Luminar’s largest customer: Tesla

[enemji]

This begs the question of why Tesla nevertheless purchased $2.1 million of Lidar in Q1 2024, if they supposedly don't need them.

My hunch is that "need" is being used in a fairly extreme way here, in the sense that I don't "need" coffee, or wifi, or underwear. I don't doubt that Tesla could eventually theoretically solve FSD (U) with pure vision, but I think they could solve it much sooner with vision+lidar+radar (which is now far more economical than when they started working on autonomy in 2016), and I suspect they're beginning to realize that.

a robotaxi would require a second source of validation during an investigation that resulted from an accident.

 

[diplomat33]

This begs the question of why Tesla nevertheless purchased $2.1 million of Lidar in Q1 2024, if they supposedly don't need them.

One possibility is that Elon is simply lying and Tesla does still need lidar for ground truth validation but he does not want to admit it because it would undermine his narrative about how far ahead Tesla is with the end-to-end approach. Another possibility is that Tesla is planning to add lidar to the robotaxi vehicle (not because they need it per se but simply for extra redundancy). So again, Elon will save face by saying that the lidar is simply for redundancy or to make regulators happy or something. A third possibility is that the lidar is not for cars at all but rather for something else like the Optimus Bot or something.

 

[Ben W]

a robotaxi would require a second source of validation during an investigation that resulted from an accident.

And the rest of the fleet wouldn't?

In other words, if lidar is a regulatory requirement for Robotaxi, that's likely the nail in the coffin for HW3/HW4 cars ever achieving Robotaxi capability.

This, I suspect, is the real reason why Elon remains so dead-set on making pure vision work. His “1 million Robotaxi-capable cars on the road by 2020” comment is going to come back to haunt him if that fleet proves not to be Robotaxi capable, and if it can be shown that millions of consumers bought the cars (and hundreds of thousands bought FSD) based on that promise.

 

[spacecoin]

And the rest of the fleet wouldn't?

In other words, if lidar is a regulatory requirement for Robotaxi, that's the nail in the coffin for HW3/HW4 cars ever achieving Robotaxi capability.

Regulators shouldn't legislate based on certain technology or solutions, they should legislate on performance (ie safety) and they likely will.

Current science is the nail in the coffin for hw3/hw4 autonomy. Unless there is a few major breakthroughs that system will never be autonomous at meaningful speeds.

 

[stopcrazypp]

And the rest of the fleet wouldn't?

In other words, if lidar is a regulatory requirement for Robotaxi, that's likely the nail in the coffin for HW3/HW4 cars ever achieving Robotaxi capability.

This, I suspect, is the real reason why Elon remains so dead-set on making pure vision work. His “1 million Robotaxi-capable cars on the road by 2020” comment is going to come back to haunt him if that fleet proves not to be Robotaxi capable, and if it can be shown that millions of consumers bought the cars (and hundreds of thousands bought FSD) based on that promise.

On the other hand, if it becomes a regulatory requirement, that gives Tesla a legal out given Tesla had a regulatory disclaimer from the start.

This was mentioned in other threads, but since 2019, Tesla have removed all references to L3/L4 behavior or possible robotaxi from the FSD order page, so Tesla had long been hedging. The awkward thing is that the earlier FSD buyers were actually promised more on the order page.

 

[spacecoin]

On the other hand, if it becomes a regulatory requirement, that gives Tesla a legal out given Tesla had a regulatory disclaimer from the start.

I find that outcome incredibly unlikely.

 

[Ben W]

Regulators shouldn't legislate based on certain technology or solutions, they should legislate on performance (ie safety) and they likely will.

They shouldn’t, but they do. Anti-lock brakes are required by law, regardless of stopping distance without them. So are side-view mirrors (to Tesla’s dismay), despite digital alternatives that could provide equal or better visibility and reliability and improved fuel efficiency. So it wouldn’t put it past regulators to require lidar for robotaxis, on the grounds that any autonomous car, regardless how safe without lidar, could be made that much safer with it, and that adding lidar is no longer a significant economic burden.

