Autonomous Car Progress

[AlanSubie4Life]

To me it seems like there are just too many insane and unusual situations that can arise in traffic jams, that a safe system there is also likely to be capable of L4. I just think any L3 system with less-than-desirable failure modes with the 10-second warning is unlikely to be able to properly deal with these unusual situations (which could arise very quickly even in a 30mph traffic jam).

Aside from the unusual situations…say they limit it to 30mph to mitigate that risk (I don’t think it would completely mitigate risk of very unusual situations (which happen all the time of course) though)? What happens when suddenly speeds go from 10mph to 80mph? Is it going to plug along at 30mph for 10 seconds hassling the driver to take over? Seems undesirable.

 

[daniel]

I don't think that is necessarily true. I think it would depend on the type of L3 and ODD. For example, a "highway traffic jam" L3 would not be good enough to be city L4. So a "highway traffic jam" L3 would not mean that you've almost solved L4.

Also, L3 can get away with not solving a lot of FSD problems because you can just throw it back to the human driver if needed. For example, L3 does not need to do traffic light detection since you could have the car just request the driver take over every time you approach a traffic light. But you can't do that with L4. L4 needs to solve FSD problems like traffic lights because you can't just pull over to the side of the road for every traffic light.

Technically, an L4 car could have the same ODD as an L3 car. (That's one of my biggest complaints about the SAE definitions: the ODD can be so narrow it's useless. An L3 system that requires you to take over at every traffic light or stop sign is, IMO useless. Note also, that the L3 car has to give you ten seconds (not by definition, but to be realistic) and that's probably not happening if it cannot reliably deal with traffic lights.) Those robotaxis in Chandler, AZ, are Level 4, no disputing it, but they're useless unless you need a ride share entirely inside those few square miles within Chandler.

Note L4 does not require the car to be able to pull to the side of the road, stopping in its current lane is an acceptable fall back.

The difference is for L4 there is no take over time. The car must be able to handle all situations it may come across and safely fall back even if the driver never responds.

Put in other words if that L3 car fails spectacularly after the buffer time (10 seconds in this case) that is still within its design constraints, but that's not the case for L4. Another way to look at it is that for L4, the request for take over is optional for the driver to respond, but not so for L3. A standard used by journalists is if the "driver" can sleep in the car: you can in L4, you can't in L3.

An L4 car has to be able to stop safely. That will very seldom be the middle of the road. I cannot imagine a situation where it would be, unless the road is completely blocked.

It's true that the defined constraints are different, what I'm saying (in agreement with others who have said it before me) is that a car capable of handling every situation that arises with less than ten seconds reaction time is probably capable, or very nearly capable, of letting the driver sleep in the back. Maybe for practical purposes we'd want more testing time/miles, but I suspect that the technical challenges of L3 are very nearly as difficult as those of L4.

I will give you that L3 in the city will be much more difficult than L3 on the highway, but that's the ODD, not the level.

I also think that ten seconds is the bare minimum. A really good L3 would give you 20 or 30 seconds. Maybe 5 or 10 seconds of a non-jarring tone before it really screeches at you and scares the bejesus out of you. And a car that can give you 30 seconds, could probably just as well pull over and wait for you to wake up.

 

[diplomat33]

Technically, an L4 car could have the same ODD as an L3 car. (That's one of my biggest complaints about the SAE definitions: the ODD can be so narrow it's useless. An L3 system that requires you to take over at every traffic light or stop sign is, IMO useless. Note also, that the L3 car has to give you ten seconds (not by definition, but to be realistic) and that's probably not happening if it cannot reliably deal with traffic lights.) Those robotaxis in Chandler, AZ, are Level 4, no disputing it, but they're useless unless you need a ride share entirely inside those few square miles within Chandler.

There are way too many possible ODDs. It would not be realistic for the SAE to try to define the ODDs for the different levels. And the SAE needs to make sure that the levels are broad enough to encompass any possible system that a vehicle might have. The SAE levels have to work for Tesla's AP, for Waymo's robotaxi, the autonomous pizza delivery robot or the autonomous airport shuttle etc... So the SAE classifies the levels based on who performs the DDT and DDT-fallback because that is very important to know. And usefulness is somewhat subjective. That's up to consumers to decide if an automated driving system that an automaker is selling is useful to them or not.

