Waymo

[diplomat33]

What are Waymo's plans for New York?

AFAIK, Waymo has not said anything about any plans for NYC. So we don't know. They could have no plans for NYC or they could just be hiding their plans for now. We know they tested in NYC awhile back and then stopped. They put out a statement that they collected good data. So their testing at the time was likely just data collection and not any ride-hailing plans. Waymo tends to only announce their plans when they are ready to actually start early deployments.

 

[enemji]

AFAIK, Waymo has not said anything about any plans for NYC. So we don't know. They could have no plans for NYC or they could just be hiding their plans for now. We know they tested in NYC awhile back and then stopped. They put out a statement that they collected good data. So their testing at the time was likely just data collection and not any ride-hailing plans. Waymo tends to only announce their plans when they are ready to actually start early deployments.

Probably a push back from NYC Taxi companies?

 

[diplomat33]

Probably a push back from NYC Taxi companies?

Could be. I imagine NYC Taxi companies are pretty powerful. And perhaps related, NY city government has very strict AV regulations, stricter than SF, that make it very difficult to launch a robotaxi service in NYC. Personally, I don't think Waymo should bother with NYC, at least not right now. There are plenty of more AV friendly cities that would welcome Waymo One.

 

[rbt123]

AFAIK, Waymo has not said anything about any plans for NYC. So we don't know. They could have no plans for NYC or they could just be hiding their plans for now. We know they tested in NYC awhile back and then stopped. They put out a statement that they collected good data. So their testing at the time was likely just data collection and not any ride-hailing plans. Waymo tends to only announce their plans when they are ready to actually start early deployments.

Probably snow related more than anything else. Chicago, Boston, DC, Seattle are also strong taxi markets but require mostly solving the snow problem for a commercial deployment to be successful.

There's easier opportunities, commercially, in Florida and Texas to implement first.

 

[jabloomf1230]

Probably snow related more than anything else.

NYC gets on the average 12 days with snow a year. New York City recorded 0.4 inches of snow early Wednesday morning, Feb. 1, 2023, ending a 328-day snowless streak in the city. Snow is not the reason AD companies avoid NYC. Horrendous traffic and nutsy drivers are the reasons.

 

[bkp_duke]

NYC gets on the average 12 days with snow a year. New York City recorded 0.4 inches of snow early Wednesday morning, Feb. 1, 2023, ending a 328-day snowless streak in the city. Snow is not the reason AD companies avoid NYC. Horrendous traffic and nutsy drivers are the reasons.

Yeah, NYC drivers would shooting the cars if they caused roadblocks like what SF has seen. Zero tolerance for that stuff in NYC.

 

[rbt123]

NYC gets on the average 12 days with snow a year. ...

Agreed that snow in the air is isn't frequent, though taking the fleet out of service for 12 days per year would not result in a rider-friendly service. This is a guaranteed way to get bad press.

Snow/slush, or even just snow fighting tools (salt, sand) on the ground obscuring lines is also a potential issue. Splashing it up into equipment was also an issue in older hardware (not sure about Gen 5). Unusually poor GPS reception, due to reflections off numerous skyscrapers, compounds this problem as exact positioning will lean heavily on visual data.


Vegas, Miami, Orlando, Tampa, Houston, Dallas, Austin, San Diego, Atlanta, etc. are all low-hanging fruit for a commercial service; comparable to currently serviced areas.

Horrendous traffic and nutsy drivers are the reasons.

If NYC was one of a small number of northern cities that they were avoiding then I'd probably agree with you. DC has huge commercial opportunities due to a large number of tourists and business travellers.

IMO, commercialization success requires senior development to be working on very common elements like higher-speed operations (highways: earlier detection of objects/issues from fewer data-points); and junior developers work to confirm or resolve neighbourhood specific features, such as LA's 6-way stop; and in the background you consider but don't work much on region specific items (roadway lines obscured by slush). This continues until most easy targets are at least planned for commercial service.

 

[rbt123]

Snow/slush, or even just snow fighting tools (salt, sand) on the ground obscuring lines is also a potential issue. Splashing it up into equipment was also an issue in older hardware (not sure about Gen 5). Unusually poor GPS reception, due to reflections off numerous skyscrapers, compounds this problem as exact positioning will lean heavily on visual data.

I completely missed changes to planning due to changes in traction (ice). This problem exists a bit for rain (hydroplaning) or loose gravel (do any Waymo areas have a gravel road?) but not to the same degree. Ice, particularly black-ice, isn't avoidable in colder climates. Manoeuvring and braking needs to be handled safely while on it.

Again, this issue is fixable with sufficient development time, but doesn't seem like a priority compared to supporting basics like Highways.

They may even require the addition of ground facing radar to adequately detect black-ice.

 

[diplomat33]

Waymo will share some of their latest ML research at ICCV:

https://twitter.com/x/status/1707487139309449360

https://twitter.com/x/status/1707487144384602264

 

[ThomasD]

Is the vehicle smart enough to recognize rough roads and adjust speed or change lanes on a multi lane road to prevent passengers from being bounced around?