Current science is the nail in the coffin for hw3/hw4 autonomy. Unless there is a few major breakthroughs that system will never be autonomous at meaningful speeds.

Ironically, speed is not the issue. Highway L4 at 70mph is MUCH easier than city-streets L4 at 30mph. The limiting factor for pure vision is the environment: dirty cars/cameras, raindrops on the lens, sun glare, fog, condensation, very low light. L4 has to work for the dirtiest cars in the fleet, not just the cleanest. A daily car wash (and only good weather) is not a reasonable requirement.

 

[spacecoin]

Highway L4 at 70mph is MUCH easier than city-streets L4 at 30mph.

Not according to every robotaxi company on the planet. Different challenges and higher stakes game tho. You can't just stop and wait for remote assistance on the highway. That's illegal (and unsafe).

 

[Daniel in SD]

Anti-lock brakes are required by law, regardless of stopping distance without them.

anti lock brakes are required because they allow you to turn with brakes fully applied. When they were first developed I don’t think they even decreased stopping distance (because they didn’t modulate all the wheels separately and they had slow modulation)

I don’t think they’ll ever require specific sensor technology for AVs.

 

[Ben W]

Not according to every robotaxi on the planet. Different challenges and higher stakes game tho.

City-streets is essential for Robotaxi to solve; highway is not, hence all the attention on city streets. (You can’t get from curbside to highway without navigating city-streets.) But there’s a reason Tesla Autopilot was available on highway for years before any city-streets functionality became available at all. Mercedes’ current industry-leading L3 offering is highway-only. Disagree on stakes necessarily being higher for highway; there are no pedestrians to run over or opposing traffic to crash into on divided highway, and Robotaxi won’t be excessively speeding, which is a significant cause of highway crashes/fatalities.

 

[aset]

Anti-lock brakes are a requirement because they do work. Lidar/radar efficiency is not substantiated by anything other than numerous complaints about their removal.

 

[Daniel in SD]

City-streets is essential for Robotaxi to solve; highway is not, hence all the attention on city streets. (You can’t get from curbside to highway without navigating city-streets.) But there’s a reason Tesla Autopilot was available on highway for years before any city-streets functionality became available at all. Mercedes’ current industry-leading L3 offering is highway-only. Disagree on stakes necessarily being higher for highway; there are no pedestrians to run over or opposing traffic to crash into on divided highway, and Robotaxi won’t be excessively speeding, which is a significant cause of highway crashes/fatalities.

Highways are simpler most of the time. The problem is the times when there are pedestrians on the highway or there is someone driving the wrong way or there is who know what falling off vehicles or overturned trucks. Trying to get 100 million miles between fatal collisions with all the crazy stuff that happens is not easy. Higher speeds mean that the system must be much better at prediction.

 

[stopcrazypp]

City-streets is essential for Robotaxi to solve; highway is not, hence all the attention on city streets. (You can’t get from curbside to highway without navigating city-streets.) But there’s a reason Tesla Autopilot was available on highway for years before any city-streets functionality became available at all. Mercedes’ current industry-leading L3 offering is highway-only. Disagree on stakes necessarily being higher for highway; there are no pedestrians to run over or opposing traffic to crash into on divided highway, and Robotaxi won’t be excessively speeding, which is a significant cause of highway crashes/fatalities.

I think it has more to do with the fact there is a higher risk of a fatality at higher speeds. That's why even existing L3 solutions try to limit the amount of users (Acura only has it available on a limited number of leased high end vehicles, Mercedes only has it on S-Class and EQS with a special option package and subscription).

From below, fatal accidents at speeds higher than 35mph make up 79% of fatal accidents. Of that it's split 27% for 40-50mph, 47% for 55+mph.

