Autonomous Car Progress

[Bladerskb]

So FSDbeta 10.2 has rolled-out and it's gone to people in the 100 Safety Score Button group as well. I'll be waiting in the 99 Safety Score Button group.

I can't believe Tesla beat Waymo to this point. I was completely certain Waymo would roll-out their service in Mountain View years ago, but now it's Tesla first out of the gate. If anything, I've been too conservative in my Tesla predictions and far too generous with Google/Waymo.

We've been watching FSDbeta videos for a while now and many anonymous nicks FUDsters claimed FSDbeta was NEVER going to be deployed. Well, they've been proven wrong again and NEVER was 10/11/2021. FUDsters will whine about reality being unfair, as usual, and they'll tweak the NEVER/FRAUD/SCAM narrative to some other nonsense.

 

[fcastrillo86]

Can we have a waiting room thread for us non-100 scoring folks lol

 

[rxlawdude]

Driver visualizations for GM Ultra Cruise.

Which is vapor for at least two years. Gotta love your penchant for non-Tesla corporate hype.

 

[diplomat33]

Nice highway merge by Zoox:

 

[JHCCAZ]

Which is vapor for at least two years. Gotta love your penchant for non-Tesla corporate hype.

To be fair, these various visualizations are an interesting part of autonomous driving software (UI) development and to me they belong in this thread.

For example one of the Chinese robo-taxi visualizations posted a few months ago looks pretty nice, and I have to say that the Tesla design is coming a little closer to these others - increasing contrast, steadiness and completeness of rendering objects in the immediately relevant traffic environment.

One concern is whether the visualization design is pushed towards informing the L2 driver of the car's actual perception including uncertainties, or towards a cool and confident rock-steady Birds-Eye View cartoon that gives an L4 passenger something reassuring to watch.

Recall the disconcerting shimmering and jumping of beta 8, then the steadier but "graded-shade" lines and borders rendered in beta 9, intended to represent the actual level of the NN's confidence. Now in 10.2, we're seeing a yet-steadier and more confident rendering. I like it and it's great if it represents real Improvement in the NN confidence, but not if it actually suppresses the uncertainties. That is, we don't want it to contribute to any false sense of security in the L2 human driver, a pretty picture of a clear lane ahead as the car boldly accelerates towards Road Closed barriers as in Rob Maurer's first 10.2 video.

Maybe the much-derided vibrating visualizations of 8.x were trying to tell us something important, something that was made easier to watch in 9.x but still delivering a true uncertainty picture. In an L2 display, we don't want them to hide that issue.

I would have exactly the same question about the ultra Cruise visualization as shown. Yes it looks good, but note that the presenter emphasized, even in the short clip, that this is an L2 system. So we could fairly ask the same question: is the pretty cartoon a good representation of current perceived reality, or is it filtered in a way that might inspire false confidence? To what extent are the road renderings a product of the car's own perception, or to what extent are they informed by pre-existing maps that only tell us "this is what the roads and intersections should look like"?

 

[rxlawdude]

To be fair, these various visualizations are an interesting part of autonomous driving software (UI) development and to me they belong in this thread.

For example one of the Chinese robo-taxi visualizations posted a few months ago looks pretty nice, and I have to say that the Tesla design is coming a little closer to these others - increasing contrast, steadiness and completeness of rendering objects in the immediately relevant traffic environment.

One concern is whether the visualization design is pushed towards informing the L2 driver of the car's actual perception including uncertainties, or towards a cool and confident rock-steady Birds-Eye View cartoon that gives an L4 passenger something reassuring to watch.

Recall the disconcerting shimmering and jumping of beta 8, then the steadier but "graded-shade" lines and borders rendered in beta 9, intended to represent the actual level of the NN's confidence. Now in 10.2, we're seeing a yet-steadier and more confident rendering. I like it and it's great if it represents real Improvement in the NN confidence, but not if it actually suppresses the uncertainties. That is, we don't want it to contribute to any false sense of security in the L2 human driver, a pretty picture of a clear lane ahead as the car boldly accelerates towards Road Closed barriers as in Rob Maurer's first 10.2 video

Maybe the much-derided vibrating visualizations of 8.x were trying to tell us something important, something that was made easier to watch in 9.x but still delivering a true uncertainty picture. In an L2 display, we don't want them to hide that issue.

I would have exactly the same question about the ultra Cruise visualization as shown. Yes it looks good, bulut note that the presenter emphasized, even in the short clip, that this is an L2 system. So we could fairly ask the same question: is the pretty cartoon a good representation of current perceived reality, or is it filtered in a way that might inspire false confidence? To what extent are the road renderings a product of the car's own perception, or to what extent are they informed by pre-existing maps that only tell us "this is what the roads and intersections should look like"?

