Autonomous Car Progress

[diplomat33]

I should probably stay out of this but let me throw in my 2 cents:

1) Tesla is not ahead on FSD. Tesla has a good driver assist system which is not the same thing as an autonomous driving system and is still working on finishing their vision neural nets that other companies have already done. Tesla's AP miles don't count towards autonomous driving miles because AP is not an autonomous driving system. Shadow mode does not count either because shadow mode is just a piece of the software running in the background, it is not the car driving in autonomous mode. Meanwhile, Waymo has L4 autonomous cars on public roads now in certain geofenced areas.

2) Tesla's fleet size is a valuable tool, yes, but it is not some magic bullet. It does not give Tesla any inherent special advantage over other companies. It is a source of data for machine learning and it is a source of feedback that can help Tesla determine where the software is failing. But other companies also have large datasets for machine learning and ways of getting feedback on how their software is performing.

I think the advantage of Tesla's fleet is probably more in deploying FSD, rather than developing FSD, thanks to Tesla's OTA update capabilities. When Tesla does finish a FSD feature or even finishes FSD itself, it can be rapidly deployed to a lot of cars. So this is an advantage in deploying FSD but not an advantage in developing FSD.

 

[diplomat33]

I would add that Tesla's large fleet may give Tesla an advantage in validating FSD but only AFTER they develop an autonomous driving system because at that point, yes, it will allow Tesla to collect a lot of autonomous driving miles very quickly. So after Tesla completes the software for autonomous driving and deploys it with driver supervision, then Tesla should be able to know fairly quickly if it is reliable or not. But the miles that the fleet drives only count towards autonomous driving after the software is done for autonomous driving, not before. Any miles driven while AP is still a driver assist or while software is in shadow mode do not count towards autonomous driving miles.

 

[mongo]

I would add that Tesla's large fleet may give Tesla an advantage in validating FSD but only AFTER they develop an autonomous driving system because at that point, yes, it will allow Tesla to collect a lot of autonomous driving miles very quickly. So after Tesla completes the software for autonomous driving and deploys it with driver supervision, then Tesla should be able to know fairly quickly if it is reliable or not. But the miles that the fleet drives only count towards autonomous driving after the software is done for autonomous driving, not before. Any miles driven while AP is still a driver assist or while software is in shadow mode do not count towards autonomous driving miles.

One could also make the claim that any miles driven on a previous version of autonomous software do not count toward the latest release. However, the miles driven do provide the edge cases for the training and testing of new versions. Once autonomous, if the HW is powerful enough, it could run the new version in parallel with the old version.

 

[diplomat33]

One could also make the claim that any miles driven on a previous version of autonomous software do not count toward the latest release.

Definitely. I think the miles reset to 0 each time a new software version rolls out. But again, the large fleet size means that Tesla can collect a lot of miles on the new software version in order to validate it pretty quickly.

However, the miles driven do provide the edge cases for the training and testing of new versions. Once autonomous, if the HW is powerful enough, it could run the new version in parallel with the old version.

Yes. I wonder if the HW3 computer could do this.

 

[Bladerskb]

Two videos demonstrating how advanced Mobileye cars are using only cameras.

Inside a Mobileye-Powered Autonomous Car (B-Roll)

 

[mongo]

Definitely. I think the miles reset to 0 each time a new software version rolls out. But again, the large fleet size means that Tesla can collect a lot of miles on the new software version in order to validate it pretty quickly.

Yes. I wonder if the HW3 computer could do this.

Hopefully, Without parallel software operation, do you force drivers back to nag based operation on each release?

 

[diplomat33]

Two videos demonstrating how advanced Mobileye cars are using only cameras.

Inside a Mobileye-Powered Autonomous Car (B-Roll)

I read that Lucid is planning to use Mobileye for their autonomous driving on their upcoming Lucid Air EV. Do you know anything more about that? The Lucid website says this:

"In the future, Lucid's assistive technology will help you and your family get things done. Over-the-air software upgrades will allow the Lucid Air to transition through progressive levels of autonomy. In the future, your car will be able to retrieve your groceries, pick up your kids from practice, or provide you a moment to sit back and relax as you are safely driven home. This time is yours." Lucid

Do you think Lucid Air will be L4 autonomous when it is released?

 

[diplomat33]

Hopefully, Without parallel software operation, do you force drivers back to nag based operation on each release?

I imagine that Tesla will just keep the nags until they are ready to remove them once and for all, rather than remove them, put them back, remove them, put them back with each new software version.

