Autonomous Car Progress

[Bladerskb]

I really can't wait for Tesla to get into China and then when those comparisons happen with XNGP and others we will see how good FSDbeta is.

I currently have Huawei's NCA above XNGP, although Huawei were unable to reach their goal of releasing ADS 2.0 that supports all cities by end of 2023 (so maybe Q1 2024?).
But yes the comparison between XNGP, NCA and FSD will definitely be interesting!

I would be surprised if FSD Beta ends up better, given it's so fitted to US conventions and roads. Tesla also currently uses a more inferior navigation map provider in China last I looked. I don't think that necessarily predicts the performance if both were on equal footing. Probably releases in Europe will be more telling (there are some Chinese makes selling in Europe).

Another way to look at it: Tesla has millions of cars in China so they have unlimited access to data. They can just as easily train models from scratch using their existing NN architecture and China's data or they can finetune the US model. Either way.

Infact their driving policy going full end to end also helps because they don't have to write explicit code. They can just collect video data based on their understanding of China laws, roads, driving culture and train the China based driving policy model on that.

So I would say their performance in China would be representative of how good their system is overall. The main difference in China for them would be that things are a-lot more chaotic. So it will showcase their ability or in-ability to handle chaos.

Just so you know, I hold this same view for Mobileye aswell. They have 150k+ cars in China, so the delays they are having shows their in-ability and not just the fact that its not their home turf (Israel). So them driving in China cities will showcase their ability or in-ability to handle chaos.

 

[stopcrazypp]

I currently have Huawei's NCA above XNGP, although Huawei were unable to reach their goal of releasing ADS 2.0 that supports all cities by end of 2023 (so maybe Q1 2024?).
But yes the comparison between XNGP, NCA and FSD will definitely be interesting!


Another way to look at it: Tesla has millions of cars in China so they have unlimited access to data. They can just as easily train models from scratch using their existing NN architecture and China's data or they can finetune the US model. Either way.

Infact their driving policy going full end to end also helps because they don't have to write explicit code. They can just collect video data based on their understanding of China laws, roads, driving culture and train the China based driving policy model on that.

So I would say their performance in China would be representative of how good their system is overall. The main difference in China for them would be that things are a-lot more chaotic. So it will showcase their ability or in-ability to handle chaos.

Just so you know, I hold this same view for Mobileye aswell. They have 150k+ cars in China, so the delays they are having shows their in-ability and not just the fact that its not their home turf (Israel). So them driving in China cities will showcase their ability or in-ability to handle chaos.

Actually they can't. China recent data law crackdown does not let them take that data to their data centers in the US for training. That means they have to build up the infrastructure in China first to get anywhere close to the training resources the US has enjoyed for a while. Mobileye will have the same issue Tesla has (unless they have already built up a massive data center in China that mirrors their abilities elsewhere; I don't follow them so perhaps you know better). This is flipped for the Chinese makes, they have a home advantage in China just like Tesla does in the US.

While Tesla is talking about switching to E2E in the US, that's still very new, and they might not use that for their first release in China. Their previous implementation includes a lot of heuristics which are all fitted to US roads and conventions.

Europe is more of an equalizer because they have roads and conventions that are not quite the same as either the US nor China. Tesla also has less restrictions on using data for training, although I'm not entirely sure if the same necessarily will apply to Chinese makes going forward (given the whole hysteria over cracking down on apps like Tiktok for example).

 

[cwerdna]

Facing roadblocks, China's robotaxi darlings apply the brakes | TechCrunch

TechCrunch spoke to insiders in China's autonomous vehicle industry. Here's what has them worried, motivated and hopeful for in 2024.

techcrunch.com

 

[heltok]

Actually they can't. China recent data law crackdown does not let them take that data to their data centers in the US for training. That means they have to build up the infrastructure in China first to get anywhere close to the training resources the US has enjoyed for a while. Mobileye will have the same issue Tesla has (unless they have already built up a massive data center in China that mirrors their abilities elsewhere; I don't follow them so perhaps you know better). This is flipped for the Chinese makes, they have a home advantage in China just like Tesla does in the US.

While Tesla is talking about switching to E2E in the US, that's still very new, and they might not use that for their first release in China. Their previous implementation includes a lot of heuristics which are all fitted to US roads and conventions.