 

[diplomat33]

The 2 Waymo research papers at ICCV are:

- "Unsupervised 3D Perception with 2D Vision-Language Distillation for Autonomous Driving"
- "MotionLM: Multi-Agent Motion Forecasting as Language Modeling".

https://twitter.com/x/status/1707494834594791647

 

[diplomat33]

Is the vehicle smart enough to recognize rough roads and adjust speed or change lanes on a multi lane road to prevent passengers from being bounced around?

Not sure if Waymo has done this. If I had to guess, I would say probably yes. But I have no proof.

 

[diplomat33]

The 2 Waymo research papers at ICCV are:

- "Unsupervised 3D Perception with 2D Vision-Language Distillation for Autonomous Driving"
- "MotionLM: Multi-Agent Motion Forecasting as Language Modeling".

https://twitter.com/x/status/1707494834594791647

I looked over the 2 papers. They are very interesting ML research. Waymo is doing a lot more with unsupervised learning and large language models.

The first paper deals with perception that can automatically draw 3D bounding boxes around objects, static and moving, and auto label them with actual words without any human labeling. In the tweet, we see some examples. So the NN drew a bounding box around the object and labeled it "bulldozer", or "stop sign" or "USPS truck" without any human labeling. The main idea is to make the perception more generalized so that you can deploy your AV in new areas and it can more reliably handle perception edge cases it has never seen before.

The second paper deals with behavior prediction. Waymo has developed a new way of using large language models to do motion prediction. They basically treat motion like words in a sentence. So they have "language tokens" like going forward, turning left, turning right etc... Just as language models can predict the next word in a sentence, MotionML can predict the next motion of multiple objects. The big advantage is that it does not require any anchor and no heuristics. And MotionML can predict how the motion of one object will affect the motion of another and vice versa. So it is a much more efficient and generalized way of doing behavior prediction.

Here is a quote from the intro that really stood to me:

"In driving scenarios, road users may be likened to participants in a constant dialogue, continuously exchanging a dynamic series of actions and reactions mirroring the fluidity of communication. Navigating this rich web of interactions requires the ability to anticipate the likely maneuvers and responses of the involved actors. Just as today’s language models can capture sophisticated distributions over conversations, can we leverage similar sequence models to forecast the behavior of road agents?"

 

[enemji]

Is the vehicle smart enough to recognize rough roads and adjust speed or change lanes on a multi lane road to prevent passengers from being bounced around?

I am seriously concerned with large potholes, dips and undulations, not to mention sinkholes. i.e. those artifacts that are not advertised by road signs.

 

[ThomasD]

If the car comes across an item that is has never identified before does it send a message to a remote operator and does that person tell the car what the item is?

 

[diplomat33]

If the car comes across an item that is has never identified before does it send a message to a remote operator and does that person tell the car what the item is?

No. Waymo has perception that can handle that on its own without needing to ask for remote assistance.

 

[ThomasD]

I am seriously concerned with large potholes, dips and undulations, not to mention sinkholes. i.e. those artifacts that are not advertised by road signs.

You mean roads like this. If the car ran into this type of road issue without any road closure signs or any type of Barriers to avoid the road. would the car self-report the road conditions so other vehicles could avoid the road?

 

[enemji]

You mean roads like this. If the car ran into this type of road issue without any road closure signs or any type of Barriers to avoid the road. would the car self-report the road conditions so other vehicles could avoid the road?

View attachment 978258

Exactly

 

[diplomat33]

Comparison of Waymo and Cruise ride-hailing service in SF:

TL;DR

Price
Waymo and Cruise were both around the same price and closely correlated to Uber and Lyft. The robotaxis were usually ~5% cheaper than their human driver alternatives. The 1.7 mile trip below was on a weekend and took about 10 minutes. It cost $11.31.

Pickup Experience
Waymo and Cruise both had decent pickup experiences. The car stays locked until you unlock it on the app. You can then enter the car and click "Begin Ride" in either the app or in-car screen. With Cruise, the rider must look for the name of the car, which is printed on the rear of the car, to determine his ride. With Waymo, they've repurposed the spinning lidar hat to show your initials when arriving.

In-Car Experience
For this category, it isn't even close. Waymo has the nicer car, seats up to four people (versus three in a Cruise), and has better visualizations of the route.

Dropoff Experience
Waymo and Cruise both struggled with the drop-off. On streets that are not crowded, they both do great, stopping immediately outside of the drop off location. On crowded streets, neither robotaxi did well. During a nighttime ride into FiDi, Cruise missed the last turn before our drop-off, so tried to do the entire 20-minute loop again to reach the original drop-off. Waymo never missed a turn, but it did struggle to figure out where to drop us off. On a separate night trip into Nob Hill, it couldn't find a "safe" spot in front of our destination. It ended up driving 0.4 miles away, passing multiple vacant parking spots, to drop us. The entire ride was only 1.1 miles, so this was a bit frustrating.

Overall
Waymo wins in the battle of the SF robotaxis for now. I'm excited to see more entrants, but currently the Waymo experience seems most polished with room to grow and improve.

Full report:

https://twitter.com/x/status/1709957606339731750

 

[Yelobird]

Granted I don’t use Wamo type service I was always curious on random opportunities. I would assume as a value you would no longer have to Tip the ride right? Or why. Second (crazy) could someone use it as a hauler? Meaning if the wife wanted to buy 30 bags of compost could she hail it at a Home Depot, load it up with a destination home and it would arrive for someone to unload? Without a passenger? Crazy but curious.