Dr. Mary (Missy) Louise Cummings should be opposed for NOT really being expert on FSD for cars

[Sporty]

Right, Musk has referred to LiDAR as a crutch that helps you get going fast but then you are stuck in Chandler for years.

PS : Personally, I don’t think LiDAR is the issue. The problem is reliance on HD Maps.

In the short term sure. But after vision driving works as well a human maybe adding LiDAR could improve it even more. Eg if LiDAR brings some new type of data into the system that wasn’t available before, in theory that could be leveraged intelligently to do better. (Like maybe drive in foggy conditions better). It might be hard to do which is why maybe you don’t start with this complexity. It also might not be worth it from a business sense. Ie what’s the value vs cost at that point in time.

Note I was trying to keep the “new data” generic as a concept. LiDAR seems to be a loaded word for ev cars. There are probably lots of inputs to consider but, for example, maybe incorporating nearby other data could be useful. Eg data from other cars. Obviously other cars would be more intermittent, but I can imagine useful scenarios.

And to stay on point.. these things should be on the back burner until the basics are accomplished. And I mean in a way a normal car driver would be comfortable using them.

 

[MontyFloyd]

I think most agree that eventually vision only should be able to do the job. The question I wonder about is will that be accomplished first or second?

Assuming the resources are available to handle the load of the additional data, it’s hard to argue that giving a system more (valid) data should negatively impact the quality of the resulting answer. I get how that can happen as it’s happened to me. But it’s a problem/limitation in the implementation and at best should be a wash.

I think T is vision only right now because there’s proof that it works, it’s cheaper on the per-car side, and perhaps there’s some intuitiveness behind it. Eg asking humans to develop a system that we more easily relate too. Just my opinion there. In the future I can imagine that adding other data types could increase the performance. But before doing that, make it basically work first.

You hit the nails on the head!

YES, the resources are available! HOWEVER they are multi core, multi CPU system needed to crunch far more than onboard computer can do. From what I read Tesla system uses object recognition, for every object, including items flying across the road.

There is a downside to more sensors, that is more data to process and how to value the data. Is the low resolution data more useful than high resolution color camera data? Too many sensors = CPU overload.

Cost is the ultimate factor. This is a guess, but all 8 cameras on a Tesla is the same cost of a single radar unit ($5000 I read).
LiDAR is on order of $50,000 for full system today (suite of 5 LiDAR scanners).

 

[Sporty]

You hit the nails on the head!

YES, the resources are available! HOWEVER they are multi core, multi CPU system needed to crunch far more than onboard computer can do. From what I read Tesla system uses object recognition, for every object, including items flying across the road.

There is a downside to more sensors, that is more data to process and how to value the data. Is the low resolution data more useful than high resolution color camera data? Too many sensors = CPU overload.

Cost is the ultimate factor. This is a guess, but all 8 cameras on a Tesla is the same cost of a single radar unit ($5000 I read).
LiDAR is on order of $50,000 for full system today (suite of 5 LiDAR scanners).

I was under the impression the cores on the car aren’t at capacity yet. Don’t know about the io / memory buses. In any case the “removing radar as an input and going vision only to remove phantom breaking” (to me) is a implementation problem. I’m fine with it but I think it’s more a matter of behind the scenes priorities.

I don’t know much about LiDAR. Did they solve the multiple LiDARs interference problems on the consumer units? Certainly $50k is a show stopper unless there’s no other way.

 

[MontyFloyd]

I was under the impression the cores on the car aren’t at capacity yet.

I expect the system on Tesla are balanced between input and processing capacity (with spare overhead).
What I was talking about was a more sophisticated system with more data inputs, and more fine analytical ability.

Don’t know about the io / memory buses. In any case the “removing radar as an input and going vision only to remove phantom breaking” (to me) is a implementation problem. I’m fine with it but I think it’s more a matter of behind the scenes priorities.

I don’t know much about LiDAR. Did they solve the multiple LiDARs interference problems on the consumer units? Certainly $50k is a show stopper unless there’s no other way.

Have not seen a "consumer" LiDAR car. Everything I have seen is private.
I have yet to seen any LiDAR equip car for general public, even the millionaires.

 

[Daniel in SD]

Have not seen a "consumer" LiDAR car. Everything I have seen is private.
I have yet to seen any LiDAR equip car for general public, even the millionaires.

2022 Mercedes EQS and S class with the "Drive Pilot" option are supposed to have LiDAR. I don't think it will be available in the US though.

