Waymo

[Terminator857]

... but I feel like Waymo will deploy for a much higher % of the public than Tesla Vision will have in 12-36 months

Agree, but Waymo still feels slow in deployment. When will they cover 90% of Bay Area or even just the peninsula? I'm still thinking in 5+ years Tesla will crack the nut and deploy widely, likely with more advanced hardware.

 

[jabloomf1230]

I guess that the NY Times got bored with attacking Musk and Tesla. Today's hatchet job focuses on self driving taxis:

‘Lost Time for No Reason’: How Driverless Taxis Are Stressing Cities https://www.nytimes.com/2023/11/20/...s-cars-cities.html?smid=nytcore-android-share

 

[Doggydogworld]

So are you disregarding the timeline in LA or do you take that into account?

• They went from full-scale testing in LA in October 19th 2022

IMHO Waymo is still a year away from a meaningful paid service in LA. I don't think they've even applied for a CPUC permit there yet.

• Downtown LA. Starting October 2023 which is ~11 months. Now this has to be around 150 sq miles give or take.

They are not serving 150 square miles in LA. They're demo'ing small subsets of that area in one-month slices.

Also what do you think will be the timeline of Austin which begun mass testing in April 2023....?

They tested extensively (not just "a couple cars") in Austin for years before dialing their effort there back about two years ago. Even with that head start I think a meaningful paid service is at least a year away.

3 years is an exaggeration, but not by much.

 

[Bladerskb]

Let me ask you. How many years will it take Waymo to cover top 100 metropolitan areas in US ?

I'm not trying to be contentious or argue, As you have noticed my activity on this forum has been very low for quite a while.
As 7 years of repeating the same thing over and over again. At this point who ever belives will continue to believe no matter what.

I'm simply interested in how you came to your conclusion and whether recent timelines in LA and Austin was taken into consideration.

In regards to your question, previously I would have said 2028. But I don't believe that anymore, while Waymo is the clear front runner in SDC and there doesn't look to be a close second. I don't believe they will be the ones to democratize it. I stopped believing back in 2022.

These comments from Waymo engineers basically sums up what I mean:

"Waymo has 2 CEOs. TWO CEOs."​

​

"Yep, it’s silly. It does align with what the company excels at most : IndecisivenessAnd what better way to express that than to not decide on a single CEO.Trickle-down indecisiveness."​

​

"If the two CEOs cannot capitalize on the progress we've made, they are not fulfilling their roles and should step down. It was an experiment that didn't succeed. Two slow snails don't combine to form a fast snail."​

I do fully believe that Commercializing a new technology while being owned by Google is literally a death sentence.

There are several recent examples you can use.

First of all, ChatGPT. Google trained almost all of the world's ML engineers. They have the most and best compute (TPU), they invented the ML Platform and library (tensorflow), they are responsible for almost all ML breakthroughs including the famous Transformers. They also have the best ML scientists.

Yet they were beaten by OpenAI with ChatGPT to the point there was a code RED meeting that brought in the founders because their entire business were threatened. They completely restructured their company to address the threat from ChatGPT (combining Google Brain with Deepmind), etc.

Then releasing Bard (which is only 1 product) which was inferior to even GPT 3.5 let alone 4. On the other hand OpenAI and MS has like 50 LLM AI products(copilot, copilot excel, copilot outlook, etc) and counting and Google has what one (Bard)?

It literally took an extinction level event for them to actually do something.

Another perfect example is the debacle of project tango.

They diddled around with project tango for years and did nothing with it. releasing just one devkit prototype. They had the best tech in the world but did nothing with it. Then out of nowhere BAM, apple showed up with ARKit (3 years later). While Google was still fiddling around and doing absolutely nothing. After the reveal of ARkit, Google in a rush scrapped the entirety of tango and created ARcore to compete. Infact most of the code for Arcore was tango but just renamed. Infact there were some codes and api that weren't renamed in the mad rush when it was first released.

