Autonomous Car Progress

[diplomat33]

But according to them ...

So, after it "worked really, really well" they started working on the FSD ?

No. They already had FSD. But they keep working on it to make it better because FSD is not solved yet.

I'm sorry your skepticism doesn't extend beyond Tesla. Its very clear they are lying about something there.

I don't see any lying. They are saying that they think the FSD is great but still needs work.

During the ScaleAI conference, Jesse Levinson explained it this way, starting at the 14:30 mark in the video:

What we found is that when we go to a new part of the city or even a brand new city, as long as the topology is something that we've seen that type of thing before, even if it's otherwise different, or the scenery is a bit different, we do great. We really just, even on our first try, we can drive with no interventions. For example, the very first time, we ever tried to drive in Las Vegas, which was way back in 2019 now, our very first drive, several miles, no take-overs were necessary. Which is pretty cool right? We'd never driven in Las Vegas before and with just a map of the city, we were able to drive quite well. That does not mean that we could drive a billion times with no intervention. And there was of course work to do.

Source: Scale Exchange

So they could drive for the very first time in a newly mapped area "several miles with no take-overs" which he calls "great". So I take it that "works really well" is just a reference to being able to drive some miles with zero interventions right after mapping a new area. Basically, Jesse thinks the FSD is working really well when it can immediately start doing some drives with zero interventions right after mapping a brand new area for the very first time. But he says that it does not mean that they could drive a billion times with no interventions. He understands that driving some miles with zero interventions does not prove your FSD is actually reliable enough for full deployment. You still need to do work to make sure it is good enough before you can actually deploy it to the public.

 

[Dan D.]

Why this Amazon-owned company is bringing its autonomous vehicles to Seattle​

Why this Amazon-owned company is bringing its autonomous vehicles to Seattle

The Amazon-owned autonomous car company is expanding into Seattle. Here's a look under the hood at how this type of tech works.

www.popsci.com

Zoox, the self-driving car outfit, is expanding to Seattle and its rainy weather.

Let's hope these companies get a finished product eventually that scales to other markets. Good that they are making some kind of progress and still issuing press reports.

The "toaster" vehicles from Zoox

and Cruise

could work well in the Las Vegas Convention Center Boring tunnel. Though I'd still rather see the large capacity vehicle from Tesla that they originally planned. Are they still working on that? Would they ever give up using the Level 0 Tesla cars in favour of other companies' autonomous vehicles (or put a Level 3+ system in there)?

 

[Bladerskb]

 

[Dutchy Ron]

Tesla bulls—and Elon Musk, too—are in open revolt against Biden’s new NHTSA adviser​

Musk is in open revolt against Biden's new NHTSA advisor

Elon Musk and his supporters believe a Duke University professor who would serve as a key advisor to the National Highway Traffic Safety Administration holds extreme "bias" against the company.

fortune.com

NHTSA isn't the only one in the U.S. government with a bone to pick with Tesla. Jennifer Homendy, the new head of the National Transportation Safety Board, told the Wall Street Journal last month that the carmaker shouldn't roll out FSD beta to its broader U.S. customer base before first addressing safety deficiencies.

 

[hgmichna]

That's all well and good, but I presume the software was designed to keep the car centered by default (which is a challenge in itself already for a lot of them, many just ping-pong between lanes), and the more advanced reactions you describe may not exist at all (may lead to things like phantom braking or undesired drifts in the lane).

I cannot compare with other lane-keeping cars, because I have never experienced one, but I'm under the impression that my Tesla Model 3 SR+ at least since the last update to version 2021.32.21 moves away from trucks while overtaking them on a multi-lane freeway. (Can anybody confirm this?) It also conspicuously positions itself towards the edge of the freeway shortly before automatically leaving it to an off-ramp. So the idea that a car does not necessarily have to be in the center of the lane is already there.

If a discussion about this ensues, we should start a separate thread. But I guess we already agree that an automomous car should steer more cleverly than just sticking rigidly to the center of the lane.

 

[Knightshade]

Tesla bulls—and Elon Musk, too—are in open revolt against Biden’s new NHTSA adviser​

Musk is in open revolt against Biden's new NHTSA advisor

Elon Musk and his supporters believe a Duke University professor who would serve as a key advisor to the National Highway Traffic Safety Administration holds extreme "bias" against the company.

fortune.com

NHTSA isn't the only one in the U.S. government with a bone to pick with Tesla. Jennifer Homendy, the new head of the National Transportation Safety Board, told the Wall Street Journal last month that the carmaker shouldn't roll out FSD beta to its broader U.S. customer base before first addressing safety deficiencies.

