Waymo

[diplomat33]

Isn’t it easier to change the business model to adopt to the current tech ?

Sure you could do that. But if the tech has not yet achieved the safety Waymo wants for driverless at scale then what business model do you think would work? IMO, there would only be two options: Waymo scales ride-hailing but with safety drivers. That would be very costly since Waymo would have to pay all the safety drivers. And the bigger the fleet, statistically, the chance of accidents, even with safety drivers, goes up. A serious accident with the public riding in the car would be very bad PR. Or you only deploy ride-hailing in limited geofenced areas where you know the driverless is good enough but that limits your customers. So IMO, it makes the most sense for Waymo to "solve the tech" first before scaling. If they can get the FSD safe enough in a large enough ODD, then they can scale driverless and reach enough customers to turn a profit. And remember that Waymo does not have to solve FSD everywhere. They only need to solve FSD for a big enough ODD to reach enough customers, enough to turn a profit. I think that is why they switched their focus to NYC. If they can do driverless in dense urban driving, it opens up a big enough ODD to make a profit. Waymo would still geofence the ride-hailing but they could widen the geofence to denser urban areas with more customers.

Furthermore, Waymo's philosophy is that the driver should be removed from the equation. They don't trust FSD with driver supervision because of the experiment Google did way back in 2012 with their "hands off" highway driving system where the drivers were caught on camera slacking off. That's the whole reason Waymo is so focused on L4 in the first place.

 

[EVNow]

But if the tech has not yet achieved the safety Waymo wants for driverless at scale then what business model do you think would work?

Drive from suburbs to the airport.

Forget NYC, they aren't even driving in actual profitable areas of SF. Their geofence is outside the densest Uber drives in SF.

 

[diplomat33]

Drive from suburbs to the airport.

Sure, but the Waymo Driver is capable of a lot more than just doing simple airport runs.

Forget NYC, they aren't even driving in actual profitable areas of SF. Their geofence is outside the densest Uber drives in SF.

Once Waymo has "solved" dense urban driving, they can expand the SF geofence to include the more denser places in SF. They can expand the Chandler geofence to include more denser parts of Phoenix. They can deploy in other cities, etc...That's the whole point of testing in NYC and training the Waymo Driver to be better at dense urban driving.

 

[EVNow]

Sure, but the Waymo Driver is capable of a lot more than just doing simple airport runs.

Are you sure ? They just don't do freeways anymore.

Airport into city hotels may not be all the simple, either.

Once Waymo has "solved" dense urban driving, they can expand the SF geofence to include the more denser places in SF.

I think this idea of "solving dense urban driving" is illusory. It is not a binary ... just a question of error rates and may be known unhandled scenarios they are yet to take on.

I don't think they will soon start operating in Boston (let alone Bombay) after the "solve dense urban" driving in SF/NYC.

 

[diplomat33]

I think this idea of "solving dense urban driving" is illusory. It is not a binary ... just a question of error rates and may be known unhandled scenarios they are yet to take on.

I don't think they will soon start operating in Boston (let alone Bombay) after the "solve dense urban" driving in SF/NYC.

That is why I put solving dense urban driving in quotes. I know it is never 100% solved. But I do think that once Waymo achieves safe driverless in dense urban driving like NYC, that experience will help them in other cities. Waymo will still map and test first in any new city before they deploy but it will reduce the deployment time.

 

[EVNow]

That is why I put solving dense urban driving in quotes. I know it is never 100% solved. But I do think that once Waymo achieves safe driverless in dense urban driving like NYC, that experience will help them in other cities. Waymo will still map and test first in any new city before they deploy but it will reduce the deployment time.

Cruise has been saying for a long time they chose SF because it was difficult. They would be right if they can expand all over SF and other cities faster than Waymo.

It is an eternal conflict - do you start with something simple or do you start with something more representative.

 

[EVNow]

Funny thing is the “dense urban” parts of SanFrancisco that Waymo is not yet ready to cover - WholeMars drives with not that many issues

 

[diplomat33]

Funny thing is the “dense urban” parts of SanFrancisco that Waymo is not yet ready to cover - WholeMars drives with not that many issues

Dense urban driving is more than just driving straight in busy traffic or driving on a tight street with parked cars on the side. Waymo can already do that with no issues. But driving with "not that many issues" is not good enough for driverless. To deploy driverless, you have to be handle every situation in dense urban driving with ZERO issues. That is why Waymo does not cover "dense urban driving" in SF yet. Waymo can already handle dense urban driving with not a lot of issues but they need to handle it better before they can deploy driverless. As Waymo explained in their tweet to me, they are working to "improve our Drivers ability to perceive and predict the actions of other road users in dense urban areas". That is because there are a lot of actions of other road users in dense urban driving, you need to be able to predict and respond to. Omar's FSD Beta might have handled that one drive with "not that many issues", but you need to test other routes, other situations where maybe a cyclist does something new or another car does something new. There are a lot of cases you need to train for!

