Autonomous Car Progress

[diplomat33]

You were sure Tesla will a couple of years back.

Yeah but that was back when I was a newbie Tesla owner who was ignorant. Back then I believed Elon's FSD predictions because I did not know any better. I know better now. And I know a lot more about autonomous driving now than I did then.

When it comes to AVs, the immutable law is that it pays to be skeptical.

Sure, nothing is certain. It is entirely possible that Mobileye could face delays. In fact, chances are that they will face delays because that is part of life. But Mobileye has a viable approach. It will just be a matter of how well they can execute. If they execute their plan, there is a very good chance that they can do "eye's off" by 2025-26 IMO.

 

[diplomat33]

Brad Templeton's 2022 review of AVs:

0:00 - Intro
0:54 - The bad news
7:50 - There's good news
12:30 - Not so ugly

 

[scottf200]

You were sure Tesla will a couple of years back.

To be fair, Tesla/Elon has changed their emphasis on what is important for FSD a few times. Prime example is HD maps usage and how they can build them.
I've had a few FSD drives recently with upcoming turn (3 come to mind immediately) and FSD made bad enough lane choices where the turn was MISSED and it got confused on how to even handle it while we were nearly into the intersection. These cases were when the nav and displayed map was clearly showing exactly how far the turn was.

 

[kabin]

I like what I'm seeing from MobileEye. It's scientific. Analyze the problem, determine requirements/specifications, break it up into solvable bites, create modular hardware to specifications, and use the best sensors for each scenario. There's no apparent dogmatic biases. I love it!

From what I can tell Tesla did it all backwards with little apparent upfront R&D, a rush to release/manufacture, quick to overstate capabilities/improvements, and arguably never acknowledge limitations. It was a rooky effort. The main driving factors were (are) cost and making deliveries. Tesla assumed a solution, hired academics predominantly in one field, built some nice hardware from inaccurate requirements/specifications, touted and promoted a design, and rolled the dice with manufacturing.

TSLA has the money to eventually get there but it will take much longer, development costs will be more, and in the mean time FSD customers may have been sold a bill of goods. What was a $10-15k FSD option with safe redundant hardware, promoted as eventually capable for robotaxis income is now arguably an overpriced dead-end, clunky, hands/eyes-on advanced driver assistance system.

 

[EVNow]

Yeah but that was back when I was a newbie Tesla owner who was ignorant. Back then I believed Elon's FSD predictions because I did not know any better. I know better now. And I know a lot more about autonomous driving now than I did then.

Sure, nothing is certain. It is entirely possible that Mobileye could face delays. In fact, chances are that they will face delays because that is part of life. But Mobileye has a viable approach. It will just be a matter of how well they can execute. If they execute their plan, there is a very good chance that they can do "eye's off" by 2025-26 IMO.

Don’t you find it strange that no US OEM is working with them on this ? Doesn’t sound very promising …

I also like their technical approach, but their GTM (go to market) strategy is very poor.

 

[kabin]

Brad Templeton's 2022 review of AVs:

0:00 - Intro
0:54 - The bad news
7:50 - There's good news
12:30 - Not so ugly

The slurred mumbling speech is rough but he seems to make some good points.

 

[diplomat33]

Don’t you find it strange that no US OEM is working with them on this ? Doesn’t sound very promising …

I also like their technical approach, but their GTM (go to market) strategy is very poor.

Here is an infographic of Mobileye's partnerships so far. This is not counting the secret "premium German automaker" that they teased at CES.

We see that VW, Ford and Zeekr are the only ones working with Mobileye on ADAS. There are other companies that I've never heard of that are working with Mobileye to deploy robotaxis or autonomous delivery vehicles.

We can also see that VW and Ford are only working with Mobileye on REM enhanced ADAS, not SuperVision. Not sure why VW and Ford are not jumping on board with SuperVision. Personally, I think VW and Ford would be smart to get on board with SuperVision especially now that they dumped Argo. Furthermore, Ford's Blue Cruise sucks so they could definitely use a better ADAS. Without Argo to give them a path to L4 and with their own ADAS lagging behind, I would think partnering with Mobileye for SuperVision and later Chauffeur would be the smart move. Maybe we will hear something from them about SuperVision at some point?

I am frustrated that Mobileye does not seem to care about bringing SuperVision to the US market any time soon. I get that they are an Israeli company so geographically, they are closer to Europe and Asia. That might explain their deployment strategy so far (SuperVision on Zeekr's in China, robotaxis in Tel Aviv and Munich), But the US is a big market so it seems odd to ignore it. And Mobileye is touting SuperVision as the most advanced ADAS ever and yet they appear to be slow at actually deploying it. So I concede their go-to-market strategy seems poor.