Europe is more of an equalizer because they have roads and conventions that are not quite the same as either the US nor China. Tesla also has less restrictions on using data for training, although I'm not entirely sure if the same necessarily will apply to Chinese makes going forward (given the whole hysteria over cracking down on apps like Tiktok for example).

Likely Tesla will train their e2e in US, then use those weights as a starting point to fine-tune the neural network to learn the differences between US and China. The amount of data you need to learn to drive in China if you are an expert driver in US is a lot less than the amount of data needed to become an expert driver in the US.

 

[scottf200]

The amount of data you need to learn to drive in China if you are an expert driver in US is a lot less than the amount of data needed to become an expert driver in the US.

At first I didn't follow but I think I missed your underlined part. Bloom helped with additional slants.

>>>> Bloom on Discord >>>
This statement suggests that if you are already an expert driver in the United States, the additional information or skills you need to acquire to drive in China is considerably less than what you originally needed to learn to become an expert driver in the US. This could be due to several reasons:

1. Transferable Skills: Many driving skills are universal, such as controlling the vehicle, understanding traffic flow, and reacting to road conditions. These skills transfer directly, regardless of the country.

1. Learning Curve: An expert driver has already gone through the steep learning curve of becoming proficient in driving. Learning new traffic laws or road systems is an incremental addition to their existing knowledge base.

1. Adaptation: An expert driver is likely better at adapting to new driving environments, which means they can apply their expertise to quickly understand and adjust to the differences in driving conditions and regulations in China.

1. Focused Learning: The learning required for an expert US driver to drive in China would be focused on specific differences, such as traffic laws, road signs, local driving etiquette, and perhaps adapting to driving on the opposite side of the road if that's the case.

However, this statement does not account for cultural and environmental differences that might pose significant challenges, such as language barriers, navigation in unfamiliar cities, or understanding local driving behavior, which can be quite distinct from one country to another.
<<<<

 

[heltok]

At first I didn't follow but I think I missed your underlined part. Bloom helped with additional slants.

>>>> Bloom on Discord >>>
This statement suggests that if you are already an expert driver in the United States, the additional information or skills you need to acquire to drive in China is considerably less than what you originally needed to learn to become an expert driver in the US. This could be due to several reasons:

1. Transferable Skills: Many driving skills are universal, such as controlling the vehicle, understanding traffic flow, and reacting to road conditions. These skills transfer directly, regardless of the country.
2. Learning Curve: An expert driver has already gone through the steep learning curve of becoming proficient in driving. Learning new traffic laws or road systems is an incremental addition to their existing knowledge base.
3. Adaptation: An expert driver is likely better at adapting to new driving environments, which means they can apply their expertise to quickly understand and adjust to the differences in driving conditions and regulations in China.
4. Focused Learning: The learning required for an expert US driver to drive in China would be focused on specific differences, such as traffic laws, road signs, local driving etiquette, and perhaps adapting to driving on the opposite side of the road if that's the case.

However, this statement does not account for cultural and environmental differences that might pose significant challenges, such as language barriers, navigation in unfamiliar cities, or understanding local driving behavior, which can be quite distinct from one country to another.
<<<<

Thanks. Yeah, my post was missing a few commas to make it more readable. ^^

 

[stopcrazypp]

At first I didn't follow but I think I missed your underlined part. Bloom helped with additional slants.

>>>> Bloom on Discord >>>
This statement suggests that if you are already an expert driver in the United States, the additional information or skills you need to acquire to drive in China is considerably less than what you originally needed to learn to become an expert driver in the US. This could be due to several reasons:

1. Transferable Skills: Many driving skills are universal, such as controlling the vehicle, understanding traffic flow, and reacting to road conditions. These skills transfer directly, regardless of the country.
2. Learning Curve: An expert driver has already gone through the steep learning curve of becoming proficient in driving. Learning new traffic laws or road systems is an incremental addition to their existing knowledge base.
3. Adaptation: An expert driver is likely better at adapting to new driving environments, which means they can apply their expertise to quickly understand and adjust to the differences in driving conditions and regulations in China.
4. Focused Learning: The learning required for an expert US driver to drive in China would be focused on specific differences, such as traffic laws, road signs, local driving etiquette, and perhaps adapting to driving on the opposite side of the road if that's the case.

