Autonomous Car Progress

[spacecoin]

Your 40 years trajectory is laughable considering how things have trended last 5-10 years. I'm not arguing the current perception solution is definitely good enough to allow better than human driving performance, but I don't see any big roadblocks stopping it from being that within a few years. Heck even 4x the resolution andand compute power in HW4 and resultant 4x in depth map resolution will help.

Everyone seem to think that driving without human input is a recent development. Clearly the field has advanced a lot in the last 5-10 years, but getting to "holy cow it drives by itself" to where Waymo is took 10 years with all the crutches.

Given the advances in the field, perhaps someone can do it in half the time from where Tesla is now, but I doubt it. Vision-only robotaxi will likely never happen as sensors keep evolving and falling in prices like Moores law. Only very stupid teams would insist on camera only, as you will get a "free" order of magnitude reliability improvement by adding more modalities.

If you've paid attention to the FSD videos, perception is almost never the issue. This was actually surprising even to me.

Well, it's a lot of things, but the perception isn't the problem right now for getting to L2 door to door. I'd argue that reliability is the main problem with only 90% of the drives without a critical disengagement and a disengagement every 12 miles.

This is where Tesla's perception is right now:

https://twitter.com/x/status/1669546988025819136

Oh yeah, I could care less about HW3. Better camera resolution & placement and higher compute power will make HW4.0 superior for FSD within a year or 2.

HW4 has the same camera placement. Some lenses are different.

These are all true, and can easily mask that the software still needs to be tweaked in that timeframe to work well in those cities.
I'm sure that when Waymo starts driverless rides in LA, it will span a wide region of LA, right? Not a small section, right?

Again, scaling up the operation is the hard part, and validation. I expect it to be a faster process than in SF. Waymo's target is to 10x from the number of active vehicles in the next 12 months.

 

[powertoold]

I wouldn't waste my brain cells

 

[diplomat33]

Safety validation is all we are talking about in the end. It's the performance in the real-world. In Waymo's case, when they begin safety validation in a new city they are doing so presumably without much training data in that city.

I believe you are incorrect. Waymo does a lot of driving and data collection prior to safety validation. For example, they did months of driving around NYC just to collect data. They did months of autonomous driving in Bellevue, WA to collect data in rain. They also did lots of autonomous driving in LA before they even announced that they were going to launch robotaxis in LA. And now they are back in Orlando to do more testing in heavy rain. That is proof that Waymo collects a lot of data before validation. If we look back at SF or LA or Phoenix, Waymo did a lot of autonomous driving testing before they officially started the safety validation process. So they do not start safety validation with almost no data in that city. When they do safety validation, they are already able to do reliable autonomous driving in that city.

Is your assertion that Waymo does not make any changes to their software over the course of this safety validation phase? That they do not use the data collected to modify their software?

No, that is not my assertion. Of course they make changes. If during the safety validation process, they uncover issues that need to be fixed, they will obviously make changes to the software to address the issues. And Waymo is always improving their autonomous driving, just like everybody does. Waymo is always doing research, developing better NN and deploying those NN in OTA updates to improve their autonomous driving. My point is that the main purpose of the validation process is to prove safety before opening up the robotaxi service to the general public, not to teach the system how to drive. But certainly, they will make changes to the software to improve safety/performance based on the data they collect.

 

[stopcrazypp]

If you've paid attention to the FSD videos, perception is almost never the issue. This was actually surprising even to me.

That surprised me too in viewing a lot of FSD videos. People here in the forums were all screaming about how horrible the perception is, but as it developed, it seems most of the issues are not perception related at all. Rather it's stupid things that affect even L4 cars packed to the gills with sensors.

For example, presuming it was true that pileup that happened on the Bay Bridge was with FSD Beta active, the cause was the car stopping in the middle of the road for no reason. This is something that apparently still happens regularly for other L4 cars (as per plenty of news stories posted in this thread)!

 

[ZeApelido]

Well, it's a lot of things, but the perception isn't the problem right now for getting to L2 door to door. I'd argue that reliability is the main problem with only 90% of the drives without a critical disengagement and a disengagement every 12 miles.

This is where Tesla's perception is right now:

https://twitter.com/x/status/1669546988025819136

HW4 has the same camera placement. Some lenses are different.

I definitely am not a proponent of "only cameras bro!". Sensor fusion theory is well known, of course adding sensors with somewhat decorrelated information will improve accuracy if modeled well. They are already adding back in radar, who knows maybe one day lidar. I'm just not worried about this, as there's plenty of time to add sensors if they wanted. And plenty of perception engineers to hire if they need to since many AV startups have or will have gone under - long term I have no concerns about perception!

I saw that video - another case unfortunately we don't have the UI recorded - you would be surprised how many of these sorts of cases show the perception tracking the object and it's the dumb planner doing the wrong thing.