My point is that at this point in time, most really don't care about other's vapor.

 

[JHCCAZ]

My point is that at this point in time, most really don't care about other's vapor.

Still I'm interested in seeing the compilation of what everyone is saying, even if it's hype. We are (mostly) adults here and we can handle the concept that many of these yet-to-be products are being edited and presented in an optimistic if not dishonest light. As you know there is no shortage tireless proponents on both sides of the Tesla vs competitors analysis and vetting of these videos.

That's why I'm happy to see the Ultra Cruise video but I present the question of how we interpret what it means.

 

[rxlawdude]

Still I'm interested in seeing the compilation of what everyone is saying, even if it's hype. We are (mostly) adults here and we can handle the concept that many of these yet-to-be products are being edited and presented in an optimistic if not dishonest light. As you know there is no shortage tireless proponents on both sides of the Tesla vs competitors analysis and vetting of these videos.

That's why I'm happy to see the Ultra Cruise video but I present the question of how we interpret what it means.

Tireless proponents. And tireless opponents.

 

[stopcrazypp]

To be fair, these various visualizations are an interesting part of autonomous driving software (UI) development and to me they belong in this thread.

For example one of the Chinese robo-taxi visualizations posted a few months ago looks pretty nice, and I have to say that the Tesla design is coming a little closer to these others - increasing contrast, steadiness and completeness of rendering objects in the immediately relevant traffic environment.

One concern is whether the visualization design is pushed towards informing the L2 driver of the car's actual perception including uncertainties, or towards a cool and confident rock-steady Birds-Eye View cartoon that gives an L4 passenger something reassuring to watch.

Recall the disconcerting shimmering and jumping of beta 8, then the steadier but "graded-shade" lines and borders rendered in beta 9, intended to represent the actual level of the NN's confidence. Now in 10.2, we're seeing a yet-steadier and more confident rendering. I like it and it's great if it represents real Improvement in the NN confidence, but not if it actually suppresses the uncertainties. That is, we don't want it to contribute to any false sense of security in the L2 human driver, a pretty picture of a clear lane ahead as the car boldly accelerates towards Road Closed barriers as in Rob Maurer's first 10.2 video.

Maybe the much-derided vibrating visualizations of 8.x were trying to tell us something important, something that was made easier to watch in 9.x but still delivering a true uncertainty picture. In an L2 display, we don't want them to hide that issue.

If you look at AI day presentation, it's pretty clear the increase in confidence of the visualization is not faked, but rather it's from multiple improvements (using multi-cam instead of single cam, multiple frames, taking into account temporary occlusions, using RNNs to keep persistent previous predictions, etc).
This summary video covers most of it.

I would have exactly the same question about the ultra Cruise visualization as shown. Yes it looks good, but note that the presenter emphasized, even in the short clip, that this is an L2 system. So we could fairly ask the same question: is the pretty cartoon a good representation of current perceived reality, or is it filtered in a way that might inspire false confidence? To what extent are the road renderings a product of the car's own perception, or to what extent are they informed by pre-existing maps that only tell us "this is what the roads and intersections should look like"?

XPeng has similarly confident visualizations (zero uncertainty) and it's that way because the road visualization is based entirely on maps (other than objects placed there like cones/cars etc which are placed in the visualization in real time). They partner with Amap which makes the "MD" maps for the system. The system can only be activated in mapped areas, so it never has to deal with an unmapped area. I imagine GM's system is similar, given they also say it can only be activated in mapped areas.

XPENG TO DEPLOY AMAP’S 3RD-GEN IN-CAR NAVIGATION SYSTEM

Tesla takes a fundamentally different approach in that while they also have "MD" maps to aid the system in identifying lanes and road sign/signal locations, the visualization and most of the car's understanding of the road structure is generated by the BEV (birds eye view) NN based on road features it recognizes from the cameras in real time. That's why even in the most confident version, it's still blurry and it makes plenty of mistakes even in roads Tesla have long mapped in detail (like in SF Bay Area).

 

[mark95476]

The competition is totally coming. No, not Super Cruise!

GM has ~2 years to render all sorts of marketing fluff and you'll be hyping it all along. I'm sure GM's marketing department will be exceptional at driver visualization renderings while Tesla is busy releasing firmware updates and adding more drivers. Lets hope in 2 years GM will have an actual product and won't have to orphan their Ultra Cruise trade mark.