 

[mongo]

I imagine that Tesla will just keep the nags until they are ready to remove them once and for all, rather than remove them, put them back, remove them, put them back with each new software version.

If new versions of SW invalidate previous validation miles, how does that work? That's where I think there has to be a parallel operation step and might be part of Tesla's dual chip dual core AP setup.

 

[DanCar]

I imagine that Tesla will just keep the nags until they are ready to remove them once and for all, rather than remove them, put them back, remove them, put them back with each new software version.

It is a big step to remove them. That means liability moves from driver to Tesla and Tesla will get sued out of existence. More likely the timing between nags will increase.
If nags do get removed, will most likely happen in the most pristine conditions. Bumper to bumper traffic on divided limited access freeway.

 

[SandiaGrunt]

areas Waymo does not operate in (parking lots, multiple states)

Where on earth did you hear these things?

From PopSci: “In fact, parking lots are a distinctive enough environment that Waymo, the self-driving car company that’s a sibling to Google, specifically trains its vehicles to deal with them by setting up real-world scenarios in a controlled environment.”

From Wikipedia: “As of 2018, Waymo had tested its system in six states and 25 cities across the U.S over a span of more than 9 years.”

 

[CarlK]

Where on earth did you hear these things?

From PopSci: “In fact, parking lots are a distinctive enough environment that Waymo, the self-driving car company that’s a sibling to Google, specifically trains its vehicles to deal with them by setting up real-world scenarios in a controlled environment.”

From Wikipedia: “As of 2018, Waymo had tested its system in six states and 25 cities across the U.S over a span of more than 9 years.”

One word "Geofensed". Waymo and Waymo types using Lidar and HD mapping do not operate in any place that is not fully (and newly) mapped. That also means they can "test" in parking lots but they will not be able to "operate" in parking lots like Tesla can do now.

 

[CarlK]

I should probably stay out of this but let me throw in my 2 cents:

1) Tesla is not ahead on FSD. Tesla has a good driver assist system which is not the same thing as an autonomous driving system and is still working on finishing their vision neural nets that other companies have already done. Tesla's AP miles don't count towards autonomous driving miles because AP is not an autonomous driving system. Shadow mode does not count either because shadow mode is just a piece of the software running in the background, it is not the car driving in autonomous mode. Meanwhile, Waymo has L4 autonomous cars on public roads now in certain geofenced areas.

2) Tesla's fleet size is a valuable tool, yes, but it is not some magic bullet. It does not give Tesla any inherent special advantage over other companies. It is a source of data for machine learning and it is a source of feedback that can help Tesla determine where the software is failing. But other companies also have large datasets for machine learning and ways of getting feedback on how their software is performing.

I think the advantage of Tesla's fleet is probably more in deploying FSD, rather than developing FSD, thanks to Tesla's OTA update capabilities. When Tesla does finish a FSD feature or even finishes FSD itself, it can be rapidly deployed to a lot of cars. So this is an advantage in deploying FSD but not an advantage in developing FSD.

Yes you should stay out of this since I see you at least as a reasonable person and don't always argue for argument's sake. I could be wrong though.

No one, let me repeat NO ONE else except Tesla has vision deep learning neural net in operation and likely to have one in the near future (beyond laboratory stage).

No one seems even want to bother to watch Fridman and Karpathy's videos to learn how this thing works yet many people don't hesitate to make up their own layperson fake theories trying to confuse the issue.

 

[mongo]

areas Waymo does not operate in (parking lots, multiple states)

Where on earth did you hear these things?

From PopSci: “In fact, parking lots are a distinctive enough environment that Waymo, the self-driving car company that’s a sibling to Google, specifically trains its vehicles to deal with them by setting up real-world scenarios in a controlled environment.”

From Wikipedia: “As of 2018, Waymo had tested its system in six states and 25 cities across the U.S over a span of more than 9 years.”

Um, you just confirmed it

The US has 50 states. 50 -6 = 44 states that Waymo does not operate in, I'd call that multiple (or a majority).
Want to compare how many cities they do vs do not operate in?

Teslas are summoning in as real world a parking environment as it gets. Not specific or test parking lots, any parking lot.
Do Waymo routes currently use parking lots? I only see articles from August saying it's hard.

 

[DanCar]

... Tesla is not ahead on FSD. ...