Fatality Facts 2021: Urban/rural comparison

A summary of fatality statistics about urban/rural comparison compiled by IIHS from 2021 Fatality Analysis Reporting System (FARS) data.

www.iihs.org

 

[Ben W]

Highways are simpler most of the time. The problem is the times when there are pedestrians on the highway or there is someone driving the wrong way or there is who know what falling off vehicles or overturned trucks. Trying to get 100 million miles between fatal collisions with all the crazy stuff that happens is not easy. Higher speeds mean that the system must be much better at prediction.

If your point is that highways are easier for L2/L3 but harder for L4, that makes some sense. The metric should probably be at-fault fatalities, not all fatalities; hitting a wrong-way freeway driver is hard to blame on the right-way car. Crazy stuff happens on city streets too though, and latency / reaction time is MUCH faster for the FSD computer than a human, so even highway speed is probably not the limiting factor. Granted we are not there yet; my Model Y on FSD v12.3.3 / v11 (HW3) recently ran over a 5-gallon paint bucket on the highway (fortunately empty, no damage) when I took my eyes off the road for a split second. But this is solvable, I think, and far easier than many city-streets scenarios. I expect the E2E highway stack to make tremendous improvements over v11 here.

 

[Unclepaul]

Tesla could also be using the LIDAR in production line and testing.

 

[Daniel in SD]

If your point is that highways are easier for L2/L3 but harder for L4, that makes some sense. The metric should probably be at-fault fatalities, not all fatalities; hitting a wrong-way freeway driver is hard to blame on the right-way car. Crazy stuff happens on city streets too though, and latency / reaction time is MUCH faster for the FSD computer than a human, so even highway speed is probably not the limiting factor. Granted we are not there yet; my Model Y on FSD v12.3.3 / v11 (HW3) recently ran over a 5-gallon paint bucket on the highway (fortunately empty, no damage) when I took my eyes off the road for a split second. But this is solvable, I think, and far easier than many city-streets scenarios. I expect the E2E highway stack to make tremendous improvements over v11 here.

That would be a horrible metric. So we deploy AVs and the number of fatalities increases? If I’m a passenger an AV I don’t care whose fault it is when it crashes.
Highways are easier for L3 if it’s just a traffic jam assist features. I’m skeptical that they’ll get full speed with no lead car working.

 

[enemji]

my Model Y on FSD v12.3.3 / v11 (HW3) recently ran over a 5-gallon paint bucket on the highway (fortunately empty, no damage)

and that is the balancing equation. More often than not, applying a reaction such as hitting the brakes or switching lanes at the last moment are much more expensive (cost of human life) than replacing a bumper of a fender

 

[Ben W]

and that is the balancing equation. More often than not, applying a reaction such as hitting the brakes or switching lanes at the last moment are much more expensive (cost of human life) than replacing a bumper of a fender

The car in front of me (which FSD was following) successfully maneuvered to avoid it. My car should have as well. The advantage of FSD’s 360-degree situational awareness and fast reaction time is that it can swerve perfectly safely in such situations, if trained to do so, and if there is a way to do so. This is not a trolley problem, nor are the vast majority of such situations. I don’t agree that “more often than not” is an accurate characterization.

 

[Ben W]

You can't just stop and wait for remote assistance on the highway. That's illegal (and unsafe).

Sure you can, if the car can’t safely be driven. (E.g. mechanical failure or flat tire; but presumably a fault in the autonomy system would be treated similarly.) That’s what the shoulders are for. And it’s why L4 cars-in-training on highway must have human backup drivers, until they’ve demonstrated a near-perfect level of reliability. (City streets are no different.) Production L4 cars must be able to slow and stop safely and legally in these situations EVERY time; otherwise, they’re just sparkling L3.

 

[spacecoin]

Sure you can, if the car can’t safely be driven. (E.g. mechanical failure or flat tire; but presumably a fault in the autonomy system would be treated similarly.)

Deploying a (poor) L4 system shouldn't qualify in that category imho. RA isn't a failure, it's a system limitation. I'm pretty sure regulators will be concerned quickly with a bunch of Robotaxis stranded on limited access highways.

(moderator edit)