Google shows off ARCore, its answer to Apple’s ARKit | TechCrunch

Now Apple is now adding lidar hardware to their cameras to get tango-que like accuracy. Basically running circles on Google. It made them ditch tango to copy what they were doing, then now gone back and add hardware tech to increase the accuracy to tango levels.

That is just one side of the Tango debacle. Kinect which is similar to tango came out in 2010 and was a hit, with a new version 2 releasing in 2013 which was a dud. But MS didn't stop there. The hardware made its way into Hololens 1 in 2015 and was used as inside out tracking. This made Hololens the first inside out tracked headset. Then it made its way into Windows Mixed Reality VR Headset in 2017 and made it the first inside out tracked VR headset and now a next gen version of it is in 2019 Hololens 2.

Tango on the other hand was mishandled resulting in Google doing nothing with it. Google glass came out and didn't use it, Google's Daydream (which is now cancelled, XD) its VR headset came out and didn't use it. Fast forward now and everyone has inside out tracking, occulus, valve for their devices. The other oems doing VR headsets are using the kinect tech (acer,etc). MS had a headstart and they surprisingly took advantage of it and created a long portfolio. Even Apple with their Apple Vision. Google beat everyone to it yet completely blew it and did absolutely ZERO. Zip. Nada. Nothing. With It.

I could keep going and list all Google failed new technological products, Stadia, blah blah.
But the point is, Waymo tech is clearly not the issue. Its the management.

I have said previously, Cruise clearly had an expansion plan.
Waymo on the other hand has no clue what they are doing, yet their tech is 3-4 years ahead of Cruise.
If Cruise had Waymo's tech, they would be in 100 cities by 2026.

Again - once Tesla "gets to FSD", will it deploy slower or fast than Waymo ?

Faster

Its mind boggling how two companies with different strategies are being compared as if they have the same strategy and challenges. Calling this comparison dumb is giving it too much credit.

Well its not a comparison. Like I said I was just curious to see how you came to your conclusion and whether recent timelines in LA and Austin was taken into consideration.

 

[diplomat33]

New podcast with Dolgov:

Episode Detail - Wonder - Feel the Story

wondery.com

A lot of the podcast is talking about the history of Waymo and the Google Self-Driving Project. But towards the end, there were a few interesting tidbits:
- Waymo has 250 driverless cars in SF and "about the same" in Phoenix.
- Waymo has done more than 5M driverless miles now.
- Waymo's business model is building a generalizable driver and deploying it on three AV applications: robotaxis, trucking, and eventually personally owned cars. Robotaxis and trucking will come first but he says Waymo is interested in fully autonomous consumer cars in the future.
- Dolgov believes that in 10 years, public will be able to summon a robotaxi in most major US cities. He says autonomous ride-hailing is a big market.
- Dolgov does not think it is unrealistic that the public will be able to buy a fully autonomous consumer car in 10 years. "a lot can happen in 10 years".
- Dolgov envisions a future where manual driving is a hobby that people do on a track but mundane, daily driving is all autonomous. He also envisions a future with smart infrastructure that could communicate with autonomous cars to make traffic smoother and safer.

 

[Terminator857]

> Dolgov does not think it is unrealistic that the public will be able to buy a fully autonomous consumer car in 10 years. "a lot can happen in 10 years".
Of course. Question is will it be Tesla , Mobileye, Waymo, etc. Which will be the least expensive option? That will likely be the winner.

 

[ItsNotAboutTheMoney]

> Dolgov does not think it is unrealistic that the public will be able to buy a fully autonomous consumer car in 10 years. "a lot can happen in 10 years".
Of course. Question is will it be Tesla , Mobileye, Waymo, etc. Which will be the least expensive option? That will likely be the winner.

Well, I still think the cheapest Robotaxi service would be the winner. There's such massive demand destruction possible for vehicle ownership once you have cheap point-to-point autonomous taxis/rental/group transit.

 

[Terminator857]

Well, I still think the cheapest Robotaxi service would be the winner. There's such massive demand destruction possible for vehicle ownership once you have cheap point-to-point autonomous taxis/rental/group transit.

In other words, we will be buying fewer cars once cheap robotaxi service is available.