One explanation could be that the big name automakers already have their own systems. For example, Tesla, Ford, GM, Volvo, Nissan, Toyota, Mercedes all have their own L2 systems. I can't really think of any big name automakers in the US that would still be looking for a L2 system. Now, I personally think their L2 is not as capable as what Mobileye is offering but perhaps they don't want to lose years of development for a system that is as of yet unproven like SuperVision. I can't imagine say GM giving up Super Cruise that is already deployed on a lot of cars, and the upcoming Ultra Cruise they are working on, to deploy a brand new largely untested system. So, it is possible that they are waiting for Mobileye to prove SuperVision is good enough before they give up their systems to go with Mobileye.

If Mobileye has a weakness, I think it is their dependence on automakers to get their tech out there. They could have the best AV tech in the world but if OEMs, for whatever reason (could be business reasons that have nothing to do with the tech itself), refuse to buy Mobileye's tech, then their business model fails. So Mobileye has to convince automakers to trust their tech or they can't deploy anything. It might explain why Mobileye emphasizes their validation, "true redundancy", and safety approach because they need to convince automakers that the risk of putting Mobileye AV tech on their cars is acceptable. I could see OEMs asking the legit question "if our cars crash while on SuperVision, are we going to get sued or you?" Maybe that is why Shashua acknowledged the liability risks if there is a reproducible error that causes an accident and emphasized that Mobileye is committed to zero reproducible errors.

 

[EVNow]

Here is an infographic of Mobileye's partnerships so far. This is not counting the secret "premium German automaker" that they teased at CES.

We see that VW, Ford and Zeekr are the only ones working with Mobileye on ADAS.

Where is the Zeekr like OEM partner in US ?

 

[diplomat33]

Where is the Zeekr like OEM partner in US ?

None. I did not say that there was. I just shared the partnerships that Mobileye has.

 

[EVNow]

None. I did not say that there was. I just shared the partnerships that Mobileye has.

That’s what I’m talking about. City navigation, anywhere and with minimum / no supervision is the holy grail … if they think they can do it in China, why don’t they have anyone in US ? OEMs don’t believe them ?

 

[Bladerskb]

Haven't "premium German automakers" been teasing L3 for ever ?

I doubt anyone deploys actual hands free / eyes free city driving by 25-26.

I don’t think anyone will have a working city streets L4 consumer car by 2026 other than Waymo.

Mobileye could have a car with all the hardware needed for L5 by 2026. But I don’t see them having working eyes off software for anything more than highway on-ramp/off-ramp by 2026.

I feel like city streets will remain L2.

Waymo is the only one who I feel like will have the software ready to do it in 2026. The hardware is another question.

 

[diplomat33]

That’s what I’m talking about. City navigation, anywhere and with minimum / no supervision is the holy grail … if they think they can do it in China, why don’t they have anyone in US ? OEMs don’t believe them ?

We don't know. But if you read my post, I tried to provide some possibilities:

1) Many US OEMs may want to develop their own "city navigation, anywhere, with minimum/no supervision" on their own, with their own in-house L2 system. See Ultra Cruise. So maybe US OEMs want to do their own in-house system?

2) Mobileye has not proven that they can do "city navigation, anywhere, with minimum/no supervision" yet. AFAIK, they have not achieved their stated goal of 10^7 hours MTBF. They believe their approach will get there but they have not actually reached it yet. So maybe US OEMs do not want a system that is yet unproven?

3) Maybe Mobileye's go-to-market strategy is just really bad and they are not good at convincing US OEMs? Or maybe they are focused on Europe and Asia right now and US OEMs is not part of their go-to-market strategy?

But in my previous post, I agreed with you that it is odd that Mobileye is not deploying SuperVision faster. I even mentioned that I am frustrated that Mobileye does not appear to be interested in bringing SuperVision to the US. It is odd to me.

 

[scottf200]

That’s what I’m talking about. City navigation, anywhere and with minimum / no supervision is the holy grail … if they think they can do it in China, why don’t they have anyone in US ? OEMs don’t believe them ?

What is the 'Compute' column defined as here: Wiki - Consumer AV - Status Tracking Thread
There is no description in that first post.

 

[scottf200]

That’s what I’m talking about. City navigation, anywhere and with minimum / no supervision is the holy grail … if they think they can do it in China, why don’t they have anyone in US ? OEMs don’t believe them ?

We don't know. But if you read my post, I tried to provide some possibilities:

1) Many US OEMs may want to develop their own "city navigation, anywhere, with minimum/no supervision" on their own, with their own in-house L2 system. See Ultra Cruise. So maybe US OEMs want to do their own in-house system?
...
But in my previous post, I agreed with you that it is odd that Mobileye is not deploying SuperVision faster. I even mentioned that I am frustrated that Mobileye does not appear to be interested in bringing SuperVision to the US. It is odd to me.

What about the cost that Mobileye charges for their different components and software being a factor?
Is SuperVision just outrageously priced, along with ongoing subscription and per vehicle cost?

As well, isn't China the largest car sales growth by far?

 

[diplomat33]

What about the cost that Mobileye charges for their different components and software being a factor?
Is SuperVision just outrageously priced, along with ongoing subscription and per vehicle cost?