Mercedes-Benz receives world's first internationally valid system approval for conditionally automated driving

Mercedes-Benz is the first automotive company in the world to meet the demanding legal requirements of UN-R157 for a Level 3 system.

media.mbusa.com

It's made my Valeo:

In a world-first, Valeo’s second-generation LiDAR will equip the new Mercedes-Benz S-Class, allowing it to reach level 3 automation

Valeo announces that the new Mercedes-Benz S-Class, capable of reaching Level 3 automation, is the first car in the world to be equipped with its LiDAR, the second-generation Valeo SCALA™.

www.valeo.com

 

[nvx1977]

Prolly worth posting here again, so everyone has at least a base understanding of how lidar and radar work. Higher-end versions will have better performance but quickly becomes cost-prohibitive.

Lidar: actively sending out lasers and capturing bounceback to scan surroundings of solid objects. Very accurate gauging distance and velocities. Cannot determine density of object (a grocery bag looks like a stone, etc). Cannot read signage, as it cannot detect color. Excellent at not being fooled by color contrast (shadows). Weak in poor weather due limitations of visible light frequencies.

Radar: similar to active nature of lidar, except photons are in the radio band instead of visible spectrum. Advantages: things that are translucent or opaque to visible light are transparent to radar, so it can see thru things like a leading car. Not affected by poor weather. Longer wavelength means lower resolution. Objects tend to look like point clouds, potential for false positives.

If cost were no object, I tend to think that sensor fusion between camera and lidar is more reliable than camera and radar. C/L combo gives you high res distance, velocity, and color, but cannot overcome heavy weather. If camera and lidar are rendered useless by weather, the use of a backup radar will not allow the car to drive autonomously because alone, radar does not have the requisite resolution to make any good perception decision.

 

[MontyFloyd]

Exactly

If cost were no object, I tend to think that sensor fusion between camera and lidar is more reliable than camera and radar.

Of course, but who can afford $150,000 Model 3? (for example)

One detail needs to note: Radar has a huge wavelength region (from mm to tens of meters). Need to have a small enough wavelength for accuracy, and focus enough to dismiss random signal returns, thus erroneous data.

 

[nvx1977]

Exactly


Of course, but who can afford $150,000 Model 3? (for example)

One detail needs to note: Radar has a huge wavelength region (from mm to tens of meters). Need to have a small enough wavelength for accuracy, and focus enough to dismiss random signal returns, thus erroneous data.

Yeah, at mm wavelengths, we're talking about the new-age 4D radar. If camera and lidar go down due to weather, you still can't drive autonomously because radar can't read signs. And 4D radar, like lidar, is still cost-prohibitive if the car is to be sold to an end-consumer.

 

[FloridaJohn]

Yeah, at mm wavelengths, we're talking about the new-age 4D radar. If camera and lidar go down due to weather, you still can't drive autonomously because radar can't read signs. And 4D radar, like lidar, is still cost-prohibitive if the car is to be sold to an end-consumer.

So what happens if autonomous driving comes down to using expensive sensors or not being able to achieve full autonomy?

Right now we are assuming that given enough time, vision can be made to work. But what if that turns out to not be possible?

 

[nvx1977]

So what happens if autonomous driving comes down to using expensive sensors or not being able to achieve full autonomy?

Right now we are assuming that given enough time, vision can be made to work. But what if that turns out to not be possible?

Then L4+ in consumer-owned cars doesn't happen, and L4+ may be limited to robotaxi companies who operate in limited areas where demand for that service is highest.

There's still a lot of value in L2 ADAS, and getting that much safer than humans doesn't seem as hard as "full autonomy." I put that in quotes because I think that means something different for everyone (can sleep in car, zero accidents, 100% of cars driving themselves and communicating with each other, handles every possible situation, even if a volcano pops up in the middle of the highway, etc)

 

[MontyFloyd]

So what happens if autonomous driving comes down to using expensive sensors or not being able to achieve full autonomy?

Right now we are assuming that given enough time, vision can be made to work. But what if that turns out to not be possible?

If the conditions are so bad the Li and RA sensors are not able, what makes you think a Eyeball Mk 1 connected to Carbon Neural Matrix Wetware with Smarts 1.0 do any better?

When conditions are that bad, pull over. I do.

From what I see with FSD, in nearly all weather the system works extremely good.

 

[JHCCAZ]

Cost is the ultimate factor. This is a guess, but all 8 cameras on a Tesla is the same cost of a single radar unit ($5000 I read).
LiDAR is on order of $50,000 for full system today (suite of 5 LiDAR scanners).

