Autonomous Car Progress

[AKAlex]

I think it's undeniable that Tesla will dominate if their system works well enough on HW3 for robotaxi operation.
I don't think it will but I'm told I should never underestimate Elon!
A big question I have about Tesla is how they're going to manage the transition to new hardware if it proves necessary.

Yea, as far as I can tell right now, the bottlenecks are first software then on-board compute. I think the concern over cameras specs/placement is most likely misplaced. Software is the most difficult to deal with, as it hasn't really been done yet. This a novel challenge. As far as update of the compute, we saw this happen in the case of the HW2.5 to 3.0 upgrades. Doing this upgrade is just a matter of economics (Which are favorable).

 

[bradtem]

that might be true in the general context of the thread but per your post you were making a very subjective statement about the word "works" and how it applies to FSD Beta. and i stand by my argument that FSD beta works, it might not work well or in every case, but the fact that it doesn't just drive every person straight into a wall proves it technically works lol, its just flawed as of yet. and furthermore, we cannot make direct comparisons with other companies and how well they "work" since the only thing we have to go by is their own materials in most cases. so it could very well be the case that they all work at the same level or that one is vastly superior but since FSD beta is the only one any of US can actually use, i'd say its a pretty safe bet tesla's is better when all context is accounted for, otherwise you'd see a mad rush of auto makers picking up that specific thing instead of the extremely limited trickle of pseudo-tesla features you see now. remember you will never likely see geofenced robotaxi hardware/software on a production car you can buy, so if my goal is a self driving car that tech does not "work" for me.

Well, in the self-driving world, as opposed to the ADAS world, "works" does not include "might not work well in every case." Perhaps if you came to this from a consumer or ADAS worldview, you might not immediately think like a self-driving developer. In that field, "works" means "makes so few mistakes as to reach an acceptable safety threshold for unmanned operation in the ODD." And that's a very, very, very, very, very few mistakes.

It might be better if we could say "works perfectly" but perfectly is not the goal, but it's surprisingly close to the goal and so would be much closer to a description of it.

Tesla is taking the longshot approach, hoping for breakthroughs in computer vision and machine learning. If they can pull this off, it will be good for them, though not as good as you might imagine. Tesla has been competing with old-world car companies and eating their lunch, because their lunch is not hard to eat. In self-driving, it faces Google, Apple, and a few dozen other high tech companies around the world who are every bit as savvy about high tech and AI as Tesla is, more so in many cases. (I mean do you seriously think Tesla knows more about machine learning than Alphabet DeepMind? More about custom AI processors than Intel, Nvidia or Google? ) Tesla can be a player against those companies, but not the way it was better at EVs than the dinosaurs of the ICE world.

 

[diplomat33]

FSD is capable of operating on ~98% of roads in the US with limited reliability. Waymo can operate on ~0.1% of roads in the US with high reliability.

To be clear, Waymo can operate on 100% of roads. It is only available to the public in small areas so far. A lot of Tesla fans seem to confuse the two. Waymo works in more areas than where it is deployed to the public.

 

[diplomat33]

I have a similar suspicion of some of the data that is coming out of those companies. Without the context around disengagements, it's very difficult to draw any conclusions.

Manufacturers provide a description of each disengagement in the CA DMV report. Some companies provide more context about their disengagements than others. But overall, I think we get a good idea of why the disengagements happened.

Here are some examples:

AIMotive:

During a merge, the test vehicle failed to keep an appropriate distance between a merging car or the merging car failed to yield to us. Conditions: Non-inclement weather, dry roads, no other factors involved.

Apple:

Incorrect prediction led to undesirable motion plan.

Cruise:

Other road user behaving poorly; lane encroachment

Precautionary takeover to address controls; yielding to cross traffic

Pony.AI:

Planning; as a precaution, the safety operator intervened as a vehicle did not follow the appropriate traffic procedures and proceeded straight from the ADV's left while the ADV was making a right turn.

Waymo:

Disengage for unwanted maneuver of the vehicle that was undesirable under the circumstances.