 

[diplomat33]

I have said previously, Cruise clearly had an expansion plan.
Waymo on the other hand has no clue what they are doing, yet their tech is 3-4 years ahead of Cruise.
If Cruise had Waymo's tech, they would be in 100 cities by 2026.

I would definitely like to see Waymo get more aggressive with their scaling, especially now that their tech is getting really good. Why not deploy a small autonomous ride-hailing service in say 20 US cities at once? Start small but then build each one up. It could help them get more diverse data which could help speed up the generalization of the waymo driver. But it would also give them a wider presence and as they build each service up, the services would become more useful the public.

I would also like to see Waymo announce plans for an "eyes off, highway-only" system for personal cars. They seem to be making a lot of progress with validating highway driving. Why not start a partnership with a carmaker now to deploy a "eyes off" system in say 2026? Then, spend the next 3 years developing and deploying the system. Deploying the Waymo Driver as an "eyes off" highway system on personal cars would be a great application of the tech, generate revenue, as well as be a great way to democratize the tech. For example, Mobileye has announced their deal with Polestar to deploy Chauffeur by 2026.

I think that 2024 will be the deciding year for Waymo for better or worse. They need to go big in 2024. Hopefully, it will be for the better.

> Dolgov does not think it is unrealistic that the public will be able to buy a fully autonomous consumer car in 10 years. "a lot can happen in 10 years".
Of course. Question is will it be Tesla , Mobileye, Waymo, etc. Which will be the least expensive option? That will likely be the winner.

Cost is important but I don't think it is the only thing that matters. Safety, reliability, convenience, ODD etc also matter. Furthermore, when buying a car, consumers also look at things that have nothing to do with autonomy like the style of the car both inside and out, amenities, luxury, size, comfort, range, maintenance etc... It is conceivable that a company could deploy full autonomy on a personal car that does not sell a lot because consumers don't like the car for reasons other than autonomy (ex: car is ugly, lacks range, poor build quality etc). So I don't think it is as simple as the cheapest autonomy will win. Consumers will weigh a lot of factors when deciding what autonomous car to buy.

 

[Bitdepth]

New podcast with Dolgov:

Episode Detail - Wonder - Feel the Story

wondery.com

A lot of the podcast is talking about the history of Waymo and the Google Self-Driving Project. But towards the end, there were a few interesting tidbits:
- Waymo has 250 driverless cars in SF and "about the same" in Phoenix.
- Waymo has done more than 5M driverless miles now.
- Waymo's business model is building a generalizable driver and deploying it on three AV applications: robotaxis, trucking, and eventually personally owned cars. Robotaxis and trucking will come first but he says Waymo is interested in fully autonomous consumer cars in the future.
- Dolgov believes that in 10 years, public will be able to summon a robotaxi in most major US cities. He says autonomous ride-hailing is a big market.
- Dolgov does not think it is unrealistic that the public will be able to buy a fully autonomous consumer car in 10 years. "a lot can happen in 10 years".
- Dolgov envisions a future where manual driving is a hobby that people do on a track but mundane, daily driving is all autonomous. He also envisions a future with smart infrastructure that could communicate with autonomous cars to make traffic smoother and safer.

It is refreshing to have someone that is very knowledgeable in this field speak realistically about the subject matter.

 

[Doggydogworld]

In other words, we will be buying fewer cars once cheap robotaxi service is available.

Far fewer. Could be almost zero. At a 6:1 ratio 2 million new Robotaxis a year would displace 12 million new consumer cars per year. We only buy 13-14m consumer cars a year, on average, plus a couple million commercial cars/vans/trucks.

The trick is getting cheap robotaxi service. So far it's absurdly expensive.

 

[diplomat33]

Maybe after they maximize their operating territories they will determine how much, if any, excess design margin they have for possible system simplification and cost reduction.