 

[daniel]

There are way too many possible ODDs. It would not be realistic for the SAE to try to define the ODDs for the different levels. And the SAE needs to make sure that the levels are broad enough to encompass any possible system that a vehicle might have. The SAE levels have to work for Tesla's AP, for Waymo's robotaxi, the autonomous pizza delivery robot or the autonomous airport shuttle etc... So the SAE classifies the levels based on who performs the DDT and DDT-fallback because that is very important to know. And usefulness is somewhat subjective. That's up to consumers to decide if an automated driving system that an automaker is selling is useful to them or not.

Yeah, we had this discussion before. I'm just saying that the SAE levels are useless to me. Waymo's car is Level 4. But since I don't live in Chandler it's of no use to me. I need to know if a car will work on the roads where I live, and no car company is going to publish a nation-wide map showing which roads their car will and will not drive on, or clear and precise promises concerning the weather limitations.

FWIW, I don't think Tesla ever says its cars are Level 2. It says you have to remain alert and ready to take over, which is the definition of Level 2, but it never mentions levels. I doubt that any car company will reference the SAE levels. They'll make vague claims, worded so as to give them plausible deniability in court, that seem to be promising more than they really are.

Websites like this are where we'll really find out what the cars can do.

 

[JHCCAZ]

Yeah, we had this discussion before. I'm just saying that the SAE levels are useless to me. Waymo's car is Level 4. But since I don't live in Chandler it's of no use to me. I need to know if a car will work on the roads where I live, and no car company is going to publish a nation-wide map showing which roads their car will and will not drive on, or clear and precise promises concerning the weather limitations.

FWIW, I don't think Tesla ever says its cars are Level 2. It says you have to remain alert and ready to take over, which is the definition of Level 2, but it never mentions levels. I doubt that any car company will reference the SAE levels. They'll make vague claims, worded so as to give them plausible deniability in court, that seem to be promising more than they really are.

Websites like this are where we'll really find out what the cars can do.

I watched the Volvo presentation about their 2022 XC90 with Lidar. Their intent (not the promised feature at intro) is to provide completely supervised highway automation, where you can sleep or do whatever and any mishaps are Volvo's responsibility. And by the way I've seen no "ten second rule" in that expressed concept. The implication is that it'll be more like L4 while on the freeway, but will give you minutes of notice (I guess gentle followed by not-so-gentle alerts) to wake up and re-engage, in preparation for exiting into the non-autonomous ODD, and would have a safe fall-back if you fail to re-engage in time.

But when asked a number of times, they were very clear that they don't want to communicate their system's capabilities using the SAE definitions because those aren't clear enough nor precisely applicable. Instead they insisted that they would use the language of
"supervised" (assistance) vs "unsupervised" (autonomy) and the corollary
"driver is responsible" vs "Volvo is responsible."

In reading the recent SAE-level discussion here, I do think that it's become too focused on the SAE definitions and possible divergent interpretations. There's some arguing over behaviors that meet the letter of L3 or L4 but are not very safe, convenient or valuable. I think such issues will get sorted in the marketplace, and here we shouldn't be overly fixated on SAE levels as the most important descriptor of each new system - that issue is leading to tedious side arguments and in most cases, isn't even in the published description from the manufacturer. I do see that Daimler.com (MB) has officially described upcoming Drive Pilot as
"an SAE Level 3 conditional automated driving system"
I think this (referencing SAE levels explicitly) is a mistake on their part, and I predict that they will figure that out and remove such language for the reasons already mentioned.

Regarding the ten second rule and all its implications, I predict that if widely deployed it will turn out to be rarer and safer than the dangerous interpretations, and if not it will join the growing list of L3-ish highway systems that are Coming Soon but don't actually appear in any meaningful volume of production cars.

 

[diplomat33]

no car company is going to publish a nation-wide map showing which roads their car will and will not drive on

GM did. They published a map for SuperCruise.

Ford did. They published a map for BlueCruise.

 

[JHCCAZ]

GM did. They published a map for SuperCruise.

Ford did. They published a map for BlueCruise.

Interesting, thanks. These are more or less the US Interstates (roughly 50k miles) and similar-construction major freeways. Good for de-fatiguing occasional road trips or for those whose long commutes are dominated by stretches of those limited-access roads.

I don't mean to minimize the benefit because these routes do have a disproportionately high density of population and traffic. But it has limited benefit if you aren't in the group that frequently drives these major routes.

GM has the more developed map and, including the planned additions, covers 200k miles, roughly 5% of US road miles.

