Will Tesla ever do LIDAR?

[diplomat33]

We all know that Musk has been quite adamant that he is against LIDAR. And a few years back, I think it was somewhat understandable. Back then, LIDAR was clumsy and expensive. There was just no way that Tesla could afford to put LIDAR in every car they sell, not to mention the issue of ruining the aerodynamics and aesthetics of the cars with a LIDAR tower on the roof. So I think back then, it made sense for Tesla to go the camera vision only route. After all, if they could manage to achieve the same result with camera vision only for a fraction of the cost of LIDAR, why not?

But today, these problems with LIDAR are pretty much solved. LIDAR is cheaper and smaller. Tesla could integrate LIDAR in the car in a way that does not ruin the aerodynamics or the aesthetics of the car at a much more affordable cost. Plus, there is no question that even if Tesla does manage to achieve great things with camera vision alone, LIDAR would offer more redundancy and make true Full Self-Driving much more robust. In other words, even if camera vision works, why not have that extra redundancy of camera vision, radar and LIDAR to give the car an even fuller picture of the environment to make FSD even better? There is no downside. So I am thinking that Tesla will eventually cave in a few years and add LIDAR to the FSD hardware.

Thoughts?

 

[Tam]

...Tesla will eventually cave...

I don't see how. They are too cheap and only provide one radar!

Once they change their mind and get more radars maybe I'll dream about the next step: LIDAR

 

[Barklikeadog]

if they ever want FSD they will

 

[iTech]

No

 

[BoerumHill]

Wish casting. IMO

 

[ewoodrick]

Why? Remember, the Tesla has over 9 eyes, you only have two. You do just as well.

The issue with FSD isn't really the detection, it's what you do once you see it all.

 

[S4WRXTTCS]

My belief is Lidar will become a regulatory requirement in some countries for any L3 driving or above.

So Tesla will be forced to include it.

 

[S4WRXTTCS]

Why? Remember, the Tesla has over 9 eyes, you only have two. You do just as well.

The issue with FSD isn't really the detection, it's what you do once you see it all.

9 eyes without eye lids.

Good luck seeing for very long without eyelids.

Especially the rear eye lid. I'm constantly cleaning the poor thing because it has no way to clean itself.

I also think my 2 eyes are better than it's 9 eyes because at least I can move my eyes around, and look down. Those 9 eyes don't prevent it from getting bitten because it can't see its feet.

It's also not just the eyes that matter, but the processing unit that it goes to. The 9 eyes don't seem to help it prevent it from crashing into things. Yeah, sure it just needs a better neural network. The same argument that was used last year, and the year before.

How many more fire trucks must be injured until people will realize that maybe this whole radar+camera isn't working?

 

[365gtb4]

Lidar doesn't work in the rain, fog or snow where radar does.

 

[Barklikeadog]

9 eyes without eye lids.

Good luck seeing for very long without eyelids.

Especially the rear eye lid. I'm constantly cleaning the poor thing because it has no way to clean itself.

I also think my 2 eyes are better than it's 9 eyes because at least I can move my eyes around, and look down. Those 9 eyes don't prevent it from getting bitten because it can't see its feet.

It's also not just the eyes that matter, but the processing unit that it goes to. The 9 eyes don't seem to help it prevent it from crashing into things. Yeah, sure it just needs a better neural network. The same argument that was used last year, and the year before.

How many more fire trucks must be injured until people will realize that maybe this whole radar+camera isn't working?

they like to brag about all the new options. why not come out and brag: it can tell the difference between a semi trailer and the sky!
probably because it still can't.

 

[S4WRXTTCS]

Lidar doesn't work in the rain, fog or snow where radar does.

I wouldn't call it a binary on/off thing.

Yes, it's true that rain, fog, snow, dust, etc can reduce the effectiveness of Lidar.

But, that doesn't immediately rule it out. Plus there are developments all the time in improving the ability of Lidar to work in adverse conditions. All you have to do is google "Will Lidar work in fog" to get videos various systems.

In any case the name of the game is sensor fusion, and redundancy.

All the AP deaths that I'm aware of were in conditions where Lidar would have operated optimally.

 

[S4WRXTTCS]

they like to brag about all the new options. why not come out and brag: it can tell the difference between a semi trailer and the sky!
probably because it still can't.

verygreen actually posted a video recently of a cross traffic situation with a semi trailer.

The results were mixed.