Another interesting bit from the Dolgov podcast is that he says the 5th Gen on the I-Pace uses 29 cameras, in addition to all the lidar and radar. 29 cameras feels like overkill to me. Surely, they could reduce some of the cameras and still be safe enough for driverless. For comparison, Mobileye proposes a robotaxi configuration with 12 cameras, 6 radar and 9 lidar. I have to believe that there is a middle ground between Mobileye's low end and Waymo's high end that would still be safe enough. I do think that Waymo has excess design margin that they will be able to cut to save costs. As Waymo gains more experience, I think they will learn where they can reduce sensors in a way that still maintains the high safety they need. I am curious about the sensors on the Zeekr vehicle. The sensor suite looks slimmer than the I-Pace. It will be interesting to see if Waymo is already working on trimming the sensor package on their next vehicle to reduce cost.

 

[rbt123]

Surely, they could reduce some of the cameras and still be safe enough for driverless.

Almost certainly could in clear weather when all hardware is functional. The real question is can they remove cameras when some are already out of order due to obstruction or damage.

True reliability requires airline type efforts: a very sensitive smoke alarm to catch smokers but still having a metal ashtray in the bathroom so they can safely put out the cigarette after they're caught. The no-smoking sign alone is sufficient 99.999% of the time.

If they need 2 forward facing cameras for minimal safety, and need to account for mud/rain/snow temporarily obstructing 50% of vision (they clean but not instantly), then the minimum installed requirement might really be 4 or 5 forward facing cameras to ensure 2 are always fully functional. Safe/durable infrastructure is built for exceptional circumstances like withstanding that 2 day period when a 1 in 100 year storm rolls through, not just what will work on an average day.

Determining whether they can remove sensors requires knowing object detection accuracy during the worst weather events.

 

[Bladerskb]

IMHO Waymo is still a year away from a meaningful paid service in LA. I don't think they've even applied for a CPUC permit there yet.

Well we are discussing driverless readiness and not necessarily paid public service.
For example when will it be driverless for employees, then media/beta testers then public.
Similar to new versions of FSD Beta that goes to employees first then influencers, early beta testers then public

They are not serving 150 square miles in LA. They're demo'ing small subsets of that area in one-month slices.

They can't serve all the geofence that they can drive driverless in. Because of the limited amount of cars they have left.
But the full geofence in LA, from Santa Monica to D-LA to the other smaller locations has to be atleast 100 sq miles.

However with 250 cars being in SF and around ~250 in PHX and 600-700 Pacificas being put out of service plus they no longer produce more I-Paces.
Its quite clear they went this "Tour" route because they couldn't provide an adequate service without super long delays if they serviced the entire area they currently support at once. There simply are not enough cars. So they thought they could do a tour and get all enthusiasts in each specific location that would have signed up for their waitlist rides anyways to ride.

Now could they have relocated some cars just do a full service in Santa Monica? Yeah sure, but the wait time to get paid driverless approval is 6+ months which would probably be delayed AGAIN due to protests. So they would have ended up servicing the people in LA waitlist that only lived in Santa Monica and then they wouldn't be able to showcase that they cover multiple areas in LA.

So its also a show & tell.
But maybe I'm wrong.

They tested extensively (not just "a couple cars") in Austin for years before dialing their effort there back about two years ago. Even with that head start I think a meaningful paid service is at least a year away.

3 years is an exaggeration, but not by much.

They officially left in 2019, but were they even technically there to begin with?

The system (sensors and compute) in the 2015 car vs 2023 is literally multiple orders of magnitude different in every category. Lidars, Cameras, Radars and Compute. You can't even compare them, that means everything HAD to be remapped.

Secondly, is there even any code from 2015 that's running in the 2023 software (8 years later). I very much doubt it.
There's night and day difference in ML, capabilities and performance.

Were they handling urban areas? No.
Were they handling rain? No.
How much left turns were they doing?
Were they handling parking lots? No
etc..

Lastly what locations were they actually testing is also important. As back then they focused more on residential streets. Plus what is the size of the locations. When you look at it you end up with a comparison of a-couple sq miles to probably around 50 sq miles with their return.

"We’ve chosen Austin, Texas, as a new testing location for our project, and one of our Lexus SUVs is there now, with safety drivers aboard, driving a few square miles north and northeast of downtown Austin. It’s important for us to get experience testing our software in different driving environments, traffic patterns and road conditions—so we’re ready to take on Austin’s pedicabs, pickup trucks, and everything in between."

