Autonomous Car Progress

[rxlawdude]

I don't think Smart Summon activates the hazard lights. (It does use turn signals sometimes when turning, I think.)

It definitely uses turn signals.

 

[rxlawdude]

Depends on what you define for "right". which is the crux of the argument (and why as I mentioned trying to pick "winners" in a comparison chart doesn't work given people have different criteria).

If "right" is releasing L5, then none of the players are "right" given none of them have reached L5 (or even just wide release L4). The basic paths being taken I see is:
1) Tesla's approach: work to release end-to-end L2, then improve reliability of that until it is to L4 then eventually L5 level. The idea is that they can convert their fleet at a flip of a switch as soon as the software is ready.
2) Others: release L4 in geofenced area relying heavily on HD maps, and achieve L5 by mapping the entire world or at least specific countries. Note that many companies don't have a explicit goal to achieve L5, they would be perfectly happy with L4 (robotaxies can operate just fine in geo fences), in which case the goal would instead be wide release L4 (releasing L4 in many more places than just experimental fleets, like Uber/Lyft's current coverage area vs when they originally started a decade ago).

Remember that mapping ain't the be-all-end-all - witness Waymo's problems with construction post mapping, for example.

 

[rxlawdude]

That's the biggest problem with all of Mobileye's solutions, it takes forever to reach the consumer.

Of course, Tesla was the first (or one of) with Gen 1 of ME's product.

 

[powertoold]

Remember that mapping ain't the be-all-end-all - witness Waymo's problems with construction post mapping, for example.

What I find most interesting about Waymo's challenges is that despite having everything fully mapped, they're still avoiding many unprotected lefts and lane merges, to name a few. It really doesn't bode well for the other fsd developers with similar approaches hoping to catch up to Waymo.

 

[diplomat33]

This is Ford's Co-pilot/Mobileye in 2021... Utterly ridiculous, but I'm repeating myself what I said that Mobileye hasn't progressed in years. They don't even have the basics to match Tesla base AP.

That is not Mobileye's system. That is Ford's system, powered by the Mobileye chip. That is an important distinction. That is the mistake that you and @powertoold are making. These are NOT Mobileye's in-house system. They do not represent what Mobileye has. They represent what the legacy automakers have in terms of ADAS.

Mobileye has progressed a lot. Mobileye's system is called Super Vision. It is what you see in the NYC and Munich FSD drives.

 

[powertoold]

That is the mistake that you and @powertoold are making. These are NOT Mobileye's in-house system. They do not represent what Mobileye has. They represent what the legacy automakers have in terms of ADAS.

Can you expand on this? So consumer-deployed ADAS systems powered by Mobileye aren't representative of Mobileye? Are you saying legacy auto developed their own lane keeping and detection NNs? I'm confused.

 

[diplomat33]

Can you expand on this? So consumer-deployed ADAS systems powered by Mobileye aren't representative of Mobileye? Are you saying legacy auto developed their own lane keeping and detection NNs? I'm confused.

Yes. The legacy automakers develop their own ADAS software that is just powered by the ME chip. The only ADAS that is actually from Mobileye is Super Vision.

 

[linux-works]

If only others were as honest as you were in the quote above. Because most are sopranos in the torch and pitchfork choir.

sopranos?

more like uncle Jun.

 

[powertoold]

Yes. The legacy automakers develop their own ADAS software that is just powered by the ME chip. The only ADAS that is actually from Mobileye is Super Vision.

Do you have some links to back this up? It's news to me.

So from what you're saying, Mobileye is like Nvidia. Nvidia just develops the GPU, and videogame developers create the game to run on the GPU.

Even with the upcoming Zeekr car, it still says "powered by Mobileye SuperVision" but they have their own name for the ADAS ("CoPilot"). Does this mean Zeekr is also developing their own system based on the Mobileye hardware?

Mobileye, Geely to Offer Most Robust Driver-Assistance Features

The Lynk & Co Zero Concept EV will feature Lynk & Co’s CoPilot solution powered by Mobileye SuperVision surround-view advanced driver-assistance system with over-the-air update capabilities.

www.intel.com

 

[Daniel in SD]

What I find most interesting about Waymo's challenges is that despite having everything fully mapped, they're still avoiding many unprotected lefts and lane merges, to name a few. It really doesn't bode well for the other fsd developers with similar approaches hoping to catch up to Waymo.

They haven't mapped cars or pedestrians.

 

[diplomat33]

Do you have some links to back this up? It's news to me.

So from what you're saying, Mobileye is like Nvidia. Nvidia just develops the GPU and videogame developers create the game to run on the GPU.