 

[AKAlex]

To be clear, Waymo can operate on 100% of roads. It is only available to the public in small areas so far. A lot of Tesla fans seem to confuse the two. Waymo works in more areas than where it is deployed to the public.

I'm sure you're right about Waymo working in more areas than just where it is available to the public. I would assume that there are areas where it is being tested that have not been perfected yet. However, I had not heard that it was working on all roads now! Could you point me in the right direction to learn more about this?

 

[OncomingStorm]

Well, in the self-driving world, as opposed to the ADAS world, "works" does not include "might not work well in every case." Perhaps if you came to this from a consumer or ADAS worldview, you might not immediately think like a self-driving developer. In that field, "works" means "makes so few mistakes as to reach an acceptable safety threshold for unmanned operation in the ODD." And that's a very, very, very, very, very few mistakes.

It might be better if we could say "works perfectly" but perfectly is not the goal, but it's surprisingly close to the goal and so would be much closer to a description of it.

Tesla is taking the longshot approach, hoping for breakthroughs in computer vision and machine learning. If they can pull this off, it will be good for them, though not as good as you might imagine. Tesla has been competing with old-world car companies and eating their lunch, because their lunch is not hard to eat. In self-driving, it faces Google, Apple, and a few dozen other high tech companies around the world who are every bit as savvy about high tech and AI as Tesla is, more so in many cases. (I mean do you seriously think Tesla knows more about machine learning than Alphabet DeepMind? More about custom AI processors than Intel, Nvidia or Google? ) Tesla can be a player against those companies, but not the way it was better at EVs than the dinosaurs of the ICE world.

i'd prefer not to get bogged down with nonsensical semantic arguments about the use of the word works, the word can really mean whatever you want it to, or rather its defined by the context of the sentence in which its used. i think we technically both agree on the reality, seems silly to go down that rabbit hole haha.

to the later half of your comment... yeah they defiantly are taking the hail mary approach lol, that said all the pieces of a self driving car theoretically exist already. our software and computing technology is sophisticated enough (or can be) and our hardware is good enough to be better than human senses (by "our" i mean humans not tesla lol). so really someone doing it is inevitable, thus the reason you see such saturation in the development by so many companies.

and to your point about the big players, yes they all have tech that could be applied to a self driving car, and in some cases that particular piece might be better (AI, processor manufacturing), but that in no way guarantees success, many of those companies have hundreds of prodcuts they've abandoned, apple was supposed to have a self driving car by now i think (yes just like tesla lol). then considering to the biggest players, a self driving car is a "project", not their sole marketable product, i'd argue that tesla's far more aggressive approach will ultimately bridge the gap faster (to the goal of a level 5 car), of course thats just my opinion

one thing i think as time goes on we will have to redefine what a self driving car actualy is. if my car can dive itself but cannot leave a 40 mile radius or is only allowed on specific routes / sections of road, is that really self driving or a gimik? likewise, if my car can drive me anywhere in the US but i have to be in the drivers seat and "ready to take over" is that not self driving? i know the "levels" would answer those questions, but i think the levels themselves are in need of further definition. there's theory and there is practical reality. just imho

 

[AKAlex]

Manufacturers provide a description of each disengagement in the CA DMV report. Some companies provide more context about their disengagements than others. But overall, I think we get a good idea of why the disengagements happened.

Here are some examples:

AIMotive:

During a merge, the test vehicle failed to keep an appropriate distance between a merging car or the merging car failed to yield to us. Conditions: Non-inclement weather, dry roads, no other factors involved.

Apple:

Incorrect prediction led to undesirable motion plan.

Cruise:

Other road user behaving poorly; lane encroachment

Precautionary takeover to address controls; yielding to cross traffic

Pony.AI:

Planning; as a precaution, the safety operator intervened as a vehicle did not follow the appropriate traffic procedures and proceeded straight from the ADV's left while the ADV was making a right turn.

Waymo:

Disengage for unwanted maneuver of the vehicle that was undesirable under the circumstances.