 

[diplomat33]

Cruise has been saying for a long time they chose SF because it was difficult. They would be right if they can expand all over SF and other cities faster than Waymo.

Yes but so far I am not seeing evidence that Cruise can scale faster than Waymo. Cruise's driverless ODD is smaller than Waymo's. Cruise has excellent driverless in their ODD, I am not taking away from their accomplishments, but focusing on the harder parts first and then deploying driverless in a smaller ODD, does not really suggest better scaling. The test will be if Cruise can quickly expand their driverless to daytime driving and to more denser parts of SF.

 

[Doggydogworld]

..... not that many issues

The FSD theme song!

 

[EVNow]

Waymo can already handle dense urban driving with not a lot of issues

So, you are saying they can match Tesla in "dense urban" traffic

Just like in Chandler, even though they lost the "race" ...

 

[diplomat33]

So, you are saying they can match Tesla in "dense urban" traffic

Yes and do a lot better than Tesla.

 

[EVNow]

Yes and do a lot better than Tesla.

Show me ...

 

[ThomasD]

It seems like these AVs will always be geofenced to a certain area. Cruise origin will most likely operate on fixed routes. I just don't see Waymo taking people farther than a few miles. These AVs will operate in areas where car ownership is low. I doubt you will ever be able to get a Waymo to pick you up at your house and take you 40 miles to the Airport or pick you up from the Airport and take you home

 

[EVNow]

Cruise origin will most likely operate on fixed routes.

In the old days we'd call them .... bus.

 

[Daniel in SD]

It seems like these AVs will always be geofenced to a certain area. Cruise origin will most likely operate on fixed routes. I just don't see Waymo taking people farther than a few miles. These AVs will operate in areas where car ownership is low. I doubt you will ever be able to get a Waymo to pick you up at your house and take you 40 miles to the Airport or pick you up from the Airport and take you home

Why can’t the geofence be 40 miles wide?
Cruise is currently offering driverless service in SF that isn’t fixed routes so why would it be different in the Origin?

 

[diplomat33]

It seems like these AVs will always be geofenced to a certain area. Cruise origin will most likely operate on fixed routes. I just don't see Waymo taking people farther than a few miles. These AVs will operate in areas where car ownership is low. I doubt you will ever be able to get a Waymo to pick you up at your house and take you 40 miles to the Airport or pick you up from the Airport and take you home

Waymo and Cruise can already do thousands of miles with no interventions. Waymo could do 40 miles to the airport and back now if they wanted to. And Cruise already does rides with no fixed routes. So it is silly to think that they will never do more than rides for a few miles on fixed routes. The current geofence areas are only "small" to make testing easier. Waymo and Cruise only deploy robotaxis with no driver when they are confident it is fully safe but also as smooth and comfortable as possible. So as they improve their autonomous driving even more, once they have the data that their robotaxis are not just 99.99999% safe but also very smooth and comfortable, I think we can expect them to expand the geofenced areas. IMO, to think they will stay in "small" geofenced areas forever is silly.

 

[willow_hiller]

The current geofence areas are only "small" to make testing easier.

We often assume that Waymo handles all roads within their known geofenced areas, but unfortunately I think it's also likely that they have geofenced roads and road-features within their geofenced areas.

That recent lawsuit against the California DMV claimed that their trade secrets included: "constraints there are on the car’s ability to traverse San Francisco’s tunnels, tight curves and steep hills."

And then if you recall, that Waymo employee who took over for one of JJRick's rides said "I don't know why this construction zone wasn't taken off the map."

So I think you're right that Waymo is prioritizing comfort and safety. But they might be prioritizing those over innovation and pushing boundaries. The fact that Waymo has such a low disengagement rate isn't evidence that their system is more advanced than others. It's evidence that they're not challenging themselves enough with difficult road situations.

 

[diplomat33]

We often assume that Waymo handles all roads within their known geofenced areas, but unfortunately I think it's also likely that they have geofenced roads and road-features within their geofenced areas.

That recent lawsuit against the California DMV claimed that their trade secrets included: "constraints there are on the car’s ability to traverse San Francisco’s tunnels, tight curves and steep hills."