Yes. Read the SuperVision pdf. It says that SuperVision is "end to end", including the full ME hardware and software stack, maps, driving policy, the works. So it is ME's entire in-house system.

End to end system, including full Mobileye design, hardware architecture & implementation, ECU design, functional stack, decision layers and end-user function

https://newsroom.intel.com/wp-content/uploads/sites/11/2020/09/supervision-product-brief.pdf

The other ADAS systems don't include this full ME stack.

That's why Super Vision is different. It is the first time that ME is actually going to deploy their in-house ADAS instead of just letting other automakers use their chip.

They haven't mapped cars or pedestrians.

Cars and pedestrians are never mapped. That is handled by the cameras, lidar and radar.

 

[powertoold]

The other ADAS systems don't include this full ME stack.

I still don't agree with what you're claiming about legacy auto implementations of Mobileye systems.

That's the whole point of Mobileye advertising EyeQ4 with a list of features. That means the chip comes with software features that automakers can rebadge or adjust to their own liking.

ZF and Mobileye Safety Technology Chosen by Toyota

www.intc.com

"Mobileye’s EyeQ4 applies enhanced computational capabilities with computer vision algorithms while rapidly processing information from the vehicle’s front-facing camera. With a variety of feature sets including vehicle detection from any angle and next-generation lane detection, EyeQ4 enables automakers to take a major step forward in autonomy, achieving the ability to support and enhance complicated driving tasks."

 

[linux-works]

mobile eye has many offerings, most of which none of us will ever see since we have not signed NDAs and most of the readership, here, is sorely lacking in actual info about the industry.

its often useful to admit when you dont have the full story and just listen for those that might. unless you really have signed the NDAs, and in which case, why are you blabbing about the details, here?

see the problem?

if the info is here, its out of date (stale) or just public info, which is, well, not very useful to judge anyone's state of the art.

just sayin'.

 

[stopcrazypp]

mobile eye has many offerings, most of which none of us will ever see since we have not signed NDAs and most of the readership, here, is sorely lacking in actual info about the industry.

its often useful to admit when you dont have the full story and just listen for those that might. unless you really have signed the NDAs, and in which case, why are you blabbing about the details, here?

see the problem?

if the info is here, its out of date (stale) or just public info, which is, well, not very useful to judge anyone's state of the art.

just sayin'.

But given there are no ADAS in consumer hands as advanced as Tesla's other than Enhanced Super Cruise (which as mentioned is only recently adding some features like auto lane change that Tesla had 6 years ago), it's pretty safe to say Mobileye haven't made dent yet (at least here in USA, in China there are a few examples of ADAS systems that are as good as Tesla's, but not necessarily using Mobileye, most of them are using Nvidia, from a quick search). No doubt their demos are impressive, but how long will it be before consumers can buy it and will it be affected by the customizations automakers may make that may make them inferior, like for example that Copilot one using EyeQ4?

 

[stopcrazypp]

I still don't agree with what you're claiming about legacy auto implementations of Mobileye systems.

That's the whole point of Mobileye advertising EyeQ4 with a list of features. That means the chip comes with software features that automakers can rebadge or adjust to their own liking.

ZF and Mobileye Safety Technology Chosen by Toyota

www.intc.com

"Mobileye’s EyeQ4 applies enhanced computational capabilities with computer vision algorithms while rapidly processing information from the vehicle’s front-facing camera. With a variety of feature sets including vehicle detection from any angle and next-generation lane detection, EyeQ4 enables automakers to take a major step forward in autonomy, achieving the ability to support and enhance complicated driving tasks."

That's an interesting point. From what I see in that datasheet, the "SuperVision" is basically two EyeQ5 High chips and 11 cameras, but it doesn't explicitly say automakers can't customize parameters or limit how the vehicle control happens. Looking at the Copilot video, the failure to make the tighter turns can easily simply be from an automaker set limit on the amount of steering angle that can be used by the system. If this sort of customization is allowed in the SuperVision product, then the same kind of failures can occur even if they offer the entire stack (including the sensors) for perception, decisions, UI.

I see the same issue in videos of the Nvidia based systems that the Chinese automakers are using, they consistently fail miserably on tighter turns because it seems like it's hitting some sort of steering angle limit (and the software isn't able to work through the turn with repetitive smaller adjustments). It is not a failure of perception given the road seems to draw correctly and they have HD/MD maps of the roads (in far better detail than Tesla, as Tesla uses regular Baidu maps, which seems far inferior to the Amap maps that a lot of others are using).

 

[powertoold]

Looking at the Copilot video, the failure to make the tighter turns can easily simply be from an automaker set limit on the amount of steering angle that can be used by the system. If this sort of customization is allowed in the SuperVision product, then the same kind of failures can occur even if they offer the entire stack (including the sensors) for perception, decisions, UI.