Driver visualizations for GM Ultra Cruise.

 

[diplomat33]

The competition is totally coming. No, not Super Cruise!

GM has ~2 years to render all sorts of marketing fluff and you'll be hyping it all along. I'm sure GM's marketing department will be exceptional at driver visualization renderings while Tesla is busy releasing firmware updates and adding more drivers. Lets hope in 2 years GM will have an actual product and won't have to orphan their Ultra Cruise trade mark.

I am not hyping anything. I did not even comment about whether I thought the visualizations were good or bad or whether Ultra Cruise will be good or good. I am not saying that I think Ultra Cruise will be better than FSD Beta. I simply posted the video so that video can see it and decide what they think. I posted the video so that people can see outside the Tesla bubble what kind of visualizations other companies are developing for their L2 end to end.

The reason I post news and "marketing" from other companies in this thread is because this thread is all about autonomous driving progress outside of Tesla. This thread is about what other companies are doing. I simply want people to know what is happening outside the Tesla bubble.

 

[ThomasD]

Has Zoox started picking up people from the Airport and taking them to their hotels yet

 

[JHCCAZ]

Here is basically a promotional video from Lucid about their DreamDrive ADAS system. As known before, it has far more sensors than Tesla's suite, including many cameras, multiple radars and front-facing lidar (my guess would be that the lidar module is from Luminar but truly I don't know). Lucid does have some sort of autonomy engineering team including a recent hire to head that program. It's not clear to me whether they are partnering with anyone like MobilEye or others, but I feel that this hardware package is not specified by a third party. Unless I missed it there's no detail about the computer engine that would run the self-driving with all these sensor inputs.

At this stage it's quite clear that this is an extensive L2 system, with aspirations for L3 on the highway and no real prediction of L4 on any particular schedule. One thing that I do like is the use of the high-end audio system, with all its speakers, to reproduce audio cues with directional fidelity to help the L2 driver stay aware of events outside the car.

I'd certainly be interested to know more about Lucid's engineering approach and their very rough expectations about timeline to get to L3 and then L4. Considering that we don't hear a lot of specific claims about advanced autonomy, it seems that their idea is to provide in abundance of hardware before the software is figured out. This is not a cheap approach but it's not a cheap car. It's in direct contrast to Tesla's approach of providing a minimal sensor suite, a powerful computer and a plan to figure it out from there.

 

[diplomat33]

It's not clear to me whether they are partnering with anyone like MobilEye or others, but I feel that this hardware package is not specified by a third party. Unless I missed it there's no detail about the computer engine that would run the self-driving with all these sensor inputs.

When Lucid first revealed the Lucid Air, they said it would use a Mobileye chip (eyeQ4 or eyeQ5). But they are doing all the software in-house. So it could be that they are using the eyeQ4/5 chip but just doing all the software in-house, or maybe they ditched the Mobileye chip too, not sure.

 

[JHCCAZ]

When Lucid first revealed the Lucid Air, they said it would use a Mobileye chip (eyeQ4 or eyeQ5). But they are doing all the software in-house. So it could be that they are using the eyeQ4/5 chip but just doing all the software in-house, or maybe they ditched the Mobileye chip too, not sure.

That's interesting because I've had the impression that Mobileye doesn't much like their customers going their own way witn-house software. Also you might think that with this awesome sensor suite, Lucid would be in a prime position to use the most complete Mobileye ADAS which is purported to be pretty darn close to an L4 solution.

From a pure sanity-check perspective, it doesn't really seem likely that Lucid would be the one to independently solve self-driving while everyone else is struggling. Having a plethora of wonderful sensors may indeed solve some issues that Tesla faces, but we know it also introduces other difficulties and is hardly a magic key. So perhaps the "in-house software" aspect is not a ground-up competitor but really more of a platform customization, tailoring the Mobileye ADAS to the specifics of the Lucid vehicle platform and adding Lucid custom UI magic as well. Again I have no special knowledge just speculation to try to fit the circumstances.

 

[Dutchy Ron]

www.automotiveworld.com/articles/accu...

Accurate HD maps can improve ADAS comfort

 

[Dutchy Ron]

Lucid Motors is using TomTom HD Maps as well

https://corporate.tomtom.com/static-files/e38f3ce0-1b8e-431b-bc73-08e9e55f66e8

 

[diplomat33]

That's interesting because I've had the impression that Mobileye doesn't much like their customers going their own way witn-house software. Also you might think that with this awesome sensor suite, Lucid would be in a prime position to use the most complete Mobileye ADAS which is purported to be pretty darn close to an L4 solution.