I think that most of these came from the first spreadsheet on the California DMV disengagement report website right? (Disengagement Reports - California DMV)

I guess I would be curious to know more context around these disengagements, a lot of the responses are total boilerplate language.

 

[diplomat33]

I think that most of these came from the first spreadsheet on the California DMV disengagement report website right? (Disengagement Reports - California DMV)

I guess I would be curious to know more context around these disengagements, a lot of the responses are total boilerplate language.

Yes, they come from the first spreadsheet. That's all we have. I agree it would be nice to have more context. Like I said, some manufacturers provide a lot of details, others do not.
"

 

[Daniel in SD]

i'd prefer not to get bogged down with nonsensical semantic arguments about the use of the word works, the word can really mean whatever you want it to, or rather its defined by the context of the sentence in which its used. i think we technically both agree on the reality, seems silly to go down that rabbit hole haha.

to the later half of your comment... yeah they defiantly are taking the hail mary approach lol, that said all the pieces of a self driving car theoretically exist already. our software and computing technology is sophisticated enough (or can be) and our hardware is good enough to be better than human senses (by "our" i mean humans not tesla lol). so really someone doing it is inevitable, thus the reason you see such saturation in the development by so many companies.

and to your point about the big players, yes they all have tech that could be applied to a self driving car, and in some cases that particular piece might be better (AI, processor manufacturing), but that in no way guarantees success, many of those companies have hundreds of prodcuts they've abandoned, apple was supposed to have a self driving car by now i think (yes just like tesla lol). then considering to the biggest players, a self driving car is a "project", not their sole marketable product, i'd argue that tesla's far more aggressive approach will ultimately bridge the gap faster (to the goal of a level 5 car), of course thats just my opinion

one thing i think as time goes on we will have to redefine what a self driving car actualy is. if my car can dive itself but cannot leave a 40 mile radius or is only allowed on specific routes / sections of road, is that really self driving or a gimik? likewise, if my car can drive me anywhere in the US but i have to be in the drivers seat and "ready to take over" is that not self driving? i know the "levels" would answer those questions, but i think the levels themselves are in need of further definition. there's theory and there is practical reality. just imho

Well, the term "self-driving" no longer means anything which is why I always talk about robotaxis so there's no confusion.
The "levels" which I try not speak about around here also say "self-driving" is a deprecated term.

 

[bradtem]

i'd prefer not to get bogged down with nonsensical semantic arguments about the use of the word works, the word can really mean whatever you want it to, or rather its defined by the context of the sentence in which its used. i think we technically both agree on the reality, seems silly to go down that rabbit hole haha.

to the later half of your comment... yeah they defiantly are taking the hail mary approach lol, that said all the pieces of a self driving car theoretically exist already. our software and computing technology is sophisticated enough (or can be) and our hardware is good enough to be better than human senses (by "our" i mean humans not tesla lol). so really someone doing it is inevitable, thus the reason you see such saturation in the development by so many companies.

and to your point about the big players, yes they all have tech that could be applied to a self driving car, and in some cases that particular piece might be better (AI, processor manufacturing), but that in no way guarantees success, many of those companies have hundreds of prodcuts they've abandoned, apple was supposed to have a self driving car by now i think (yes just like tesla lol). then considering to the biggest players, a self driving car is a "project", not their sole marketable product, i'd argue that tesla's far more aggressive approach will ultimately bridge the gap faster (to the goal of a level 5 car), of course thats just my opinion

one thing i think as time goes on we will have to redefine what a self driving car actualy is. if my car can dive itself but cannot leave a 40 mile radius or is only allowed on specific routes / sections of road, is that really self driving or a gimik? likewise, if my car can drive me anywhere in the US but i have to be in the drivers seat and "ready to take over" is that not self driving? i know the "levels" would answer those questions, but i think the levels themselves are in need of further definition. there's theory and there is practical reality. just imho

Nobody is guaranteed success. Whether we have all the components is not entirely clear, actually. We often hear arguments (especially in Tesla circles) like "Humans can drive with just their eyes" or "Human can drive without maps" as though our AI tools are anywhere near being as capable as a human in that way. Like most computer tools they are more capable than a human in some ways, much less in others. Several companies (Wayve just got $200M this week) are trying for a fully end-to-end approach, even more than Tesla is planning. They know it might take far more compute than we have available today, if it works at all, but they hope it doesn't. As does Tesla. Waymo has shown that they do indeed know how to make it work in a suburb with superhuman sensing and maps, and hope they are not far from generalizing it to many more places, but have yet to.