In the video you posted, it seems copilot fails at minor curves as well. I'm not sure if this is intentional, lol, especially if they want to compete with AP.

People constantly underestimate what Tesla has achieved, even with basic AP. Karpathy has had many slides on the difficulty of labeling lane lines.

 

[linux-works]

But given there are no ADAS in consumer hands as advanced as Tesla's other than Enhanced Super Cruise (which as mentioned is only recently adding some features like auto lane change that Tesla had 6 years ago),

china, for one, but agreed that pretty much everyone in the US would have no visibility into that. and the china stuff is mixed, but its not day and night diff from tesla. point is: tesla does not own all the engineers that can pull off lane-keeping. its not rocket science and expect to see it as a regular feature as more car companies trade people, code, standards, even api's.

and again, tesla's gain is also their loss. all players who started after have a natural advantage (they saw what others tried and what works and what does not. they avoid problems and avoid having to rework code and even hw).

 

[diplomat33]

I still don't agree with what you're claiming about legacy auto implementations of Mobileye systems.

That's the whole point of Mobileye advertising EyeQ4 with a list of features. That means the chip comes with software features that automakers can rebadge or adjust to their own liking.

Yes. That is what I am saying. EyeQ4 might come with some basic features but the legacy automakers can build upon them to create their own ADAS. So you can't say that it is Mobileye's ADAS since the it has been changed by the automakers.

ME has Super Vision that uses eyeQ5 and which they claim is their in-house end-to-end ADAS, that is what we should look at in terms of ME's true capabilities. We should judge ME based on Super Vision not some knock off ADAS that legacy makers develop with the ME chip.

Think of this way: If Tesla licensed out the FSD computer with some basic NN's on it and let automakers use it for their own ADAS. Would you consider that to be Tesla's true capabilities or would you consider Tesla's in-house FSD Beta to be Tesla's true capabilities? I think it is obvious that you would consider FSD Beta to be Tesla's true capability, not some ADAS that an automaker develops with the FSD computer.

Now, imagine if someone was pointing to some inferior ADAS that an automaker developed with Tesla's FSD computer and claimed that it was proof that Tesla had made no progress, ignoring FSD Beta. That is what you are doing with ME, using other ADAS developed with the ME chip and claiming that it is proof of no progress, completely dismissing all the FSD work that ME is doing.

I would also judge ME when they deploy L4 robotaxis. That will be ME's true FSD capabilities.

If this sort of customization is allowed in the SuperVision product, then the same kind of failures can occur even if they offer the entire stack (including the sensors) for perception, decisions, UI.

I guess that is possible but that would seem very foolish for ME to do. If automakers customize Super Vision and it causes accidents then it would look bad on ME. I mean, if Tesla licensed FSD out to automakers would you want Tesla to let other automakers tweak it or change it where maybe it causes accidents that would not have happened otherwise and it gives Tesla bad name? I doubt it.

I would think it would make more sense for ME to limit the customizations. After all, it is a full stack system that ME has developed and validated. ME has even said that the goal is to offer a full "plug and play" system for automaker. So it would make sense for Super Vision to be "take it or leave it".

 

[powertoold]

Yes. That is what I am saying. EyeQ4 might come with some basic features but the legacy automakers can build upon them to create their own ADAS. So you can't say that it is Mobileye's ADAS since the it has been changed by the automakers.

Nah, I think it's mostly a rebadge with their own logos or branding. It's why almost all the Mobileye-based systems have similar interfaces on the display.

If you can find some evidence that Ford / Toyota / etc. is training their own lane detection NNs or something, then what you're saying will be more convincing. Otherwise, all these Mobileye-based ADAS systems are just using the Mobileye detections / predictions.

 

[diplomat33]

Nah, I think it's mostly a rebadge with their own logos or branding. It's why almost all the Mobileye-based systems have very similar interfaces on the display.

If you can find some evidence that Ford or Toyota is training their own lane detection NNs or something, then what you're saying will be more convincing. Otherwise, all these Mobileye-based ADAS systems are just using the Mobileye detections / predictions.

You are still missing my real point: we should judge ME based on their cutting edge, not based on what legacy automakers have deployed with the ME chip, especially if it is the older eyeQ4 chip and not the new eyeQ5 chip. Just because some legacy automaker's ADAS with the ME chip has poor lane keeping does not mean that ME sucks at lane keeping and has made no progress. We know that because we can look at the FSD demos that ME has put out and we can clearly see that they have way better than the cruise control and lane keeping on legacy automaker cars.