ultrasonics are no good at distance or speed...

Because one of the point of using Lidar is redundant 360 degree view (not just a front Lidar but side and back Lidars as well). Other players in the industry have insisted on 360 degree vision, 360 degree radar, 360 degree Lidar.

This lack of sensor redundancy is something that has been nagging my mind about Tesla’s Level 5 AP2 suite since October 2016 and the same concern remains. But now they are two, three years the wiser and they say this suite will do Level 5 no geofence robotaxi so there are no if or buts about it, they must have a solution on how to overcome this.

 

[electronblue]

One interesting thing Karpathy said in his presentation today was Tesla vision NN is detecting small movements of cars and pedestrains (body languages) to predict their intentions like changing lanes or stepping out of the curb. Only vision system has the resolution to do that but Lidar does not. And these are all trained automatically with cars in the shadow mode. The machine will make predictions and than check with what actually happened. This is what Levandowski described as the main area to work today and Tesla already has it solved.

You do realize this is solved technology already on other vision systems like Waymo and MobilEye?

This is not a case against redundancy. It is par for the course for using vision.

Absolutely nobody these days advocates a Lidar only car. The question has been how much sensor redundancy you should have.

 

[electronblue]

What Tesla is basically advocating, ever since Musk said radar heating will be solved by not needed the front radar, is no sensor redundancy. The AP2/3 suite is basically a non-redundant sensor suite expect in a small front cone where three cameras merge which do have redundant vision.

That is a bold claim indeed for a Level 5 (no steering wheel required) robotaxi.

 

[StealthP3D]

This lack of sensor redundancy is something that has been nagging my mind about Tesla’s Level 5 AP2 suite since October 2016 and the same concern remains.

The lack of sensor redundancy has been nagging on your mind? What I find concerning is that you don't even work on Tesla's Autonomous Driving Program and yet you are worried about ridiculous concerns the leading people in the field are not. When you have three of the most brilliant minds in the field tell you LIDAR is a waste of time for autonomous driving, then you flip and say "well then, they should have 180-degree radar then" (as if you know what you're talking about and the experts at Tesla don't). It's comical. I really don't think you have much of a clue about how AI and neural nets actually work. The brilliant minds at Tesla are about 15 years ahead of you but you act as if you know more than they do while spreading your FUD.

 

[StealthP3D]

That is a bold claim indeed for a Level 5 (no steering wheel required) robotaxi.

Some people make bold claims and then back them up with bold achievements. I think this applies to Elon Musk more so than any other modern technologist. Did you see the twin boosters do a perfect landing a couple of weeks ago? It wasn't very long ago that people said it was impossible.

Personally, I don't think it's that bold of a claim. Radar is more of a training tool until they have the neural net trained. The AI get's most of it's distance information from the 3D model created by the camera. The radar assists in training the neural net and validating it's distance measurements. Once it's fully trained, the radar is not necessary. Just think if humans required radar to drive a car safely, LOL!

 

[electronblue]

The lack of sensor redundancy has been nagging on your mind? What I find concerning is that you don't even work on Tesla's Autonomous Driving Program and yet you are worried about ridiculous concerns the leading people in the field are not. When you have three of the most brilliant minds in the field tell you LIDAR is a waste of time for autonomous driving, then you flip and say "well then, they should have 180-degree radar then" (as if you know what you're talking about and the experts at Tesla don't). It's comical. I really don't think you have much of a clue about how AI and neural nets actually work. The brilliant minds at Tesla are about 15 years ahead of you but you act as if you know more than they do while spreading your FUD.

I don’t even work on Tesla's Autonomous Driving Program... what kind of a weird requirement is that for having an opinion?

Of course it is true I don’t work for any autonomous driving program. I am having a technology conversation on an enthusiast forum. It is quite normal to have views on technology in that context.

I am not flipping at all. I have always held this notion about AP2 suites lack of redundancy in 360 Lidar and radar, as well as lack of visual redundancy towards the sides and back. I think these are pretty common sense concerns.

That said also many other industry experts of course are advocating for 360 degree redundant sensors and Tesla is the lone exception to that so it is not really just my concern. You are ignoring great many real experts in the field if you ignore the calls for redundancy.

As for understanding how neural nets work I think I have better than average consumer understanding of them, probably better than you but then I don’t know you so I can’t say. Again the field is very divided on what can and should be achieved with vision NNs alone. After all NNs are used on radar and Lidar inputs as well.

 

[electronblue]

Some people make bold claims and then back them up with bold achievements. I think this applies to Elon Musk more so than any other modern technologist. Did you see the twin boosters do a perfect landing a couple of weeks ago? It wasn't very long ago that people said it was impossible.