 

[OncomingStorm]

Nobody is guaranteed success. Whether we have all the components is not entirely clear, actually. We often hear arguments (especially in Tesla circles) like "Humans can drive with just their eyes" or "Human can drive without maps" as though our AI tools are anywhere near being as capable as a human in that way. Like most computer tools they are more capable than a human in some ways, much less in others. Several companies (Wayve just got $200M this week) are trying for a fully end-to-end approach, even more than Tesla is planning. They know it might take far more compute than we have available today, if it works at all, but they hope it doesn't. As does Tesla. Waymo has shown that they do indeed know how to make it work in a suburb with superhuman sensing and maps, and hope they are not far from generalizing it to many more places, but have yet to.

agreed 100% its anyones game at this point, mostly because of how little we actually know behind the scenes but the result is the same lol.

i know a lot of people will disagree with this and has probably been hashed out in the 356 pages of this thread lol but hear me out! haha... i think tesla made the right move going vision only. the power required to integrate multiple systems is daunting, espically with a system like radar, you basically need an entire separate system to interpret the data from that and the shunt it over to the main system for further analysis of "potential" objects which it then compares to its camera inputs (just an example i can't say exactly how they handle data propagation). radar kicks back a lot of data, you get signals off everything from the car in front of you to the slight bump in the road or curb, that takes a lot of processing power away from other areas.

so ok, lets assume driving CAN be accomplished with just camera's/vision. a sophisticated enough program could theorize about objects that are as small as 1 pixel in the view (based on context, surroundings, what is over in that direction, etc.. ). in most instances this is quite a bit further than (or at least equal too) what we can see and discern. if it can build a large enough virtual world, its predictions should be much more accurate and projected further out in time, which opens up a lot of possibilities. and they will achieve this faster by maximizing computing power and going vision only now instead of wasting years with integrations that will eventually be dropped anyway.

lots of assumption and opinion in that, but i think it holds. i dont see it as impossible they would need to upgrade the cameras and computer but those are simple retrofits. one area i do think they will eventually shift gears is the HD maps or some variation of that. since that data is mostly static and already digital it could be integrated into the system extremely easily with significant benefit imo, but who knows right lol.

 

[bradtem]

agreed 100% its anyones game at this point, mostly because of how little we actually know behind the scenes but the result is the same lol.

i know a lot of people will disagree with this and has probably been hashed out in the 356 pages of this thread lol but hear me out! haha... i think tesla made the right move going vision only. the power required to integrate multiple systems is daunting, espically with a system like radar, you basically need an entire separate system to interpret the data from that and the shunt it over to the main system for further analysis of "potential" objects which it then compares to its camera inputs (just an example i can't say exactly how they handle data propagation). radar kicks back a lot of data, you get signals off everything from the car in front of you to the slight bump in the road or curb, that takes a lot of processing power away from other areas.

so ok, lets assume driving CAN be accomplished with just camera's/vision. a sophisticated enough program could theorize about objects that are as small as 1 pixel in the view (based on context, surroundings, what is over in that direction, etc.. ). in most instances this is quite a bit further than (or at least equal too) what we can see and discern. if it can build a large enough virtual world, its predictions should be much more accurate and projected further out in time, which opens up a lot of possibilities. and they will achieve this faster by maximizing computing power and going vision only now instead of wasting years with integrations that will eventually be dropped anyway.

lots of assumption and opinion in that, but i think it holds. i dont see it as impossible they would need to upgrade the cameras and computer but those are simple retrofits. one area i do think they will eventually shift gears is the HD maps or some variation of that. since that data is mostly static and already digital it could be integrated into the system extremely easily with significant benefit imo, but who knows right lol.