 

[daniel]

I watched the Volvo presentation about their 2022 XC90 with Lidar. Their intent (not the promised feature at intro) is to provide completely supervised highway automation, where you can sleep or do whatever and any mishaps are Volvo's responsibility. And by the way I've seen no "ten second rule" in that expressed concept. The implication is that it'll be more like L4 while on the freeway, but will give you minutes of notice (I guess gentle followed by not-so-gentle alerts) to wake up and re-engage, in preparation for exiting into the non-autonomous ODD, and would have a safe fall-back if you fail to re-engage in time.

But when asked a number of times, they were very clear that they don't want to communicate their system's capabilities using the SAE definitions because those aren't clear enough nor precisely applicable. Instead they insisted that they would use the language of
"supervised" (assistance) vs "unsupervised" (autonomy) and the corollary
"driver is responsible" vs "Volvo is responsible."

In reading the recent SAE-level discussion here, I do think that it's become too focused on the SAE definitions and possible divergent interpretations. There's some arguing over behaviors that meet the letter of L3 or L4 but are not very safe, convenient or valuable. I think such issues will get sorted in the marketplace, and here we shouldn't be overly fixated on SAE levels as the most important descriptor of each new system - that issue is leading to tedious side arguments and in most cases, isn't even in the published description from the manufacturer. I do see that Daimler.com (MB) has officially described upcoming Drive Pilot as
"an SAE Level 3 conditional automated driving system"
I think this (referencing SAE levels explicitly) is a mistake on their part, and I predict that they will figure that out and remove such language for the reasons already mentioned.

Regarding the ten second rule and all its implications, I predict that if widely deployed it will turn out to be rarer and safer than the dangerous interpretations, and if not it will join the growing list of L3-ish highway systems that are Coming Soon but don't actually appear in any meaningful volume of production cars.

If you can sleep, it's L4. I believe them when they say that's their "intent." I also believe them when they say the car will not have this feature when it's introduced.

We're not going to see Level 3 cars (even just highway) for a long time. They may (like Cadillac?) monitor your eyes rather than (like Tesla) your hands, but L3 means you do not have to pay attention, and that still far away for consumer cars.

GM did. They published a map for SuperCruise.

[...]

Ford did. They published a map for BlueCruise.

[...]

Sure, when your system only operates on divided, limited-access freeways you can publish a map. A map that has no roads in my state. A map that is limited to less than 1% of the roads in the country. And probably only at low speed during traffic jams?

I'm talking about a system that's supposed to be for wide-area general use. With millions of miles of roads in the U.S., either they're going to limit it so severely that the map of the interstate highway system covers it, or they're going to say something general and generic that sounds good but turns out not to apply to a ton of roads we normal folks drive on every day.

 

[JHCCAZ]

I watched the Volvo presentation about their 2022 XC90 with Lidar. Their intent (not the promised feature at intro) is to provide completely supervised highway automation, where you can sleep or do whatever and any mishaps are Volvo's responsibility....

Edit, I meant to write "...completely unsupervised highway automation..." Too late to fix it now.

 

[diplomat33]

Looks like Cruise poached some talent from Waymo:

https://twitter.com/x/status/1420472451444592642

 

[mikes_fsd]

Looks like Cruise poached some talent from Waymo:

https://twitter.com/x/status/1420472451444592642

Now they can go-to-market as fast as Waymo!
Yay!

 

[Doggydogworld]

If you can sleep, it's L4.

Yes. Or summon from across town, as Tesla promised in 2016.

We're not going to see Level 3 cars (even just highway) for a long time.

They may (like Cadillac?) monitor your eyes rather than (like Tesla) your hands, but L3 means you do not have to pay attention, and that still far away for consumer cars.

Honda did a 100 car faux launch in April. Mercedes says they will enable Drive Pilot later this year. It's Level 3 by definition - the driver does not need to pay attention.

I'm talking about a system that's supposed to be for wide-area general use. With millions of miles of roads in the U.S., either they're going to limit it so severely that the map of the interstate highway system covers it, or they're going to say something general and generic that sounds good but turns out not to apply to a ton of roads we normal folks drive on every day.

A system that only works during highway traffic jams might not be of much use to you (or me, frankly). But a lot of premium car commuters spend a lot of time in traffic jams on their local highways. We're talking 6, 8 and even 10+ hours per week. Some pay hundreds per month to access high speed lanes that bypass these jams. A car that lets them answer e-mails or whatever while stuck on highways is very valuable to these commuting professionals.

 

[mark95476]

Some people even have a vehicle and then pay to take public transportation (train, bus/coach) or a Uber/Lyft.