Personally, I don't think it's that bold of a claim. Radar is more of a training tool until they have the neural net trained. The AI get's most of it's distance information from the 3D model created by the camera. The radar assists in training the neural net and validating it's distance measurements. Once it's fully trained, the radar is not necessary. Just think if humans required radar to drive a car safely, LOL!

I never considered SpaceX’s goals impossible.

SpaceX uses Lidar after all! Just a self-deprecating joke, don’t take it too seriously. But in seriousness, I never considered what SpaceX did impossible nor did I ever consider Tesla’s BEV mission impossible. I actually believe in the Mars idea too.

My concerns are specific to what is claimed (Level 5 no geofence robotaxi) of the AP2 suite with FSD computer. As I said, I have been wondering if this one is Elon’s rare mistake. (And Elon/Tesla has overpromised on hardware before too eg P85D etc.)

But he doubled down on it now so I guess not. Looking forward to it in my car.

 

[StealthP3D]

SpaceX uses Lidar after all!

Yes, SpaceX uses a custom developed LIDAR system for docking in space. Elon Musk championed the use of LIDAR for this application but says it's not suited to autonomous driving. It's not clear to me why you think you know more than he does. He's obviously a LIDAR expert.

Looking forward to it in my car.

I think you will be pleasantly surprised with how rapidly it improves.

Which Tesla do you own? And do mind sharing your location? "Earth" is pretty broad considering it covers every single human that ever existed.

 

[StealthP3D]

I don’t even work on Tesla's Autonomous Driving Program... what kind of a weird requirement is that for having an opinion?

It's not a requirement for having an opinion. But an opinion is not a very meaningful thing if it's not an informed opinion. If you worked on Tesla's Autonomous Driving Program, your opinion would be much more informed about what they are doing and why they don't need LIDAR for autonomy.

 

[electronblue]

Yes, SpaceX uses a custom developed LIDAR system for docking in space. Elon Musk championed the use of LIDAR for this application but says it's not suited to autonomous driving. It's not clear to me why you think you know more than he does. He's obviously a LIDAR expert.

It was just a joke.

But to ponder your serious point: Musk very likely knows more about Lidar than I do but he is also very alone in the autonomous driving industry on that point — and he is now pot committed to no Lidar so there is also a financial bias on his side. So listening to Elon on that point would require ignoring a whole bunch of other experts.

Therefore all we have is our own opinion based on whatever inputs and experiences we have gathered and analyzed.

I see very little redundancy in the AP2/3 sensor suite and for me Tesla didn’t adequately explain this away in the presentation nor at any time on or since October 2016. I’m not sure Lidar in particular is a must at all (there may well be many mixes of sensors that work great) but I would feel more comfortable if there was more redundancy overall.

What I do know is that the reasons for dissing on Lidar in the presentation did not convince me at all. Of course they don’t need to convince me. I’m irrelevant. I’m just it telling like I see it — my opinion.

I think you will be pleasantly surprised with how rapidly it improves.

I hope I am because I have not been for the first two and a half years.

Which Tesla do you own?

An AP2 Model X.

 

[electronblue]

It's not a requirement for having an opinion. But an opinion is not a very meaningful thing if it's not an informed opinion. If you worked on Tesla's Autonomous Driving Program, your opinion would be much more informed about what they are doing and why they don't need LIDAR for autonomy.

That would mean taking comments from companies at face value as it would render all outside views irrelevant.

I think that is not a wise strategy for best accuracy... for obvious reasons.

 

[StealthP3D]

An AP2 Model X.

Nice. And where do you live? It helps to put your comments in context.

 

[electronblue]

Nice. And where do you live? It helps to put your comments in context.

I don’t matter but I would accept Level 5 as delivered if Tesla does Level 5 no geofence on AP2 in continental U.S. I would say that is delivered.

 

[Bladerskb]

One interesting thing Karpathy said in his presentation today was Tesla vision NN is detecting small movements of cars and pedestrains (body languages) to predict their intentions like changing lanes or stepping out of the curb. Only vision system has the resolution to do that but Lidar does not. And these are all trained automatically with cars in the shadow mode. The machine will make predictions and than check with what actually happened. This is what Levandowski described as the main area to work today and Tesla already has it solved.

Mobileye and Waymo already solved it years ago.
Props to Tesla for catching up.

 

[diplomat33]

Mobileye and Waymo already solved it years ago.
Props to Tesla for catching up.

Sure Mobileye and Waymo are ahead in terms of tech but Tesla is catching up fast. Yesterday, Tesla proved that they have excellent hardware, very good camera vision, they don't need LIDAR, and they have the fleet learning. And Tesla can deploy their software to more cars faster.

By the way, your signature is super annoying and obnoxious.

 

[StealthP3D]

I don’t matter but I would accept Level 5 as delivered if Tesla does Level 5 no geofence on AP2 in continental U.S. I would say that is delivered.

It troubles me that you feel you need to keep your location confidential.