Many believe that is will be possible to do it with vision at some date in the future. They don't know the date though. Most teams have decided they want to make the tech work now, without waiting for a set of unpredictable breakthroughs. Instead, they want to wait for highly predictable improvements, such as LIDAR, compute and other sensors getting cheaper and lower power, which is a trend that has happened reliably for many decades.

 

[Doggydogworld]

we will have to redefine what a self driving car actualy is.

Tesla already did, rendering the term useless. Self-driving now means you drive it yourself, lol. "Full Self Driving" means Level 2 ADAS. Since you can't go beyond "Full", a car that drives with nobody in the driver's seat needs a different term. The industry uses Autonomous. And this thread is about Autonomous Cars.

if my car can dive itself but cannot leave a 40 mile radius

Hugely useful, since 95%* of all driving days stay within 40 miles.

or is only allowed on specific routes / sections of road,

True autonomy restricted to congested US metro highways would free up hundreds of man-hours per year per commuter. Billions of man-hours in total.

is that really self driving or a gimik?

"Full Self Driving" is the gimmick. It frees up no man hours. It's a parlor trick.

likewise, if my car can drive me anywhere in the US but i have to be in the drivers seat and "ready to take over"

"Be ready to take over" is Mercedes Drive Pilot, if it ever ships. Answer e-mails, read a book, etc. as long as you can respond to a takeover request within 10-15 seconds.

FSD is not "be ready to take over". FSD is "watch like a hawk or suffer decapitation". FSD frees up zero man hours. I taught four teenagers how to drive, I'm done watching like a hawk.

the power required to integrate multiple systems is daunting,

All else equal, multiple sensor types reduce computational load.

in most instances this is quite a bit further than (or at least equal too) what we can see and discern.

Way, way off. Human vision near the focal point is equivalent to a 576 megapixel camera. Tesla cameras are 1 megapixel.

*Completely made up statistic, but true

 

[voyager]

Mercedes-Benz Plans To Have Level-3 Automated Driving Vehicles On U.S. Roads This Year | Carscoops

Many automakers are competing to be the first to Level 3 autonomy in the United States, and it looks like Mercedes might be the first to do so

www.carscoops.com

Level 3 marks a clear transition from simple driver aids to true autonomous driving systems, and as such, Mercedes says drivers are able to take their hands completely of the wheel and engage freely in “certain secondary activities”, which can include things like watching movies, sending emails, and communicating with colleagues.

 

[EVNow]

https://twitter.com/x/status/1497826835144777733

 

[EVNow]

https://twitter.com/x/status/1498369312192548866

 

[diplomat33]

https://twitter.com/x/status/1498369312192548866

I don't think there was any real doubt Kyle would be made CEO. He founded Cruise and was interim CEO. He was the natural choice.

Also kind of weird that Omar is complimentary towards Kyle considering Omar has such a disdain for the "lidar approach".

 

[EVNow]

Also kind of weird that Omar is complimentary towards Kyle considering Omar has such a disdain for the "lidar approach".

Something personal ? (Esp. since Vogt replied)

I’ve lots of friends in companies that compete with Tesla, MSFT too.

ps : Waymo vs Tesla is in a way weird given close friendship between Google founders and Elon.

 

[diplomat33]

Waymo got their permit to charge for rides in SF with a safety driver:

https://twitter.com/x/status/1498480455254904837

For those who don't know, CA requires AV companies to get permits from the DMV to test AVs and separate permits from the Public Utilities Commission to charge money for AV rides.

 

[EVNow]

Waymo got their permit to charge for rides in SF with a safety driver:

https://twitter.com/x/status/1498480455254904837

For those who don't know, CA requires AV companies to get permits from the DMV to test AVs and separate permits from the Public Utilities Commission to charge money for AV rides.

How much do they charge - say in comparison to Uber rates ?