I don't think so but it is hard to say. I could not find the cost that Mobileye charges for SuperVision. But looking at cost of hardware, I found a 8MP camera for about $100 and the eyeQ Ultra chip cost less than $1000 ( I could not find cost of eyeQ5 chip). SuperVision requires 11x8MP cameras and 2 eyeQ5 chips. So I would estimate the hardware cost at ~$3000. So, I don't think the hardware would be outrageously priced. Of course, maybe SuperVision costs $3000 but Mobileye charges $20,000. I think that is unlikely but you never know. We don't know what Mobileye charges for the software. Mobileye does estimate that the cost for Chauffeur will be less than $5000.

As well, isn't China the largest car sales growth by far?

I think so. So that could be a factor too. Mobileye could simply be focusing on China because car sales growth is the largest so they figure they will get more bang for their bucket there.

 

[Dewg]

I think so. So that could be a factor too. Mobileye could simply be focusing on China because car sales growth is the largest so they figure they will get more bang for their bucket there.

China also has challenging traffic patterns compared to the US, with a lot more variability and, dare I say, chaos to contend with. If any of the Amazing Race episodes I've watched where they're driving in China give any indication...

 

[diplomat33]

https://twitter.com/x/status/1612338671633707009

This does not really make sense to me. If the car is capable of handling the situation by itself, why bother asking the driver to take over? I guess they are trying to say that the car "plays it safe", that even though the car could handle the situation by itself, it will do the safer thing? But kind of the whole point of L4 is that you design the car to handle the situation by itself so you don't need a driver anymore. Asking the driver to take over when the car can handle it by itself seems odd to me.

 

[hgmichna]

I had thought about this in relation to Tesla's FSD beta as well. It is not a black-and-white decision. In a particular traffic situation you cannot simply say, the autopilot (or whatever one wants to call it) can handle this situation or it cannot. In fact, approaching a situation you can only try to quantify the risk that the autopilot will fail to navigate the situation properly.

An autopilot could continuously predict this risk while driving. If it predicts that the risk exceeds a certain threshold, it could hand over to the human driver just in case and well in time. The programmer could optimize this threshold by setting it lower or higher.

In the videos you often see that a car on autopilot handles many situations very well, but then approaches a difficult situation, like a complex street crossing or unusual obstacles in the street. Why not hand over to the human in such situations, even if the predicted risk is still low in absolute terms? Why take a risk? The machine could hand over if the perceived risk is 10% or 1% or whatever.

These difficult situations are relatively rare, at least in the US with its straightforward street geography and wide streets. In a German city center with medieval street structures it's a bit different.

Anyway, I personally would not mind at all, if the autopilot occasionally handed over to me with a warning like, "Difficult situation ahead. Please take over." Of course, this is impossible for driverless cars, but many private cars won't be driverless for a while.

 

[diplomat33]

https://twitter.com/x/status/1612451959734407169

Gotta agree. Making excuses for bad driving is not good look.

 

[Goose66]

I had thought about this in relation to Tesla's FSD beta as well. It is not a black-and-white decision. In a particular traffic situation you cannot simply say, the autopilot (or whatever one wants to call it) can handle this situation or it cannot. In fact, approaching a situation you can only try to quantify the risk that the autopilot will fail to navigate the situation properly.

An autopilot could continuously predict this risk while driving. If it predicts that the risk exceeds a certain threshold, it could hand over to the human driver just in case and well in time. The programmer could optimize this threshold by setting it lower or higher.

In the videos you often see that a car on autopilot handles many situations very well, but then approaches a difficult situation, like a complex street crossing or unusual obstacles in the street. Why not hand over to the human in such situations, even if the predicted risk is still low in absolute terms? Why take a risk? The machine could hand over if the perceived risk is 10% or 1% or whatever.

These difficult situations are relatively rare, at least in the US with its straightforward street geography and wide streets. In a German city center with medieval street structures it's a bit different.

Anyway, I personally would not mind at all, if the autopilot occasionally handed over to me with a warning like, "Difficult situation ahead. Please take over." Of course, this is impossible for driverless cars, but many private cars won't be driverless for a while.

You have described the difference between L2 and L3 automation, albeit perhaps in a contrary way from what you intended. In an L2 system such as Autosteer on City Streets (FSD beta), the driver is always responsible for the driving task (100% accountable for the decisions made). Thus FSDb is designed to perform steering and acceleration based on the surrounding environment and circumstances to the best of its ability, knowing that the driver is responsible and will intervene in a dangerous situation or an incorrect decision. Of course, as in all things, this is not absolute, and FSDb will "panic" and turn control over to the driver in some situations (e.g. "Do Not Block the Box" intersections).

In an L3 system, however, the autonomous system is responsible for the driving task and will fallback to the driver when it determines it cannot safely navigate a situation. Accordingly, the threshold for handoff to the driver in an L3 system would be much lower than such a threshold in an L2 system like FSDb. I would expect that if FSDb ever evolves to an L3 system, that the initial threshold will be very low, and will be raised slowly as the system improves over the years and decades that follow.