Agreed. I was surprised that Mobileye decided to apparently go it along on developing their Dream Drive. I agree with you that with their sensor suite and the eyeQ5 chip, they would be in a great position to use ME's out of the box solution on what I think is a nice looking car. They could offer SuperVision which is L2 "end to end" and maybe even L4. So why reinvent the ADAS wheel so to speak? It's baffling to me. Personally, I think it is a mistake.

From a pure sanity-check perspective, it doesn't really seem likely that Lucid would be the one to independently solve self-driving while everyone else is struggling. Having a plethora of wonderful sensors may indeed solve some issues that Tesla faces, but we know it also introduces other difficulties and is hardly a magic key. So perhaps the "in-house software" aspect is not a ground-up competitor but really more of a platform customization, tailoring the Mobileye ADAS to the specifics of the Lucid vehicle platform and adding Lucid custom UI magic as well. Again I have no special knowledge just speculation to try to fit the circumstances.

Yeah, it is unlikely IMO this late in the game that a small company like Lucid will "solve FSD". But they don't seem to be really interested in that since they are only promising L3 highway. So perhaps, Lucid is satisfied with just doing L3 in house and not even trying to do L4 or L5? I am speculating that since Lucid is a small company that they don't want to take any risks with trying to "solve FSD" and are content with less ambitious goals?

 

[JHCCAZ]

...
Yeah, it is unlikely IMO this late in the game that a small company like Lucid will "solve FSD". But they don't seem to be really interested in that since they are only promising L3 highway. So perhaps, Lucid is satisfied with just doing L3 in house and not even trying to do L4 or L5? I am speculating that since Lucid is a small company that they don't want to take any risks with trying to "solve FSD" and are content with less ambitious goals?

I'm sure they would love to have L4 and they probably think the car is ready for it hardware-wise (and I would agree from the sensor aspect at least). It would bring them a ton of attention of course and they would love to be first in the consumer space.

But as much as they've been an unknown quantity, both investors and potential customers are trying to decide if they are real, or yet another failure or even another scam. The smart play is to not over-promise regarding something that everyone else is late to accomplish. I can forgive them the the lush website and somewhat dreamy videos, that's normal advertising especially for a luxury focused product. Indications are that they are at least as real as Rivian, 'and more real then a bevy of concept cars that are shown every year by existing automakers. They have a factory that they built, a first model that arrived late but intact, and a stock offering that is doing just OK. But the production ramp could be their undoing (as it nearly was for Tesla).

With the fancy hardware, their ADAS is well in the game and crrtainly not lagging the market. Lucid is starting to look real and good, and what they don't need right now is some exposé article that says they're scamming on an L4 promise that they don't even need at this time.

 

[diplomat33]

I'm sure they would love to have L4 and they probably think the car is ready for it hardware-wise (and I would agree from the sensor aspect at least). It would bring them a ton of attention of course and they would love to be first in the consumer space.

But as much as they've been an unknown quantity, both investors and potential customers are trying to decide if they are real, or yet another failure or even another scam. The smart play is to not over-promise regarding something that everyone else is late to accomplish. I can forgive them the the lush website and somewhat dreamy videos, that's normal advertising especially for a luxury focused product. Indications are that they are at least as real as Rivian, 'and more real then a bevy of concept cars that are shown every year by existing automakers. They have a factory that they built, a first model that arrived late but intact, and a stock offering that is doing just OK. But the production ramp could be their undoing (as it nearly was for Tesla).

With the fancy hardware, their ADAS is well in the game and crrtainly not lagging the market. Lucid is starting to look real and good, and what they don't need right now is some exposé article that says they're scamming on an L4 promise that they don't even need at this time.

I definitely think Lucid is real and here to stay. They have an excellent product and a solid team. Although they will probably be a luxury EV niche for awhile.

I actually like that Lucid is not "pulling an Elon" and promising the cars are L5 capable before they even deliver any features They are smart to be conservative. I think delivering safe and reliable features is more important than being on the cutting edge, especially for a new company that needs to build a reputation. So, I think step one will be to deliver a solid autosteer and TACC to show that they can deliver good basic ADAS. Then, if Lucid can deliver good features, via OTA on a steady timeline and say in 1-2 years from now, deliver reliable L3 highway that is truly hands-off and eye's off, that would be fantastic IMO.

I was just pointing that ME is a respected ADAS provider. So I am a little surprised they did not go with ME's out of the box solution which could have given them a very good ADAS with minimal effort.