Autonomous Car Progress

[powertoold]

The split did not happen because Tesla was trying to improve ME's system. ME broke up with Tesla because they felt that Tesla was misusing their system to do FSD that it was not designed to do and compromising safety:



Source: Mobileye spills the beans: Tesla was dropped because of safety concerns

Yea, that's what they tell you. Did we get both sides of the story?

Regardless, I find it illogical and silly to believe the Mobileye carmakers develop their own ADAS. Also, if you know anything about NN training, you'd need the dataset to improve on it, so Mobileye is providing carmakers with their dataset? Or are the carmakers collecting their own dataset? None of this makes any sense if you say that the ADAS performance is not representative of Mobileye. It may not be 100% representative, but the prediction NNs are Mobileye's.

Edit: I also find it funny that you take Mobileye's word at face value. You need to read between the lines there. Mobileye wanted to develop FSD, and Tesla was wanting control over AP to eventually develop FSD. Obviously Mobileye doesn't want to help a competitor develop FSD faster, so they come up with some safety excuse to break up with Tesla, even though Tesla had not even deployed fsd. In fact, it took over 4 years later for Tesla to deploy it. Lol!

 

[diplomat33]

Yea, that's what they tell you. Did we get both sides of the story?

Regardless, I find it illogical and silly to believe the carmakers develop their own ADAS. Also, if you know anything about NN training, you'd need the dataset to improve on it, so Mobileye is providing carmakers with their dataset? Or are the carmakers collecting their own dataset? None of this makes any sense if you say that the ADAS performance is not representative of Mobileye.

Training data sets are easy to get and machine learning is more and more common now. So I would not be surprised if automakers are doing their own NN training.

 

[powertoold]

Training data sets are easy to get and machine learning is more and more common now.

Understatement of the thread lol

I can't take you seriously while you still believe stuff like this.

 

[powertoold]

So I would not be surprised if automakers are doing their own NN training.

I don't know where you got the idea that Mobileye sells only the chip and the carmarker develops the ADAS / NNs / logic / etc.

I thought you were more fact-based?

Basically, what you're saying disregards all those posts bladerskb made showing a list of EyeQ4 features: traffic light / lane detection / sign detection / etc. etc.

Also, what you're saying is contradicted by Amnon himself, where he says the Eye SoCs come EMBEDDED with ADAS software:

 

[rxlawdude]

... Waymo does not avoid unprotected left turns when there is a safety driver, they only avoid some situations when it is a driverless ride. ... Waymo does not want to take the chance with a driverless ride.

After all, when you are testing with a safety driver, you can take some risks.

Thank you for endorsing the existing functionality of Tesla's FSD.

Waymo, when all is said and done from your own description, is no safer than what Chuck's Tesla beta FSD videos (de facto with a "safety driver") demonstrate.

We're making some progress here!

 

[diplomat33]

Thank you for endorsing the existing functionality of Tesla's FSD.

Waymo, when all is said and done from your own description, is no safer than what Chuck's Tesla beta FSD videos (de facto with a "safety driver") demonstrate.

We're making some progress here!

Don't twist my words. I am not saying that at all.

 

[Daniel in SD]

Thank you for endorsing the existing functionality of Tesla's FSD.

Waymo, when all is said and done from your own description, is no safer than what Chuck's Tesla beta FSD videos (de facto with a "safety driver") demonstrate.

We're making some progress here!

https://storage.googleapis.com/sdc-prod/v1/safety-report/Waymo-Public-Road-Safety-Performance-Data.pdf

How safe do you estimate FSD beta is relative to this? Why?

 

[rxlawdude]

https://storage.googleapis.com/sdc-prod/v1/safety-report/Waymo-Public-Road-Safety-Performance-Data.pdf

How safe do you estimate FSD beta is relative to this? Why?

Really? You're using a self-published vendor's article as an authoritative source, rather than objective, academic third-party researchers?

Kind of like using a new drug's Phase III studies where the company's own employees conducted participant selection, stratification, analysis and drug safety monitoring (DSMB). In the industry, it's called the fox guarding the hen house.

 

[diplomat33]

I don't know where you got the idea that Mobileye sells only the chip and the carmarker develops the ADAS / NNs / logic / etc.

I thought you were more fact-based?

Basically, what you're saying disregards all those posts bladerskb made showing a list of EyeQ4 features: traffic light / lane detection / sign detection / etc. etc.

Also, what you're saying is contradicted by Amnon himself, where he says the Eye SoCs come EMBEDDED with ADAS software:

There is no contradiction. I will quote @Bladerskb :

The EyeQ vision chips doesn't do driving policy. They contain only perception NNs. Absolutely no planning or control algorithms. Ford's Copilot is either using driving policy developed by Ford or a Tier 1. Just like BlueCruise is using driving policy developed by Ford.

Mobileye for the first time is offering driving policy and its only happening through the supervision package.

So yes, ME does provide some "basic features" but the automaker can build on it. Super Vision is ME's "total package" ADAS.

I don't think you can use the automaker's ADAS to judge ME's best when they are not using ME's "total package". If an automaker was only using Tesla's "autosteer" and adding their own driving policy and controls, would you say that represents Tesla's best? Of course not!

 

[rxlawdude]

Don't twist my words. I am not saying that at all.

I did not at all twist your words. Nor the context. Waymo needs safety drivers in challenging intersections. Cool. So they are offering no real-world advantage over FSD.

What did I twist, oh Waymo Master?

There is no contradiction. I will quote @Bladerskb

Credibility lost.

 

[linux-works]

Credibility lost.

you owe me a new keyboard.