Google's self-driving car takes to the roads in Austin, Texas, in next phase of testing

If you live in Austin, Texas, and thought you spotted one of Google’s self-driving cars, you weren’t imagining things: the...

9to5google.com

Is quite different from...

"Over the next few months, we’ll begin testing our fifth-generation driver, on the Jaguar I-PACE platform, across central and East Austin, including downtown, Rainey Street, Clarksville, Bouldin Creek, the Market District, Holly, and the Capitol. We intend for our Austin operations to be a truly useful service from the start, traversing a large portion of the city night and day. The Waymo Driver will travel to many popular locations, like the heart of downtown, Barton Hills, Riverside, East Austin, Hyde Park and more."

Expanding Waymo’s testing to the city that keeps it weird

On October 20, 2015, Steve Mahan, who is legally blind, became the first person in the world to ride in a fully autonomous vehicle—with no human driver behind the wheel. And he did so on residential streets in Austin, in Waymo’s Firefly prototype vehicle—the first generation of our Waymo Driver.

waymo.com

So this is why I refute the idea that if Waymo was driving there 8 years ago, then it took them 8 years to launch (scale) in that city. That completely glosses over the details.

 

[rbt123]

The system (sensors and compute) in the 2015 car vs 2023 is literally multiple orders of magnitude different in every category. Lidars, Cameras, Radars and Compute. You can't even compare them, that means everything HAD to be remapped.

One minor nit. The maps are an abstraction, not a direct save of sensor data. There's no reason to believe the abstraction is specific to one sensor kit.

They would want to reconfirm the maps simply due to time passing and maps becoming partially outdated: this applies whether using the same sensors or not.

 

[EVNow]

I'm simply interested in how you came to your conclusion and whether recent timelines in LA and Austin was taken into consideration.

Because your original question was mostly meaningless.

Why ?

- Waymo has been testing in Seattle area for years already. So, they are not ready to rollout based on quality data or business.
- Google started their maps dev in Seattle (with ex MSFT maps guys !). Kirkland here is pretty much second home to Google. If Google wanted they could pretty much post LA type videos of Seattle today.

So, from the time they announce a service in Seattle to starting is mostly a PR timeline rather than engineering. The real and practical question would be when do you think Waymo will really operate in my area. Will it be before or after I retire and move to a warmer city ? Will it be before or after Tesla FSD is good enough to be called a very useful ADAS (Won’t be L3 because Tesla won’t take liability).

How about this question. When do you think Waymo will start a service in an Indian city ?

 

[kabin]

Another interesting bit from the Dolgov podcast is that he says the 5th Gen on the I-Pace uses 29 cameras, in addition to all the lidar and radar. 29 cameras feels like overkill to me. Surely, they could reduce some of the cameras and still be safe enough for driverless. For comparison, Mobileye proposes a robotaxi configuration with 12 cameras, 6 radar and 9 lidar. I have to believe that there is a middle ground between Mobileye's low end and Waymo's high end that would still be safe enough. I do think that Waymo has excess design margin that they will be able to cut to save costs. As Waymo gains more experience, I think they will learn where they can reduce sensors in a way that still maintains the high safety they need. I am curious about the sensors on the Zeekr vehicle. The sensor suite looks slimmer than the I-Pace. It will be interesting to see if Waymo is already working on trimming the sensor package on their next vehicle to reduce cost.

No doubt 29 is a lot of cameras and a load of data. But I could see a design using wide field of view cameras as well as strategically placed narrow field of view cameras where needed. Each camera type would have it's pro/con for given scenarios. More redundancy but depends heavily on vision.

 

[diplomat33]

No doubt 29 is a lot of cameras and a load of data. But I could see a design using wide field of view cameras as well as strategically placed narrow field of view cameras where needed. Each camera type would have it's pro/con for given scenarios. More redundancy but depends heavily on vision.