HW4 isn't the same per se - the repeater cameras are in the same place, but their FOV changes by allowing more lateral views. This fundamentally changes their utility as they will now be allowed to used for lateral cross-traffic detection at a much forward viewpoint than the B-pillar.

I believe you are incorrect. Waymo does a lot of driving and data collection prior to safety validation. For example, they did months of driving around NYC just to collect data. They did months of autonomous driving in Bellevue, WA to collect data in rain. They also did lots of autonomous driving in LA before they even announced that they were going to launch robotaxis in LA. And now they are back in Orlando to do more testing in heavy rain. That is proof that Waymo collects a lot of data before validation. If we look back at SF or LA or Phoenix, Waymo did a lot of autonomous driving testing before they officially started the safety validation process. So they do not start safety validation with almost no data in that city. When they do safety validation, they are already able to do reliable autonomous driving in that city.



No, that is not my assertion. Of course they make changes. If during the safety validation process, they uncover issues that need to be fixed, they will obviously make changes to the software to address the issues. And Waymo is always improving their autonomous driving, just like everybody does. Waymo is always doing research, developing better NN and deploying those NN in OTA updates to improve their autonomous driving. My point is that the main purpose of the validation process is to prove safety before opening up the robotaxi service to the general public, not to teach the system how to drive. But certainly, they will make changes to the software to improve safety/performance based on the data they collect.

Thanks for this information. However I read this as supporting my point - Waymo is collecting a bunch of data in each city before they deploy. And likely modifying their models using this data.

That means their software was not already robust enough to handle those cities off the bat.

There is nothing wrong with this process per se - it's just not rapidly scalable.

 

[Bladerskb]

Thanks for this information. However I read this as supporting my point - Waymo is collecting a bunch of data in each city before they deploy. And likely modifying their models using this data.

That means their software was not already robust enough to handle those cities off the bat.

There is nothing wrong with this process per se - it's just not rapidly scalable.

So Tesla is scalable because they can go ~100 miles (being generous) between safety disengagement in a random city.
But Waymo isn't scalable because when they go to a random city they are aiming for 100k+ miles between safety disengagement?

That makes absolutely zero sense.

 

[diplomat33]

Thanks for this information. However I read this as supporting my point - Waymo is collecting a bunch of data in each city before they deploy. And likely modifying their models using this data.

That means their software was not already robust enough to handle those cities off the bat.

There is nothing wrong with this process per se - it's just not rapidly scalable.

You are misinterpreting my words. Of course Waymo collects data and modifies their models. Everybody does that. It's called developing autonomous driving. Is Tesla stopping collecting data? Of course not! But Waymo does not specialize their software for each city. Waymo uses the same software in every city. And their software is robust to handle new cities off the bat. They did that in LA. I showed you the tweet that proves it. But being robust does not mean that there is not room for further improvement. Nobody has AVs that are done and don't need any more improvement. Of course, they will improve the software further.

 

[Doggydogworld]

Looks like BMW beat Mercedes to approval of higher L3 speeds in Germany (130km/h is approximately 80mph):

BMW 5 Series is the First Car with German Approval for 130 KM/H Highway Assist

The new 2024 BMW 5 Series is the first car to get Germanys approval for its semi-autonomous motorway assist.

www.bmwblog.com

This is still Level 2. It just uses eye-tracking instead of steering wheel touch to monitor driver attention.

 

[stopcrazypp]

This is still Level 2. It just uses eye-tracking instead of steering wheel touch to monitor driver attention.

I presumed that it was L3 given the speed limit is the same 130km/h that was recently pushed out, but you are right that this one is still L2. I didn't realize Germany didn't have a hands-free system yet at these speeds.

 

[ZeApelido]

So Tesla is scalable because they can go ~100 miles (being generous) between safety disengagement in a random city.
But Waymo isn't scalable because when they go to a random city they are aiming for 100k+ miles between safety disengagement?

That makes absolutely zero sense.

You are misinterpreting my words. Of course Waymo collects data and modifies their models. Everybody does that. It's called developing autonomous driving. Is Tesla stopping collecting data? Of course not! But Waymo does not specialize their software for each city. Waymo uses the same software in every city. And their software is robust to handle new cities off the bat. They did that in LA. I showed you the tweet that proves it. But being robust does not mean that there is not room for further improvement. Nobody has AVs that are done and don't need any more improvement. Of course, they will improve the software further.

What proof do you have that it is robust right off the bat, because a Waymo employee said so? Come on.

They did "what" in LA? Deploy robotaxis right off the bat? I don't think so. No one implied its going to work terribly right off the bat, just not at the same high reliability as it is in SF / Chandler right off the bat.