 

[Bladerskb]

Mobileye provides a base feature set. I guess the carmaker can improve on it if they wanted? But then again Tesla was trying to improve it to the point where they take control of it, and that's when they split.

You don't improve on it, you utilize the perception outputs to drive your car, big difference.

Regardless, I find it illogical and silly to believe the Mobileye carmakers develop their own ADAS. Also, if you know anything about NN training, you'd need the dataset to improve on it, so Mobileye is providing carmakers with their dataset? Or are the carmakers collecting their own dataset? None of this makes any sense if you say that the ADAS performance is not representative of Mobileye. It may not be 100% representative, but the prediction NNs are Mobileye's.

Mobileye's sells three variants of their chips (Low, Mid, and HI versions(more sophisticated NN)) with variety of vision NNs for ~$30 (EyeQ3) and ~$60 (EyeQ4).
Tiers 1s and automakers write planning and control algorithms that uses any number of the NN outputs of the chip.

You have tier 1s like Aptiv or ZF who will use the variants of the chips to create basic, moderate and advanced packages that they sell to OEMs. A Basic package for example would contain: AEB, FCW, LKA, ACC.

Then you have automakers like Tesla, Nio, Ford, GM, etc who write their own driving policy (planning and control) using the outputs from the EyeQ chip.
The result is Navigate on Nio pilot, GM Supercruise, Ford Blue Cruise, Tesla AP1, etc.

Its simple. Why make this complicated?

 

[diplomat33]

I did not at all twist your words. Nor the context. Waymo needs safety drivers in challenging intersections. Cool. So they are offering no real-world advantage over FSD.

What did I twist, oh Waymo Master?

Because you are making it sound like Waymo's FSD is basically no better than Tesla's FSD which is wrong. Waymo is L4, Tesla is L2. Waymo can do driverless and Tesla can't. And Waymo's FSD has a much better disengagement rate than Tesla's FSD.

 

[powertoold]

I don't think you can use the automaker's ADAS to judge ME's best when they are not using ME's "total package". If an automaker was only using Tesla's "autosteer" and adding their own driving policy and controls, would you say that represents Tesla's best? Of course not!

So where's the part where you talk about carmakers training their own NNs and collecting their own training dataset? Since it's so easy and whatnot.

 

[rxlawdude]

Because you are making it sound like Waymo's FSD is basically no better than Tesla's FSD which is wrong. Waymo is L4, Tesla is L2. Waymo can do driverless and Tesla can't. And Waymo's FSD has a much better disengagement rate than Tesla's FSD.

No, I'm not making it sound anything. Your own words literally stated that autonomous functionality at complex intersections or unprotected left turns is too risky for Waymo, even within their little geofenced utopia.

 

[Bladerskb]

Really? You're using a self-published vendor's article as an authoritative source, rather than objective, academic third-party researchers?

Kind of like using a new drug's Phase III studies where the company's own employees conducted participant selection, stratification, analysis and drug safety monitoring (DSMB). In the industry, it's called the fox guarding the hen house.

I bet you haven't even read the paper. Tesla tells you in one sentence of their entire paragraph "safety report" that they are "In the 1st quarter, we registered one accident for every 4.19 million miles driven in which drivers had Autopilot engaged." You immediately run with it and regurgitate it with all of your fellow fans. over and over again. "Tesla AP is 10x safer, blah blah blah"

Waymo on the other hand gives you a comprehensive detail of every single incident, actual data into every incident.
Your response is that its marketing trash...

This is Tesla's fans logic at its finest.

 

[Daniel in SD]

Really? You're using a self-published vendor's article as an authoritative source, rather than objective, academic third-party researchers?

Kind of like using a new drug's Phase III studies where the company's own employees conducted participant selection, stratification, analysis and drug safety monitoring (DSMB). In the industry, it's called the fox guarding the hen house.

What do you propose as an alternative system for evaluating safety?
Drug trials can be done by a third party. I can't really think of how that would work for autonomous vehicles.

 

[Matias]

Edit: I also find it funny that you take Mobileye's word at face value. You need to read between the lines there. Mobileye wanted to develop FSD, and Tesla was wanting control over AP to eventually develop FSD. Obviously Mobileye doesn't want to help a competitor develop FSD faster, so they come up with some safety excuse to break up with Tesla, even though Tesla had not even deployed fsd. In fact, it took over 4 years later for Tesla to deploy it. Lol!

Anyway, it is a pity, that Tesla and Mobileye ended their collaboration. I’m sure world would now have better autonomous cars on the road, if that was not the case.

 

[diplomat33]

No, I'm not making it sound anything. Your own words literally stated that autonomous functionality at complex intersections or unprotected left turns is too risky for Waymo, even within their little geofenced utopia.

What I said is that Waymo probably does not want to take the risk of some situations when they have passengers on board. But being "too risky" does not necessarily mean that Waymo's FSD can't do it. And we've seen Waymo do lots of unprotected left turns and always handled them very well. So we know Waymo can do them well.

 

[powertoold]

Anyway, it is a pity, that Tesla and Mobileye ended their collaboration. I’m sure world would now have better autonomous cars on the road, if that was not the case.

I'm glad they ended. Again, I posted a video where I find Amnon talking a bunch of nonsense. But then again, Mobileye has its hands tied because they're not able to collect the data they need to achieve deployable fsd.

Based on what I've seen in Mobileye's talks, they don't collect images or videos from their ADAS fleet. They constantly emphasize how little data they collect, 10kb per mile. Basically, they only collect data points to create a vector map. It's not that they do this by choice, but they're limited in how much data they can collect and transmit from their ADAS fleet.