However, this statement does not account for cultural and environmental differences that might pose significant challenges, such as language barriers, navigation in unfamiliar cities, or understanding local driving behavior, which can be quite distinct from one country to another.
<<<<

The above talks about how a human would adapt to driving in a different country, that doesn't necessarily apply to AI. Humans are very good at generalized knowledge and applying existing knowledge to other situations (even ones they have never encountered before). So far for AI that has been a huge challenge (other threads talk about the difficulty of achieving general intelligence).

Likely Tesla will train their e2e in US, then use those weights as a starting point to fine-tune the neural network to learn the differences between US and China. The amount of data you need to learn to drive in China if you are an expert driver in US is a lot less than the amount of data needed to become an expert driver in the US.

Does it really work that way if the input is directly video to control output? Or would the E2E network essentially be "contaminated" by US based training, especially for corner cases, where a E2E network trained exclusively on China based training would handily beat a US based one that was tweaked, when used on Chinese roads.

 

[heltok]

The above talks about how a human would adapt to driving in a different country, that doesn't necessarily apply to AI. Humans are very good at generalized knowledge and applying existing knowledge to other situations (even ones they have never encountered before). So far for AI that has been a huge challenge (other threads talk about the difficulty of achieving general intelligence).

Does it really work that way if the input is directly video to control output? Or would the E2E network essentially be "contaminated" by US based training, especially for corner cases, where a E2E network trained exclusively on China based training build handily beat a US based one that was tweaked when used on Chinese roads.

For image recognition it was standard practice to start with the weights from the millions of images of dogs and cats etc in imagenet, then do fine tuning on thousands of images of cataracts in eyes etc. Once you have learnt the concepts of edges, corners, lane lines, kids playing, complex intersections, it will be a lot easier to learn how to drive in China.

Same with optimus. Once they have mastered a few thousands tasks, adding another task will be a lot easier than learning the first.

Some old video talking about this from Karpathys course:

 

[stopcrazypp]

For image recognition it was standard practice to start with the weights from the millions of images of dogs and cats etc in imagenet, then do fine tuning on thousands of images of cataracts in eyes etc. Once you have learnt the concepts of edges, corners, lane lines, kids playing, complex intersections, it will be a lot easier to learn how to drive in China.

Same with optimus. Once they have mastered a few thousands tasks, adding another task will be a lot easier than learning the first.

Some old video talking about this from Karpathys course:

I don't doubt Tesla can get a very basic system out in China much quicker than a Chinese make starting from scratch, however, getting one that is considered good is a whole different matter. FSD Beta had been stuck in plateau where there are still many corner cases it can't handle, even when on US roads where it was trained exclusively. I guess we'll see how good E2E is at improving on this (not overfitting to certain roads and situations) and how adaptable it would be to different countries.

 

[diplomat33]

XNGP available in 243 Chinese cities from Jan 2, 2024.

https://twitter.com/x/status/1742206615728824440

 

[mspohr]

XNGP available in 243 Chinese cities from Jan 2, 2024.

https://twitter.com/x/status/1742206615728824440

What's xngp?

 

[spacecoin]

What's xngp?

XPeng's version of Level 2 in the city like FSDb.

"Xpeng Navigation Guided Pilot"

 

[diplomat33]

What's xngp?

As @spacecoin said, it is Xpeng's version of Tesla FSD beta. To elaborate a bit more, it is a L2 system that can drive autonomously on highways and city streets. You enter a destination and the car will drive you there hands free but you need to keep your eyes on the road and supervise at all times and be ready to take over if needed. One difference with Tesla is that Xpeng adds radar and lidar in addition to camera vision.

 

[diplomat33]

More info from Xpeng's press release:

XNGP adds 191 new cities to its coverage starting today, bringing the total to 243 cities, the Chinese electric vehicle (EV) maker announced in a statement today. Xpeng's goal is for XNGP to provide driver assistance in all scenarios, including highways, city roads, internal campus roads, and parking lots.

The feature was initially available only on highways, then known as Highway NGP. Xpeng then gradually expanded its coverage to urban areas, also known as City NGP, with the first five cities covered being Guangzhou, Shenzhen, Foshan, Shanghai, and Beijing.