And then if you recall, that Waymo employee who took over for one of JJRick's rides said "I don't know why this construction zone wasn't taken off the map."

So I think you're right that Waymo is prioritizing comfort and safety. But they might be prioritizing those over innovation and pushing boundaries. The fact that Waymo has such a low disengagement rate isn't evidence that their system is more advanced than others. It's evidence that they're not challenging themselves enough with difficult road situations.

I think we need to make a distinction between internal testing/development and public rides.

Development
Development deals with solving FSD problems and making the autonomous driving more capable. This is done with a lot of software development, research, simulation and of course a lot of real world testing with safety drivers. You need to push the boundaries in development in order to make progress. So, I am sure Waymo pushes the boundaries, within reason, when they are doing internal testing with safety drivers. They would not be doing good development if they just timidly tested in safe and easy cases all the time. But we don't see this testing because Waymo does not release videos of their internal testing.

So, I personally think that the disengagement rate is low because the Waymo Driver driving is that good, not because Waymo is gaming the disengagements with easy routes. The fact is that gaming the disengagement rate with lots of easy routes does not help you advance the autonomous driving. So it does not make sense to do that. Also, remember that Waymo runs simulations to analyze all disengagements. So Waymo can push the boundaries and have the safety drivers disengage before any problem and then just analyze what would have happened later. So there is no reason for Waymo to avoid more difficult situations when they are doing internal testing.

Public Rides
Public ride-hailing is very different than internal testing. Waymo is cautious when they do public ride-hailing for obvious safety reasons. When a mom and her kids are riding in your robotaxi to go to the store, it is not the time to push the boundaries of your autonomous driving. You want the mom and her kids to reach their destination safely and comfortably. So I am not surprised that videos from JJ Ricks and others show Waymo avoiding certain unprotected left turns, highways or other "difficult" areas etc... When doing public rides, it makes sense to "play it safe" and take the easy, safe route. We only see videos from the public riding which is why people often think the public rides represent the cutting edge of Waymo Driver. But public rides only happen after many months of testing. So public rides probably lag behind the cutting edge development software by several months. And the public rides we see in Chandler are on the 4th Gen cars. So it is also likely lagging behind the 5th Gen Waymo. That is one reason why I am looking forward to seeing videos from public rides in SF so that we can see the difference with the Chandler rides.

 

[willow_hiller]

So, I personally think that the disengagement rate is low because the Waymo Driver driving is that good, not because Waymo is gaming the disengagements with easy routes. The fact is that gaming the disengagement rate with lots of easy routes does not help you advance the autonomous driving.

Just quickly downloading and analyzing the California DMV disengagement reports for 2021 here: Disengagement Reports - California DMV

I don't think we can draw any conclusions in either direction. But I put some effort into this so I might as well share the graphs. Ultimately the data just isn't detailed enough to really dig into it very far.

First, miles per disengagement are indeed increasing as total miles are driven.



And the image is somewhat more interesting when you break it down by VIN. On average, Waymo reported 47,964 miles per disengagement for November 2021 on average, but the worst vehicle in the fleet drove 440 miles before its first disengagement that month, while the best vehicle in the fleet drove 3,344 miles without a single disengagement that month. (Points at the top of the chart represent Inf miles per disengagement).



It's a pretty extreme variance in disengagements per vehicle. So they could have a steady fleet of consumer-facing vehicles which they keep to a limited number of streets to keep public disengagements down, and a more disengagement-heavy testing fleet. Here's how that non-disengagement fleet has grown over the year:



And for the most part, miles driven is positively correlated with disengagements. Although if you plot all the VIN-months out in boxplots, it recognizes all of those 0 disengagement/high-mileage months as outliers:



And then I thought it would be interesting to break disengagements per month down into whether the disengagement was initiated by a safety-driver or by the Waymo-driver. And note, in this data, Waymo did not report a single disengagement from a driverless vehicle. So in all of these cases a safety-driver was behind the wheel:



That spike in system-initiated disengagements from April to June is a little odd. Would be interesting to see if that correlates with any expansion in the geofence, but for now it seems unexplained why it rose so dramatically and then dropped back, when driver-based disengagements stayed relatively steady.

And then lastly I'll just point out how lacking the data is on road type. They don't break miles down into Street/Freeway/Highway, but they do break disengagements down by that, so this is the best view we have of that slice:



Might be tempting to say the April-June disengagement spike was caused by Waymo testing more freeways and highways, but we can't really tell because we don't have miles per road type, just disengagements.