The neural network clearly identified the trailer in most frames, but labeled the area under the trailer as drivable.
The system identified it as "being in the way" in the wide angle camera view, but not the normal view or the narrow view.

So far no one can come to any conclusions on whether AP detected the trailer in the latest fatality accident. We don't even know if the driver had AP on.

What we do know is whatever Tesla is doing when it comes to crash prevention isn't good enough.

Even yesterday a Model S was involved in a pile up. You'd expect it to be one of the victims where it was rear ended, but no it rear ended someone else. All that technology, and it still rear ends a car in light fog (there was about 100ft of visibility at the time).

Quite frankly Tesla won't be able to go without Lidar for more than another 2-3 years.

Audi has it in the A8
Supposedly Audi will have it in the E-tron, but it's too early to tell

Lidar will continue to get cheaper while it improves in capabilities.

What's funny is Tesla will be half the reason for Lidar. What Autopilot has clearly shown is Radar+Camera doesn't work, and it will continue to happen. Why?

Radar has serious resolution issues with stopped objects. It's much more suited for tracking moving objects.

A Camera relies on a Neural Network to detect what it's looking at. This means the Neural network has to be trained on ALL the possible things that might be on a road. You can have pedestrian detection, but then the pedestrian might be carrying something that will throw the neural network off.

Increasing the risk of hitting something or someone just won't be worth the $500 to $1K per vehicle saved.

 

[StealthP3D]

No, I don't see Tesla moving to LIDAR. Sure, it would provide more data but that data has to be processed to be of any use. If they needed more data to process to able to self-drive safely, they might add LIDAR or, more likely, more cameras, but I don't see them doing this as the current crop of cameras show everything necessary to drive safely.

Sometimes less is more. It's not how much data you have, it's what you do with it.

 

[S4WRXTTCS]

No, I don't see Tesla moving to LIDAR. Sure, it would provide more data but that data has to be processed to be of any use. If they needed more data to process to able to self-drive safely, they might add LIDAR or, more likely, more cameras, but I don't see them doing this as the current crop of cameras show everything necessary to drive safely.

Sometimes less is more. It's not how much data you have, it's what you do with it.

What about rear radar?

My vote is in 2-3 years we'll see the Sensor Suite of new Tesla vehicles incorporating rear corner radars, front Lidar (in addition to the all cameras and the Radar) along with 360 degree down facing cameras.

Where those things will be pretty standard among L3 vehicles that also include self parking capabilities.

If they don't I'll probably buy an Audi or Porsche EV that does have these things.

The limited processing power in comparison to that of a human will mean that a self-driving car needs more data from a variety of sensors. We're also asking it to drive at probably 10X the safety of an average driver while having considerably less processing capability.

Right now the Audi A8 (that roughly has the sensor suite I described) is undergoing the approval process for limited L3 driving in Germany. It will set the precedence for other vehicles.

Regulatory stuff is often more political than technical. Sometimes there will be requirements that might not make sense. So even if you're in the "doesn't need Lidar" camp like you, or in the "definitely needs Lidar" camp like me it won't make any difference. It won't be up to us, but people who set the regulatory requirements in various countries.

 

[StealthP3D]

What about rear radar?

My vote is in 2-3 years we'll see the Sensor Suite of new Tesla vehicles incorporating rear corner radars, front Lidar (in addition to the all cameras and the Radar) along with 360 degree down facing cameras.

Where those things will be pretty standard among L3 vehicles that also include self parking capabilities.

If they don't I'll probably buy an Audi or Porsche EV that does have these things.

The limited processing power in comparison to that of a human will mean that a self-driving car needs more data from a variety of sensors. We're also asking it to drive at probably 10X the safety of an average driver while having considerably less processing capability.

Right now the Audi A8 (that roughly has the sensor suite I described) is undergoing the approval process for limited L3 driving in Germany. It will set the precedence for other vehicles.

Regulatory stuff is often more political than technical. Sometimes there will be requirements that might not make sense. So even if you're in the "doesn't need Lidar" camp it won't make a lick of difference.

Well, we disagree on this. Regulators will not regulate hardware, they will regulate capabilities. And LIDAR capabilities are dismal when you need it the most, in snow, sleet and heavy rain. But I don't expect a LIDAR fanboy to agree, they think LIDAR is the holy grail of self-driving.

 

[S4WRXTTCS]

Well, we disagree on this. Regulators will not regulate hardware, they will regulate capabilities. And LIDAR capabilities are dismal when you need it the most, in snow, sleet and heavy rain. But I don't expect a LIDAR fanboy to agree, they think LIDAR is the holy grail of self-driving.