Actions speak louder than words, and Waymo's actions mean they need to collect data in every city to update their model. Meaning they will do the same collection in most cities they open up to for at least several years. There is nothing wrong with this, it just means the global robust model is a work in progress. And not scalable "quickly". But you could certainly argue Waymo cannot scale quickly because of other issues, so maybe it is irrelevant to them.

Does anyone think Waymo will be working / operating in all suburbs and middle density cities across the U.S in the next 5 years? Not a chance.

The point isn't that Tesla is ahead from a pure technology standpoint, the point is that if Tesla's goal is to be able to operate basically anywhere, they have to build software that is very robust geographically, meaning they have to try to find the right algorithms and architectures sooner rather than later. This means generally worse performance in a selected subset of areas (e.g. SF) than if they had focused on getting maps / planning to work well in one area at a time.

Blader is right to say Tesla isn't scalable at 100 miles / disengagement. But we all know most of those important disengagements are due to the very fact they don't understand how to navigate all these intersections across the U.S. yet robustly. And they are actively working on solutions to that problem. Solving these problems using a robust data set means that once they are solved, they are solved for basically everywhere. That's scalable.

Perception is not the main issue. It's navigation / planning and a bit of prediction. Tesla has plenty of data to make advancements that are state of the art. It's an engineering and probably inference compute issue.

I am bullish in the medium term because I think dominance will come from the ability to work basically anywhere, and Waymo / Cruise won't even be competitors in that domain. They just simply won't be in enough places even in 3 years, it will basically be small contributions to probably 30-50 downtown / urban areas.

 

[spacecoin]

I am bullish in the medium term because I think dominance will come from the ability to work basically anywhere, and Waymo / Cruise won't even be competitors in that domain. They just simply won't be in enough places even in 3 years, it will basically be small contributions to probably 30-50 downtown / urban areas.

If Tesla is driverless (L4) in three years on existing cars in a meaningful ODD I’ll eat my shorts. For ride hailing 95% of the money is in dense urban centers. In 10-20 years we might have driverless privately owned cars. Highway L3 in what’s going to be happening before anything else 2025-2028.

Waymo is a software company first - and not a ride hailing company. They will sell the Waymo Driver suite to operators to scale out eventually.

 

[hooty]

Tesla don't have or have yet applied for a licence to operate autonomous vehicles in a public environment.
That's all you need to know

 

[nativewolf]

If Tesla is driverless (L4) in three years on existing cars in a meaningful ODD I’ll eat my shorts. For ride hailing 95% of the money is in dense urban centers. In 10-20 years we might have driverless privately owned cars. Highway L3 in what’s going to be happening before anything else 2025-2028.

Waymo is a software company first - and not a ride hailing company. They will sell the Waymo Driver suite to operators to scale out eventually.

Further to your point the real money is in 50 global metropolis. Uber is a money losing pig everywhere else and overall is a money losing pig. Ride sharing wont be the cash cow that some predict. Especially not some whacky lady with a big fund.

I can see software sales being huge so wealthy can have a chauffeur.

 

[willow_hiller]

Tesla don't have or have yet applied for a licence to operate autonomous vehicles in a public environment.
That's all you need to know

I see you're from Australia. I don't know how centralized your government is, but in the US a lot of policy like that is set at the state level. Many states don't require licenses or any reporting whatsoever for autonomous vehicle operation. And we know for a fact that Tesla is employing a fleet of professional FSD test drivers in those states and around the world.

For e.g. Tesla FSD on its way to Europe as Test Operators hiring spree begins

 

[nativewolf]

I see you're from Australia. I don't know how centralized your government is, but in the US a lot of policy like that is set at the state level. Many states don't require licenses or any reporting whatsoever for autonomous vehicle operation. And we know for a fact that Tesla is employing a fleet of professional FSD test drivers in those states and around the world.

For e.g. Tesla FSD on its way to Europe as Test Operators hiring spree begins

CA does though and as it is the state with...just guessing...over half of all Teslas. And is the most studied...and is the basis for much national legislations...CA is a state they cannot ignore. Many states don't require anything but the willingness to assume liability. That will last up until the first fatality.

Companies like testing in Phoenix- streets are all new and very uniform...suburb like setting. Waymo and Cruise geofence it even then. However, Phoenix will never be a big money maker. Money is in the big metros- where Uber does well.

 

[spacecoin]

Further to your point the real money is in 50 global metropolis. Uber is a money losing pig everywhere else and overall is a money losing pig. Ride sharing wont be the cash cow that some predict. Especially not some whacky lady with a big fund.

I assume you mean the stock pumping lady that delivered negative returns over the last 5 years despite Nasdaq being up 100%…

I can see software sales being huge so wealthy can have a chauffeur.

there is real money to be made in long haul trucking.

 

[nativewolf]

I assume you mean the stock pumping lady that delivered negative returns over the last 5 years despite Nasdaq being up 100%…

there is real money to be made in long haul trucking.

