Autonomous Car Progress

[JHCCAZ]

I have no dog in the game, that was how I saw it described by others a long while ago and it made perfect sense as a separation against typical L2 (which can not finish the whole trip). I have no problem using "door to door" if that avoids confusion with the machine learning terminology.
I will note however that not very far upthread, Mobileye uses the same "end to end" terminology.
Autonomous Car Progress

I searched a little for this, and yes I did find the term "end-to-end" attached to recent MobileEye articles (e.g. CES Intel press release and a trade-journal article covering the same) - but only in the summary headlines that may have been written by webmasters or others- not in the blog article text.

I agree it's not the biggest issue but I think it's good to take note of it. There's already a lot of loose claims and subsequent parsing of statements around here to reinforce points (or to score points ). It confused me a little, so I thought it might be confusing to others who are trying to follow the developments.

 

[stopcrazypp]

Well I think significantly safer than a human is good enough for robotaxi, afterall it would save lives. Elon is a very rational person so I actually think he agrees and the only reason he hedges is concerns about liability. Anyway he should be more clear because with all the talk of robotaxis and asset values it seems like he’s trying to mislead on the timeline.

Well I don't think the regulators that monitor such things agree that's good enough, especially to launch 265k vehicles. "Significantly" can be 10% better. Would people really be fine with that, especially if it comes to fatality or injury rates? And Elon himself said the bar was 2-3x minimum.
I'm reminded of all the argument over AP safety statistics. I think there will be similar arguments if Tesla releases their FSD Beta safety stats.

 

[Terminator857]

Camera washing nozzles:

https://www.reddit.com/r/electricvehicles/comments/sfyeo9

 

[stopcrazypp]

I searched a little for this, and yes I did find the term "end-to-end" attached to recent MobileEye articles (e.g. CES Intel press release and a trade-journal article covering the same) - but only in the summary headlines that may have been written by webmasters or others- not in the blog article text.

I agree it's not the biggest issue but I think it's good to take note of it. There's already a lot of loose claims and subsequent parsing of statements around here to reinforce points (or to score points ). It confused me a little, so I thought it might be confusing to others who are trying to follow the developments.

I dug a bit myself to get something more definitive on what Mobileye meant by it, and found this article where the CEO used the term repeatedly, and these parts makes it pretty clear he means door to door when he says "end-to-end," not the machine learning definition. So this terminology is coming from the top, not just others adding it in when making the titles.

"So it has two such chips, and it provides an end-to-end capability. We have unedited videos on the internet showing how this car drives at multiple sites in Tel Aviv, Jerusalem, Munich, Tokyo, Paris, China, Detroit. I'm going to show this at the CES. So it's really end-to-end for L2+, but it doesn't have the robustness good enough to remove the driver out of the loop. For that, you need more compute and more sensor modalities in order to create redundancy"

"full L2+ which provides an end-to-end capability of hands-free driving. It's not L3 or L4 yet, and you need the driver behind the steering wheel, but in terms of the capability, it's a full end-to-end hands-free driving."

An Exclusive Interview with Mobileye CEO Prof. Amnon Shashua: The Future of Autonomous Driving

www.anandtech.com

I did try to dig a bit to see where I saw the first reference, but I don't recall. This was the first mention I found in these forums that seems to refer to the same definition:
"Based on the previous technical R&D and the iteration of road tests for the first half of 2019, the new sensing system that is composed of 10 cameras will help Baidu's vehicles to conduct the end-to-end closed-loop autonomous driving on urban roads sans high-beam rotatable lidar, said Wang Liang."
Apollo Lite: Baidu's autonomous vehicle hardware suite without lidar

 

[Daniel in SD]

Well I don't think the regulators that monitor such things agree that's good enough, especially to launch 265k vehicles. "Significantly" can be 10% better. Would people really be fine with that, especially if it comes to fatality or injury rates? And Elon himself said the bar was 2-3x minimum.
I'm reminded of all the argument over AP safety statistics. I think there will be similar arguments if Tesla releases their FSD Beta safety stats.