I wouldn't call myself a Lidar fanboy.

Mostly I'm someone who pulls out their hair while training neural networks because they never work as well as I want them too.

So let's just say I'm a sensor fusion fanboy.

Regulators will definitely regulate redundancy.

It's all about making a system safe in case of failures and you never want a single fault failure crashing the car.

I don''t know if radar+camera will be enough.

Lidar is also great at being able to provide feedback for the training loop. Let's say for example that the Neural network was deficient at recognizing black people (it's been a problem on some DNN's caused by inadequate training data). A Lidar system can be used to notify the mothership every time the DNN misses a detection event. So the DNN would then get trained with new data that got uploaded when that trigger fired.

Having at least on vehicle model on the road with Lidar will give Tesla the ability to improve all the vehicles that have only vision+radar sensors. Plus with the right Lidar system they can build ultra high resolution maps. So they can avoid things like entering gore spaces.

We're still too far away to make the call on the whole lidar argument.

2019, and 2020 will be the time of rapid improvements in Tesla vision. So we'll know what they can pull off within a year or two. Whether they'll be able to get to L4 driving without Lidar.

If accidents keep happening as a result of frontal detection failures then Lidar will have to be used.

 

[StealthP3D]

I wouldn't call myself a Lidar fanboy.

Mostly I'm someone who pulls out their hair while training neural networks because they never work as well as I want them too.

So let's just say I'm a sensor fusion fanboy.

Regulators will definitely regulate redundancy.

It's all about making a system safe in case of failures and you never want a single fault failure crashing the car.

I don''t know if radar+camera will be enough.

Lidar is also great at being able to provide feedback for the training loop. Let's say for example that the Neural network was deficient at recognizing black people (it's been a problem on some DNN's caused by inadequate training data). A Lidar system can be used to notify the mothership every time the DNN misses a detection event.

So having at least on vehicle model on the road with Lidar will give Tesla an advantage. Plus with the right Lidar system they can build ultra high resolution maps. So they can avoid things like entering gore spaces.

We're still too far away to make the call on the whole lidar argument.

2019, and 2020 will be the time of rapid improvements in Tesla vision. So we'll know what they can pull off within a year or two. Whether they'll be able to get to L4 driving without Lidar.

If accidents keep happening as a result of frontal detection failures then Lidar will have to be used.

Haha, Like I said, I'll never agree with the LIDAR fanboys. It's simply not a required part of a safely functioning self-driving vehicle.

 

[heltok]

Lidars are awesome and so are cameras. Lidars are active, cameras are passive, lidars measure range, cameras measure light. Both have their usefulness and cost associated with them. For now Tesla has decided that a non-lidar based solution was the best approach. For the future we will see. If they decide to they will likely have done a good calculation of the benefits and costs.

There is a huge benefit for Tesla to keep the same hardware suite over a long time, it makes developing the software so much easier. At some point once vision is ~solved they might decide that, in order to push the systems performance to the next level, adding Lidar might be beneficial, but I don’t think that will be for a few more years at the least.

I guesstimate
1year: <10%
2year: <20%
5year: <50%
10year: >50%

Googles decided to bet heavily into lidar. They could afford to do it as Lidar hardware was not ready and they had to scale slower. But in the long run their solution might be better as they have a longer time to train on the end goal. We will see if early scale beats longer time with more hardware.

 

[StealthP3D]

We will see if early scale beats longer time with more hardware.

True, and I think that day is coming sooner than many realize. And don't forget, two camera lenses can do rangefinding also. We have optical rangefinders and they work very well. I have no idea whether Tesla is using the left and right forward facing cameras for stereo imaging but it wouldn't surprise me one bit.

 

[S4WRXTTCS]

Haha, Like I said, I'll never agree with the LIDAR fanboys. It's simply not a required part of a safely functioning self-driving vehicle.

It really comes down to what level of safety you're trying to obtain, and your time schedule.

In some ways Camera+Radar+Lidar is a cheat in that it enables higher levels of safety more quickly than trying to do Camera+radar only.

In the end we don't know the time frame as even a system that's ready for limited L3 (the Audi A8) doesn't have approval yet in Germany.

We also don't know the level of safety that will be required. I don't believe the 2X better than an average driver will be good enough, and I think 10X is what we should be aiming for. But, I have no idea what the public will find acceptable.