Maybe in point to point trucking. Aurora is making good progress but they have done so by removing all the complexity.

Yes that lady

 

[Bladerskb]

What proof do you have that it is robust right off the bat, because a Waymo employee said so? Come on.

They did "what" in LA? Deploy robotaxis right off the bat? I don't think so. No one implied its going to work terribly right off the bat, just not at the same high reliability as it is in SF / Chandler right off the bat.

But that reliability off the bat is higher than Tesla FSD...Which is the entire point.
Yet people say that Waymo isn't scalable but Tesla FSD with the only barometer being you can activate it.

Actions speak louder than words, and Waymo's actions mean they need to collect data in every city to update their model. Meaning they will do the same collection in most cities they open up to for at least several years. There is nothing wrong with this, it just means the global robust model is a work in progress. And not scalable "quickly". But you could certainly argue Waymo cannot scale quickly because of other issues, so maybe it is irrelevant to them.

Atleast several years? They went from full-scale testing in LA in October 19th 2022 to Driverless at night in March 7th.
That's 4.5 months. Then they went driverless in daytime (24/7) around May/June.

So thats around 6-7 months to go 24/7 Driverless in a ~50sq mile area after full scale testing.

Does anyone think Waymo will be working / operating in all suburbs and middle density cities across the U.S in the next 5 years? Not a chance.

The problem is that you are not evaluating Waymo logically.
If you did, you will clearly see the trajectory of improvement.
How long it took Waymo after they started doing full-scale testing in SF to go 24/7 driverless.
How long it took Waymo after they started doing full-scale testing in LA to go 24/7 driverless.
A logical person will compare and see the difference in time.

Then the question becomes, how low can they get the time between testing to 24/7 driverless?
Then also what timeline delta is low enough for the company to be able to attempt to deploy at multiple cities simultaneously?

Once you answer these two questions, you can logically come to a conclusion for 5 years outlook.

For example the trajectory points to I believe 3 months by end of the year / Early 2024.
And I believe with it down to 3 months, you can then simultanously do full scale testing in 3+ cities.
If you project out 3 cities every quarter from 2024 for 5 years you get 60 cities.

Trajectory of improvement is also happening in its abilities aka ODD.
It went from just handling surburb, light rain and zero construction to being able to handle environments that include city, urban, heavy rain, heavy fog and constructions. All while being driverless.

I don't see you ever mention this.

The point isn't that Tesla is ahead from a pure technology standpoint, the point is that if Tesla's goal is to be able to operate basically anywhere, they have to build software that is very robust geographically, meaning they have to try to find the right algorithms and architectures sooner rather than later. This means generally worse performance in a selected subset of areas (e.g. SF) than if they had focused on getting maps / planning to work well in one area at a time.

If that's the case and logic then Huawei is ahead of Waymo...
You can't pick and choose how to apply that logic.

Blader is right to say Tesla isn't scalable at 100 miles / disengagement. But we all know most of those important disengagements are due to the very fact they don't understand how to navigate all these intersections across the U.S. yet robustly. And they are actively working on solutions to that problem. Solving these problems using a robust data set means that once they are solved, they are solved for basically everywhere. That's scalable.

I am bullish in the medium term because I think dominance will come from the ability to work basically anywhere, and Waymo / Cruise won't even be competitors in that domain. They just simply won't be in enough places even in 3 years, it will basically be small contributions to probably 30-50 downtown / urban areas.

Unless Tesla FSD is driverless everywhere in the USA. Then it doesn't have the ability to "work basically anywhere".
This myth is the same thing that Tesla fans always hold on to, that some day Tesla FSD will magically be L5.
So no matter how good Waymo is, they will claim Tesla is better due to some non-existent magic.

 

[powertoold]

It's funny how many people here pretend to understand the details of how Waymo rolls out a city to the public. The only thing we know so far is that it takes years. Everything else is speculation.

Waymo is secretive about almost everything they do. My interpretation is that it's a money sink. Others may think it's some kind of competitive advantage, but obviously their approach is straightforward and copied by pretty much every other AV company.

 

[spacecoin]

It's funny how many people here pretend to understand the details of how Waymo rolls out a city to the public. The only thing we know so far is that it takes years. Everything else is speculation.

What Waymo is doing now isn't a massive scale-out. They are still developing the product.

Waymo is secretive about almost everything they do.

They publish many research papers and talk at CVRP every year. Can you name a single AV company that is less secretive? Tesla has published zero research afaik.

My interpretation is that it's a money sink. Others may think it's some kind of competitive advantage, but obviously their approach is straightforward and copied by pretty much every other AV company.

Does it matter if it's a money sink right now? Every VC startup is a money sink pre-revenue. How is this relevant? The investors clearly believe in Waymo still. In the same way that SpaceX investors keep pouring money into that money sink?