A system that only works during highway traffic jams might not be of much use to you (or me, frankly). But a lot of premium car commuters spend a lot of time in traffic jams on their local highways. We're talking 6, 8 and even 10+ hours per week. Some pay hundreds per month to access high speed lanes that bypass these jams. A car that lets them answer e-mails or whatever while stuck on highways is very valuable to these commuting professionals.

 

[diplomat33]

But I also blame Waymo for being utterly blind to the massively changing landscape all around them (the research and results in NN arena are not subtle or easily ignored). And for years they would go around to all the conferences speaking with arrogant authority that - our approach is the only approach!

Waymo was not blind to ML or NN or computer vision advances. Waymo pioneered much of the computer vision that Tesla is using now! Waymo uses 29 HD cameras. They have depth perception with vision-only, 4D vision and more. They have advanced computer vision. Their approach is not lidar-only as Elon and Karpathy falsely try to claim.

Waymo was not arrogant. Waymo demonstrated real FSD (driverless ride in 2015) when they claimed that their approach would work.

Waymo's FSD approach is "camera vision + lidar vision + radar vision + HD maps + AI + ML". Waymo is correct that their approach is right. The proof is the fact that so many AV companies have copied Waymo and have robotaxis or are close to deploying robotaxis on public roads. The vision-only approach has yet to deploy any robotaxis.

If anyone is arrogant, it's Elon claiming that his approach is correct and that FSD is already solved before he had any FSD at all or claiming that Tesla will have 1M robotaxis by 2020 when they had not even finished FSD. That's arrogance!

Waymo can still regain my support! Show us that you can make your solution available to everyone in US. Don't tell me about your sandbox in the desert. Sell me on how everyday people are going to benefit everywhere!

Waymo is collecting data to prove their safety. Waymo has stated that they will scale wide after they prove safety to regulators. IMO, once they do, they will scale with real fully driverless FSD everywhere. Mark my words!

 

[rxlawdude]

Waymo was not blind to ML or NN or computer vision advances. Waymo pioneered much of the computer vision that Tesla is using now! Waymo uses 29 HD cameras. They have depth perception with vision-only, 4D vision and more. They have advanced computer vision. Their approach is not lidar-only as Elon and Karpathy falsely try to claim.

Waymo was not arrogant. Waymo demonstrated real FSD (driverless ride in 2015) when they claimed that their approach would work.

Waymo's FSD approach is "camera vision + lidar vision + radar vision + HD maps + AI + ML". Waymo is correct that their approach is right. The proof is the fact that so many AV companies have copied Waymo and have robotaxis or are close to deploying robotaxis on public roads. The vision-only approach has yet to deploy any robotaxis.

If anyone is arrogant, it's Elon claiming that his approach is correct and that FSD is already solved before he had any FSD at all or claiming that Tesla will have 1M robotaxis by 2020 when they had not even finished FSD. That's arrogance!



Waymo is collecting data to prove their safety. Once they do, they will scale with real fully driverless FSD everywhere. Mark my words!

Lots of statements not qualified by "it is my opinion that..."

 

[diplomat33]

Lots of statements not qualified by "it is my opinion that..."

Because I am stating fact. I don't need to say "in my opinion" for facts.

 

[rxlawdude]

Because I am stating fact. I don't need to say "in my opinion" for facts.

The last paragraph of the first section is not fact. It's clearly your biased opinion.

It's arrogance to state your opinion as a fact.

 

[diplomat33]

The last paragraph of the first section is not fact. It's clearly your biased opinion.

It's arrogance to state your opinion as a fact.

Quote the specific passage please.

 

[rxlawdude]

Quote the specific passage please.

"If anyone is arrogant, it's Elon claiming that his approach is correct and that FSD is already solved before he had any FSD at all or claiming that Tesla will have 1M robotaxis by 2020 when they had not even finished FSD. That's arrogance!"

That's an OPINION, not a FACT.

 

[diplomat33]

"If anyone is arrogant, it's Elon claiming that his approach is correct and that FSD is already solved before he had any FSD at all or claiming that Tesla will have 1M robotaxis by 2020 when they had not even finished FSD. That's arrogance!"

That's an OPINION, not a FACT.

I think it is obvious that I am stating an opinion since I started the sentence with "if". But everything I said about Elon is fact. Elon really did claim that FSD was solved and did claim that Tesla would have 1M robotaxis by 2020. And it is also fact that Tesla did not have FSD when Elon made those claims. And I am making a correlation between Elon's statements that he did not have proof for and the definition of arrogance.