Respectfully, I think all of these cost numbers are way, way off - on the way-high side. Unless perhaps my browser refuses to render decimal points

I would believe an automotive radar unit, as used for collision-avoidance braking and adaptive cruise control in many cars (including radar-eqiipped Teslas) costs something like $50, perhaps $100 but not more in volume. This is not talking about possible future deployment of high-resolution millimeter-wave radar for external detection - though still that would not be in the several thousand dollar range if it's going to get deployed.

Tesla's 8-camera suite? Including housings, harnesses. heaters? My rough guess is a few hundred at most. $1k IMO sounds unreasonably high for what they do presently and for what could possibly be contemplated to go in a $25,000 Tesla upcoming. So the $5000 figure is, I'd say, 10x to 50x too high. (But on another topic, I'd be all for them to splurge just a little for more and higher-quality ones.)

Cost of Lidar is a bit more complex topic. The figures of tens of thousands of dollars are a holdover from early 360-degree electromechanical units built in limited volumes for prototyping, research and mission-critical cost-insensitive applications. Not for recently-designed, high-volume-targeted and primarily forward-facing Lidar arrays that you're likely to see pretty soon in production cars. I'll venture a guess that the latter might be several hundred dollars, maybe even higher at the moment but if so, with a supplier promise and a credible business plan to demonstrate a downward cost progression fairly quickly into the low hundreds. Otherwise they would not be seriously considered for consumer-vehicle ADAS use at all.

And by the way, these estimates do not imply the Elon was wrong, lying or misinformed when he made his decision and pronouncements about Lidar for FSD. Tesla had to make decisions in the 2016-2017 design time-frame to adopt and finalize a hardware suite that would be reasonable for cars with a planned highly-capable system to hit volume manufacturing starting in 2018. I'm pretty sure there was nothing (Lidar) available at any reasonable cost or production-readiness at that time. Given that, there was no realistic path for a Lidar-dependent FSD platform that could have achieved Elon's/Tesla's goal that the upcoming product generation (the cars we're driving right now) would be capable of autonomous driving. It is a related but different argument regarding whether Elon was mistaken, then and now, about Tesla's ability to achieve the goals of FSD without Lidar.

 

[MontyFloyd]

Respectfully, I think all of these cost numbers are way, way off - on the way-high side. Unless perhaps my browser refuses to render decimal points

The numbers quoted come from a trade journal our company subscribes to.
That issue is May 2021

The costs per system (not unit) matches other estimates I have seen, so YES, those are actual approximate costs for LiDAR system ($50,000).

If you do find substantially lower costs, post the source with quote.

 

[Bladerskb]

You hit the nails on the head!

YES, the resources are available! HOWEVER they are multi core, multi CPU system needed to crunch far more than onboard computer can do. From what I read Tesla system uses object recognition, for every object, including items flying across the road.

There is a downside to more sensors, that is more data to process and how to value the data. Is the low resolution data more useful than high resolution color camera data? Too many sensors = CPU overload.

Lol compute power is not the problem or an issue at all of AV today.

Cost is the ultimate factor. This is a guess, but all 8 cameras on a Tesla is the same cost of a single radar unit ($5000 I read).
LiDAR is on order of $50,000 for full system today (suite of 5 LiDAR scanners).

Lol this is why people need to get info from direct sources. Those are made up numbers.
The cameras on Tesla costs significantly below $10 a pop and the radar cost $50-$70.
Today a single unit of high resolution Lidar used in AV would cost you anywhere from $150-$1,000 depending on the scale.
The luminar iris lidar at scale cost $500 a pop, 5 of those would be $2500k.

Have not seen a "consumer" LiDAR car. Everything I have seen is private.
I have yet to seen any LiDAR equip car for general public, even the millionaires.

There are plenty of lidar cars, you just ain't looking. Xpeng P5 with two livox lidars have already sold thousands.
The NIO ET7 with innuvision lidar is starting deliveries in 2 months, followed by the ET5. Then there's the BMW IX with innoviz also starting deliveries in acouple months. Then Huawei Arcfox Hi which started deliveries late December. The list goes on.

 

[MontyFloyd]

Lol this is why people need to get info from direct sources. Those are made up numbers.
The cameras on Tesla costs significantly below $10 a pop and the radar cost $50-$70.

I never quoted what the prices of the Tesla cameras.
I should have sourced the price of radar (my bad), but a quick search shows eBay asks around $200 (you can search more).
Read article a few months ago and I probably am mistaken the cost estimate.