Correct. Waymo uses a variety of cameras in order to provide robust perception for driverless. I definitely think Waymo is right to have different cameras to have the most redundancy. But I also know people complain that Waymo's sensor suite is too expensive and wouldn't work for consumer cars. Now, I think Tesla needs more sensors to do driverless. I just wonder if Waymo could reduce sensors a bit without compromising driverless safety, still more than tesla but a bit less than current sensors, in order to reduce cost.

 

[Bladerskb]

One minor nit. The maps are an abstraction, not a direct save of sensor data. There's no reason to believe the abstraction is specific to one sensor kit.

They would want to reconfirm the maps simply due to time passing and maps becoming partially outdated: this applies whether using the same sensors or not.

Actually it is processed (dynamic objects removed) raw sensor data with abstraction on top. The richness of the abstraction has changed drastically over-time and so has the sensor quality. The more points you have the ability for you to make out more details for your map (for example a sign post), hence the richer your HD map is which leads to better accuracy and performances. So no, they wouldn't use a 2nd gen car hd map with 5th gen car hd map.

This is why Lidar HD Map are GBs in size while purely abstracted camera HD map are KBs in size.

 

[Bladerskb]

- Waymo has been testing in Seattle area for years already. So, they are not ready to rollout based on quality data or business.
- Google started their maps dev in Seattle (with ex MSFT maps guys !). Kirkland here is pretty much second home to Google.

Yes they have been testing since 2016 when they sent a single 2nd gen car there.

"One of Google's self-driving Lexus RX450h SUVs have been cruising the streets of North Kirkland over the last several weeks, creating a detailed map of the area to help future driverless cars better understand the area."

Google is testing self-driving cars in a third city: Kirkland, WA

The Pacific Northwest joins Austin and Mountain View

www.theverge.com

And the point i'm trying to make is, do you agree or disagree that having 1-5 cars in a city is different from having 50-100 cars?

They sent 5 cars to NYC.
They sent 6 cars to WA.

for example Waymo says they test in Detroit. But I worked right in the center of hart plaza (campus martius) for several years.
Guess how many Waymos I saw in Detroit? 0
Did I see Waymos in surrounding areas in the last 7-8 years? Yes, 3 times in total (Novi, Farmington Hills).
Of which I'm almost positive were the same car.

What's my point? Waymo sends a-couple cars to dozens of cities (30+) so they can collect data and build a generalized system so their dataset doesn't just consist of one city (Mountain View). This is what Tesla fans says are Tesla's claim to fame. So its weird that it doesn't count as data collection for Waymo, but somehow its seen as a failed attempt to "launch/scale".

Waymo could tomorrow announce to launch in Detroit then all of a sudden there would be dozens of Waymos (50-100) seen every day in the Motor City. Which would be up from ~zero sighting. You would presumably say that its all PR that actually it took them 8 years to launch there because they had sent a car there once. Even though there has been zero sighting of them in Detroit for years.

Which goes back to my question, do you agree or disagree that having 1-5 cars in a city is different from having 50-100 cars there?

If Google wanted they could pretty much post LA type videos of Seattle today.

What do you mean by LA type videos?

So, from the time they announce a service in Seattle to starting is mostly a PR timeline rather than engineering.

Which goes back to my question, do you agree or disagree that having 1-5 cars in a city is different from having 50-100 cars there? Or is the difference just PR as you have said?

The real and practical question would be when do you think Waymo will really operate in my area. Will it be before or after I retire and move to a warmer city ? Will it be before or after Tesla FSD is good enough to be called a very useful ADAS (Won’t be L3 because Tesla won’t take liability).

I can see Waymo offering a service there in 4 years. But i'm less interested in whether they will attempt to offer a service.
But rather if they tried, how long would it take. You seem to say they are already trying and failing even if they just sent "One Google self-driving Lexus RX450h SUV"

How about this question. When do you think Waymo will start a service in an Indian city ?

I don't think that will happen in the next 10 years. Not just for Waymo but for everyone else. Both in terms of practicality (people over there mostly use bikes and motorcycles) and also financial reasons. There's no lucrative money to be made.