NTSB has been heavily anti-tesla for many years, so this is nothing new. NTSB also has no regulatory authority, so much less a concern than NHTSA which does.

 

[EVNow]

I don't see any lying.

If it was already "working really, really well" - its a solved problem.

Thats the lie.

You are again wearing your rosy glasses for non-Tesla related ads masquerading as news items. Tesla skeptics need to look at other companies with the same critical eyes.

Basically, Jesse thinks the FSD is working really well when it can immediately start doing some drives with zero interventions right after mapping a brand new area for the very first time.

So, Musk can't use terms like "its great" or "drives very well" - but others can, right ?

Jesse, doesn't get to decide what "really, really well" means. We do. Unless he actually puts all those disclaimers every time he says "very, very well". Otherwise its a lie.

 

[EVNow]

In other news ...

Robotaxi fleets on roads likely 10 years away, Lucid CEO says

It is likely to take a decade before self-driving taxis will be deployed on roads, the chief executive of electric car startup Lucid Group Inc, Peter Rawlinson, said on Thursday. Tesla Inc's CEO, Elon Musk, said in 2019 that robotaxis with no human drivers would be available in some U.S...

news.yahoo.com

10 years is the new 5 years.

 

[MP3Mike]

They are not taking 2 more years just to do more mapping. They finished mapping.

If you are using HD maps you are never finished mapping...

 

[idriveacar]

If you are using HD maps you are never finished mapping...

Regardless of maps or no maps, you're never finished. New signs and road semantics will always be introduced. One off signs are a thing. Because AGI isn't yet possible, no car is going to automatically understand new signs or markings until they're programmed to do so.

Tesla ignores this entirely by being an L2 system. But the L4 robo-taxies actually have to solve these problems before launch. By mapping and localization, they've developed a robust means of identifying new road signage and semantics, and will likely launch with pipelines in place to quickly identify and label these changes.

I think Tesla's approach is appropriate for an ADAS; but when you remove the driver you have a whole set of new challenges. Robo-taxies can solve these new challenges differently and (quite likely) more effectively because they have an entirely different structure (fleet of purpose built vehicles in limited ODDs).

I don't understand this anti mapping sentiment. Tesla would be foolish to not have their cars report lane structures of roads back to HQ, effectively building their own set HD maps anyway. What's the functional difference between that and what the L4 companies already do?

Also it only takes a matter of days to map an entire city once, and then the vehicles themselves can work to keep them up to date without any additional overhead.

 

[diplomat33]

If it was already "working really, really well" - its a solved problem.

Thats the lie.

You are again wearing your rosy glasses for non-Tesla related ads masquerading as news items. Tesla skeptics need to look at other companies with the same critical eyes.

When it comes it autonomous driving, "working really well" does not equal "solved problem". That's all I am saying. Jesse says that they have not solved FSD. He is not pretending that their FSD is solved.

But Jesse probably should not use terms like "works really really well" that gives people the wrong impression that the FSD is ready when it is not.

I think I do try to hold Tesla and non-Teslas to the same FSD standard. But the reality is AV companies like Waymo and others acknowledge L4 autonomous driving and are deploying some robotaxis whereas Tesla only acknowledges L2. So they are literally not at the same level of autonomous driving.

So, Musk can't use terms like "its great" or "drives very well" - but others can, right ?


Jesse, doesn't get to decide what "really, really well" means. We do. Unless he actually puts all those disclaimers every time he says "very, very well". Otherwise its a lie.

You are reading into my words. I am not holding Elon to a different standard. I think Jesse is wrong to say that it is working really really well because it can do "several miles with no interventions" after mapping. Doing several miles with no intervention is not "working really well" in my book.

 

[diplomat33]

In other news ...

Robotaxi fleets on roads likely 10 years away, Lucid CEO says

It is likely to take a decade before self-driving taxis will be deployed on roads, the chief executive of electric car startup Lucid Group Inc, Peter Rawlinson, said on Thursday. Tesla Inc's CEO, Elon Musk, said in 2019 that robotaxis with no human drivers would be available in some U.S...

news.yahoo.com

10 years is the new 5 years.

I guess it depends on what he means by "fleets". Are we talking 1,000 robotaxis, 10,000 robotaxis, or 1M robotaxis?

But I think he is probably a bit pessimistic in his prediction. We already have a small number of robotaxis on public roads and several companies like Waymo, Cruise, Zoox, Aurora, Argo are working on expanding to several more areas. Considering how much progress has been made in autonomous driving and how many AV companies are actively working on deploying robotaxis, I think we will definitely see more and more robotaxis in the next 5 years.