Today a single unit of high resolution Lidar used in AV would cost you anywhere from $150-$1,000 depending on the scale.
The luminar iris lidar at scale cost $500 a pop, 5 of those would be $2500k.

Again, I was quoting price of system, FROM A TRADE JOURNAL which our company uses in our work.

There are plenty of lidar cars, you just ain't looking. Xpeng P5 with two livox lidars have already sold thousands.
The NIO ET7 with innuvision lidar is starting deliveries in 2 months, followed by the ET5. Then there's the BMW IX with innoviz also starting deliveries in acouple months. Then Huawei Arcfox Hi which started deliveries late December. The list goes on.

Links to them, please.

Bottom line, I put an actual quote down from an industry journal with HARD numbers.

Which is very rich because your fist line says "this is why people need to get info from direct sources" which is exactly what I did IRT LiDAR.
LOL

 

[azred]

Interesting discussion and I don’t have technical expertise that several posters seem to have. However, I have FSD Beta on a Model 3 and an “old“ 12/2016 S and my non-technical concern is that I may need better and/or more cameras. Both cars are scared to death of unprotected lefts, which suggests they simply can’t see those situations very well. And of course currently it doesn’t take much rain for FSD Beta to balk. I suppose these issues could simply be Tesla taking a careful approach…or not.

 

[MontyFloyd]

There are plenty of lidar cars, you just ain't looking. Xpeng P5 with two livox lidars have already sold thousands.
The NIO ET7 with innuvision lidar is starting deliveries in 2 months, followed by the ET5. Then there's the BMW IX with innoviz also starting deliveries in acouple months. Then Huawei Arcfox Hi which started deliveries late December. The list goes on.

Your definition of "plenty" is unusual.
You only list one that is in production and a few others just starting Jan 2022 or coming in months.
As for one in production:
Xpeng P5 came out 2 months ago, 244 sold in September (probably 10K ~ 20K now) but only China and Europe.
12 ultrasonic sensors, 13 cameras, 1x 5mm RADAR and 2x LiDAR (!!!), all for a price between $25,000 to $35,000.

Details on P5 LiDAR is limited, looks like a narrow arc or perhaps a cone of scanning, some closeups here.
Interestingly this article also states LiDAR "Not long ago, a lidar sensor cost about $75,000." but I am sure that talks about systems like the WayMo suite of 4 or more 90* to 360* wide scanning systems.

How Xpeng can profitably sell a car with more than 2x the sensor array of a any Tesla for $10k less is, well, add your adjective here.

 

[Daniel in SD]

Cellphones and robot vacuums have LIDAR these days. How much did the first digital cameras cost? The first RADAR systems? The price of new tech goes down over time as volume goes up. What part of a LiDAR system do you think is fundamentally expensive?
Volvo and Mercedes halve announced models with LIDAR too.

 

[wooter]

Cellphones and robot vacuums have LIDAR these days. How much did the first digital cameras cost? The first RADAR systems? The price of new tech goes down over time as volume goes up. What part of a LiDAR system do you think is fundamentally expensive?
Volvo and Mercedes halve announced models with LIDAR too.

Cellphones, robot vacuums and cars have very low resolution LIDAR.

The LIDARs that Waymo is using are of a higher resolution, and then still have lower pixel density than a camera.

On top of that, with a camera you can take a picture, so the possible volume is very high because camera's fit in laptops, smartphones, tablets, cars, security camera's, dash cams, ... But with LIDAR you can't take a picture. You can only create a 3D point cloud without color. So the possible volume is less, since in what scenarios does your future smartphone, tablet, laptop, security cameras, dash cams need 3D point clouds without color?

 

[Daniel in SD]

Cellphones, robot vacuums and cars have very low resolution LIDAR.

The LIDARs that Waymo is using are of a higher resolution, and then still have lower pixel density than a camera.

On top of that, with a camera you can take a picture, so the possible volume is very high because camera's fit in laptops, smartphones, tablets, cars, security camera's, dash cams, ... But with LIDAR you can't take a picture. You can only create a 3D point cloud without color. So the possible volume is less, since in what scenarios does your future smartphone, tablet, laptop, security cameras, dash cams need 3D point clouds without color?

Yep. That’s why talking about the cost of automotive LIDAR is somewhat silly until someone actually finds a mass market use for it. How much did radar units cost before they started using them for adaptive cruise control? My guess is that they were not cheap.
It remains to be seen whether LIDAR will enable L3 highway systems but if it does I expect there will be enough volume to drive costs down. Hopefully the Volvo and Mercedes systems will actually get deployed this year.