I understand my opinion is in the minority, but I do believe Tesla is ahead on FSD for only one reason, which is risk. Waymo is far ahead in terms of tech, but they are unable to risk blood on their hands. Waymo will fail because of this. Tesla is not adverse to blood.

 

[SandiaGrunt]

Not sure what you're talking about but watch some of those Fridman's audio to learn please.

So wait... you don’t understand what I’m saying, so *I* need to go learn something?

Deep learning does not learn the same thing over and over again. Soon as a case is solved then it's done unless situation changed.

Deep nets are retrained all the time with new and better data sets. Each iteration is brand new. There is no carry-over of learning from one training run to the next. Every time you retrain a net, you risk losing some capability you had in the past.

Sad that that are still people like you two trying to reinvent established science in your dreams. No wonder we got all those flat earthers and climate change deniers.

You are sort of a caricature of Tesla fans — the worst example out of many bad examples. I’d almost declare Poe’s law, but I think you’re serious about this.

You can do everything other test cars do, and even more, without sending final output to steering wheel and pedals.

If the car’s software doesn’t control the car, then every time it disagrees with the driver, it is put into a state that it didn’t choose, and may be outside its training set or programming. This is called “distributional drift,” and it makes the proposed “shadow mode” nearly useless in practice. It’s an idea that sounds good to lay people, but you’d learn how it fails in literally the first lecture of Berkeley’s class on deep reinforcement learning.

That's the perfect scenario for using deep learning.

Everyone uses deep learning. Literally everyone.

No way your Lidar + mapping could work in there.

And yet it does work — really really well.

Waymo and Waymo types using Lidar and HD mapping do not operate in any place that is not fully (and newly) mapped.

A map is just sensor data, compiled and passed through a human review. It gives humans a chance to insert some nuance or finesse. Over time, the humans find fewer things to fix, and the map becomes just a Bayesian prior that improves safety. It will cost nothing to produce. Its marginal cost will fall to zero. The map will be essentially free; just drive the cars around you’re done.

 

[mongo]

Every time you retrain a net, you risk losing some capability you had in the past.

Only if you didn't test/ train for that capability. Otherwise, the training. testing would not be successful.

(are you are disagreeing that 44 states is considered multiple?)

 

[SandiaGrunt]

No one, let me repeat NO ONE else except Tesla has vision deep learning neural net in operation

Literally every AV company uses deep learning for vision.

No one seems even want to bother to watch Fridman and Karpathy's videos to learn how this thing works yet many people don't hesitate to make up their own layperson fake theories trying to confuse the issue.

I think you may be misunderstanding the situation here. There are several people here who know a great deal more about this subject than what you could learn from watching the Autonomy Day presentation. That presentation was specifically designed to impress laypeople (investors). It was actually laughably entry-level, approximately like the first lecture in a class on deep learning. Karpathy used to teach a class on deep learning at Stanford, and there were some eerie similarities.

 

[diplomat33]

If new versions of SW invalidate previous validation miles, how does that work? That's where I think there has to be a parallel operation step and might be part of Tesla's dual chip dual core AP setup.

Just guessing but I think it will probably work similar to how Tesla staggers software updates across the fleet now. Tesla will release a version of FSD to a portion of the fleet, see what the disengagement rate it, make changes, release a new version, rinse and repeat eventually, they accumulate say 1 billion miles of autonomous driving on the same version and the disengagement rate hits the right number and then Tesla can say that their FSD is good enough to remove driver supervision.

I understand my opinion is in the minority, but I do believe Tesla is ahead on FSD for only one reason, which is risk. Waymo is far ahead in terms of tech, but they are unable to risk blood on their hands. Waymo will fail because of this. Tesla is not adverse to blood.

That's a weird thing to say. So Tesla is more willing to let people die? That would not be a good thing in my book.

 

[CarlK]

Re: other cars blinker used

18:55 and 19:57 (maybe the blinkers on)

Glad there is someone actually is watching those and trying to learn some real science.

18:55 and 19:57 in that video is a good sample of deep learning and how it can be used to learn "body language" of cars and pedestrians on the road (at parking lot, stop signs, unmarked lanes, pedestrian crossing etc.). At this moment this is considered the final missing link of self driving implementation. Again no one else with Waymo type Lidar + mapping will be able to solve this. No one else can set up the Tesla deep learning in the near future even when they decide to go this route (which they will have to eventually).

OK I've had enough fun here. I'll leave it to you some blinds who can lead other blinds and continue to have your fun.