...

From a technical perspective, XNGP is already capable of supporting urban smart driving nationwide, Xpeng said, adding that it will make the feature available in different cities in batches after engineering tests have completed validation.

To date, XNGP has completed more than 3 million kilometers of road verification, including highway kilometers, according to the company.

The feature's simulated driving mileage has reached 145 million kilometers, with more than 16,000 core simulation scenarios and more than 40,000 specialized simulation scenarios, Xpeng said.

 

[Bladerskb]

Actually they can't. China recent data law crackdown does not let them take that data to their data centers in the US for training. That means they have to build up the infrastructure in China first to get anywhere close to the training resources the US has enjoyed for a while. Mobileye will have the same issue Tesla has (unless they have already built up a massive data center in China that mirrors their abilities elsewhere; I don't follow them so perhaps you know better). This is flipped for the Chinese makes, they have a home advantage in China just like Tesla does in the US.

I took that into account. They have been in China for years and have already been offering NOA features, so they definitely have an infrastructure/compute setup in China. Obviously it won't be nowhere as big as what they have in the States. Mobileye on the other hand is having the OEMs (in this case Geely) handle the datacenter and infrastructure requirements. For example, their REM maps are stored and built in Geely's datacenters.

While Tesla is talking about switching to E2E in the US, that's still very new, and they might not use that for their first release in China. Their previous implementation includes a lot of heuristics which are all fitted to US roads and conventions.

Maybe/Maybe not. I think we will know in a month or so.

Europe is more of an equalizer because they have roads and conventions that are not quite the same as either the US nor China. Tesla also has less restrictions on using data for training, although I'm not entirely sure if the same necessarily will apply to Chinese makes going forward (given the whole hysteria over cracking down on apps like Tiktok for example).

Yup.

 

[stopcrazypp]

I took that into account. They have been in China for years and have already been offering NOA features, so they definitely have an infrastructure/compute setup in China. Obviously it won't be nowhere as big as what they have in the States. Mobileye on the other hand is having the OEMs (in this case Geely) handle the datacenter and infrastructure requirements. For example, their REM maps are stored and built in Geely's datacenters.

I'm under the impression they have not been doing any significant Chinese nor even Europe specific training for NOA. NOA is much simpler and more universal than city driving and would require a lot less training resources than FSD Beta. China was also less strict in previous years, but recently cracked down on Tesla.

Maybe/Maybe not. I think we will know in a month or so.

I would be surprised if they launch straight out with E2E in China. In late 2023, there were rumblings they were already working on a FSD Beta release for China and the timing seems off to straight up use E2E so soon.

Yup.

 

[diplomat33]

In case anyone is interested, Wayve is going to give a talk on LINGO:

What's the latest with #LINGO? Next Monday Wayve's Principal Applied Scientist
@SinavskiOleg will be giving a keynote at #WACV2024

Find out more information about the workshop here: https://llvm-ad.github.ioAnd information on how to register here: https://lnkd.in/eASGFxnm

https://twitter.com/x/status/1742594829958119689

 

[DanCar]

Somewhat off topic: Mobileye shares down 25% today because customer orders are down.

Mobileye shares plunge after chipmaker warns of order pullback

The auto chipmaker, majority owned by Intel, warned that customer order numbers would fall short of the year-ago quarter, citing excess inventory.

www.cnbc.com

Customers stocked up on chips after global supply chain issues, per the company, and are now opting to work through excess inventory.

 

[kabin]

Somewhat off topic: Mobileye shares down 25% today because customer orders are down.

Mobileye shares plunge after chipmaker warns of order pullback

The auto chipmaker, majority owned by Intel, warned that customer order numbers would fall short of the year-ago quarter, citing excess inventory.

www.cnbc.com

Customers stocked up on chips after global supply chain issues, per the company, and are now opting to work through excess inventory.

That's a spanking and will likely go lower.

Are they still blaming Tier 1 customer's for stockpiling chips from the covid days? Or spin from failed leadership?

 

[powertoold]

I'm bad at predicting things, but I got these things kinda right:

1) Cruise basically shutting down by end of 2023
2) Mobileye sucking