 

[idriveacar]

Tesla Inc's CEO, Elon Musk, said in 2019 that robotaxis with no human drivers would be available in some U.S. markets in 2020. That proved to be inaccurate.

"Proved to be inaccurate" is certainly one choice of words (I'm far from a Tesla hater but I found that euphemism to be funny ).

Back on topic, I don't agree with Lucid's CEO's assessment. If we're talking about a system that requires no human involvement - not even remote operators - he has a point.

But the way that I see it going is robo-taxies will be deployed with limited ODDs and remote operators, who help with new signs and ambiguous situations. Over time the operator/machine interface will improve, the cars will improve, and fewer and fewer humans will need to be on staff. Eventually, the operational cost will dip below the cost of conventional taxies, and we'll see an influx of robo-taxies. I seriously doubt this is over 10 years away.

I'm always a bit saddened to hear smart people with relevant experience just use their energy to decide to which degree of "impossible" SDCs are, instead of using their energy to help overcome the limitations.

I'm not saying they're wrong about true, generalized, intelligent robots... But when you limit the scope and tackle each problem one by one, there is a solution.

 

[diplomat33]

I think we got a first glimpse of some Mobileye's SuperVision "city driving" features for the Zeekr001. From the set up of the additional screen, it looks like a test car with a professional car. And yes, I am aware that this is a marketing video from Geely. But I think it has interesting information worth sharing so I am sharing it.

Here we see Zeekr 001 perform left turn and u turn at busy intersection:

I think it shows that Zeekr001 will have "FSD Beta" type features. It looks like they are testing the city driving features for the advanced ZAD package.

Also, note that the additional screen appears to have the same FSD visualizations that we've seen on Mobileye's FSD demos.

 

[eli_]

Video is kind of bizarre, I don't understand why they have this super fake looking augmented view on the right, the graphics don't correspond to the vehicle path or steering angle or anything like that. And then there's some random bounding boxes thrown in at the very end.

 

[diplomat33]

Video is kind of bizarre, I don't understand why they have this super fake looking augmented view on the right, the graphics don't correspond to the vehicle path or steering angle or anything like that. And then there's some random bounding boxes thrown in at the very end.

Yeah, they should have just shown us what was on the screen so we could see all the real FSD visualizations from the car.

 

[Bladerskb]

Video is kind of bizarre, I don't understand why they have this super fake looking augmented view on the right, the graphics don't correspond to the vehicle path or steering angle or anything like that. And then there's some random bounding boxes thrown in at the very end.

Yeah Geely just throwing some marketing sparkles on a basically mobileye video to make it Geely branded. I don't think Geely's UI is done as we have yet to see their driving adas UI.
They definitely still have work on their end software wise.

 

[mark95476]

Road signs are regulated and relatively consistent. They have to make sense so NN training can be used to extract relevant contextual info and put it into the driving stack. That's mimicking what human drivers do while driving.

Keeping a city map up-to-date and in-sync in real-time is going to involve an immense amount of overhead. That's probably one reason why Tesla never went that route.

Regardless of maps or no maps, you're never finished. New signs and road semantics will always be introduced. One off signs are a thing. Because AGI isn't yet possible, no car is going to automatically understand new signs or markings until they're programmed to do so.

Tesla ignores this entirely by being an L2 system. But the L4 robo-taxies actually have to solve these problems before launch. By mapping and localization, they've developed a robust means of identifying new road signage and semantics, and will likely launch with pipelines in place to quickly identify and label these changes.

I think Tesla's approach is appropriate for an ADAS; but when you remove the driver you have a whole set of new challenges. Robo-taxies can solve these new challenges differently and (quite likely) more effectively because they have an entirely different structure (fleet of purpose built vehicles in limited ODDs).

I don't understand this anti mapping sentiment. Tesla would be foolish to not have their cars report lane structures of roads back to HQ, effectively building their own set HD maps anyway. What's the functional difference between that and what the L4 companies already do?

Also it only takes a matter of days to map an entire city once, and then the vehicles themselves can work to keep them up to date without any additional overhead.

 

[svusa]

so what's the deal with the new cameras for autopilot 2.0!? my vision is lately getting blurred.

 

[stopcrazypp]

Video is kind of bizarre, I don't understand why they have this super fake looking augmented view on the right, the graphics don't correspond to the vehicle path or steering angle or anything like that. And then there's some random bounding boxes thrown in at the very end.

Don't remember which brand, but some cars have a heads up augmented reality view (either as a screen or part of HUD unit), but I would hope the UI is a lot better than that video and actually lines up with the real world, or it would be kind of pointless.