No, I don't think 10% would be significantly. I think significantly means a large enough improvement that it is statistically significant, so maybe 10% could be the lower end of the confidence interval.
Yes, I could definitely see arguments over the methodology used to calculate safety. Though it's actually easier to calculate the safety of FSD if you have some vehicles using FSD 100% of the time.
My opinion is that significantly would be enough because I suspect that the system will get in way fewer at fault collisions than a human driver but have relatively more collisions where the other driver is at fault. So the public perception would be that the system is not causing collisions.

 

[HVM]

 

[diplomat33]

Brad Templeton has an excellent article on Mobileye's approach that details everything we know so far.

His conclusion:

MobilEye is one of the few companies to have it all: Experience, a huge fleet to draw mapping and training data from, extensive mapping efforts at very low cost, an FMCW LIDAR in the works, imaging radar, advanced computer vision, a trip-planning app, the ability to make its own silicon, low cost, robtaxis driving in complex cities and the most relationships of automakers of anybody in the game. While some players have better in some of these individual area, nobody has as good a combination. The key that remains to be seen is just how good their software is. Shashua said they are still working at getting their system to 1,000 hours between accidents but they are confident they will get there soon. That’s not there yet, as humans go 3,500 hours between minor dings and about 12,000 hours between police reported accidents. We’ll be watching to see how they do.

Intel’s MobilEye Levels Up To Take On Tesla And Others In Self-Driving

MobilEye has released powerful new chips and has a well-crafted strategy to create a robotaxi. I interview CEO Amnon Shashua, analyse their strategy and contrast it with Tesla and others.

www.forbes.com

 

[Matias]

FSD is an engineering problem, which Tesla is good at.

Actually it is a problem which still requires fundamental research and Tesla has not done fundamental research at all. Previous problems have been applied research.

 

[Matias]

“Shashua said they are still working at getting their system to 1,000 hours between accidents but they are confident they will get there soon. That’s not there yet, as humans go 3,500 hours between minor dings”

I thought they already have this.

 

[Terminator857]

Actually it is a problem which still requires fundamental research and Tesla has not done fundamental research at all. Previous problems have been applied research.

No one is better on applied research than Tesla. They apply it directly to us.

 

[powertoold]

Tesla using HD maps confirmed!

 

[hatch]

Actually it is a problem which still requires fundamental research and Tesla has not done fundamental research at all. Previous problems have been applied research.

Yep.. and i also think Tesla THINKS they have done more research on the matterthan anyone else. Because they have so much more data from the car than anyone else could dream of having anytime soon. But if they don't even understand the fundamental problem, and the relationship between human beings and computers, they can't know what the data is really showing. They can't be starting from the right place or making the correct assumptions to begin the process. The result is a wild goose chase, compared to the realities of real world FSD.

 

[diplomat33]

Cruise has now opened up their driverless rides in SF to the general public:

If you live in the SF service area, you can sign up for a driverless ride here: We’re Cruise, a self-driving car service designed for the cities we love.

 

[diplomat33]

Cruise also just got an additional $1.35B investment from Softbank Vision Fund:

https://twitter.com/x/status/1488561898752782336

 

[Daniel in SD]

Tesla using HD maps confirmed!

Or that street was in the training data and no other similar street was.

 

[momo3605]

Tesla using HD maps confirmed!

Or Tesla hard coding Lombard street for the Youtube clicks confirmed

 

[Terminator857]

Tesla using HD maps confirmed!

Is there a generally accepted definition of HD maps or is that still the source of much debate? Just trolling.

 

[diplomat33]

Is there a generally accepted definition of HD maps or is that still the source of much debate?

 

[MP3Mike]

Tesla using HD maps confirmed!

Or maybe it is responding to the warning sign showing a curvy road ahead?

 

[Daniel in SD]

Is there a generally accepted definition of HD maps or is that still the source of much debate? Just trolling.

Yes, any maps more detailed than what Tesla is currently using.
also trolling