Waymo

[bkp_duke]

Do you have a link for this? The city is clearly not happy with being the testing ground for the two companies, and they're critical of expansions, but I can't find anything about an attempt at a ban. At least, not one that has caught the attention of the web.

Two companies race to deploy robotaxis in San Francisco. The city wants them to hit the brakes.

The city’s transportation officials sent letters this week to California regulators asking them to halt or scale back the expansion plans of Cruise and Waymo.

www.nbcnews.com

That's from a quick google search. But something came up in my news feed this week about the City Council trying to put an ordinance in place to either prevent or greatly curtail them.

EDIT - found one more recently dated, with a lot more detail:

‘Incompetent’ driverless cars are wreaking havoc on San Francisco

As Cruise and Waymo wrestle for dominance, local transportation agencies worry the...

www.sfgate.com

 

[SO16]

Waymo uses ML and fleet learning to update the maps these days. Waymo doesn't operate in unmapped areas because they wouldn't probably have the safety margins at this point to remove the driver, which is 100% of their business case.

Of course it scales. 95% of the ride hailing market is in dense urban centers incl airport.

Due to vision processing? Can you elaborate what you mean by that? Given that Tesla is at less than 95% completed rides without a critical disengagement and a DE every 10-15 miles, I'd say it's needed for the foreseeable future for driverless deployment and/or L3+.

Waymo still manually updates maps. Difficult to say how well a single vehicle driving by a construction zone would accurately update the HD map. And the first time could be an issue for the car that knows nothing about it.


A little info on Tesla vision.

Tesla AI day: Image space to Vector spaceâââthe real breakthrough

The media (both Tesla fans and haters), analysts, and content creators are focused on the Tesla bot and DOJO that were revealed on Teslaâ¦

medium.com

Again, Tesla disengagements are primarily not due to hardware/vision limitations. They are primarily from incorrect routing decisions.

I think this will be a case where Waymo works better overall for quite some time due to extra hand holding with HD maps and sensors, until it’s leapfrogged. When/if that happens, remains to be seen. Exciting times regardless.

 

[spacecoin]

Waymo still manually updates maps. Difficult to say how well a single vehicle driving by a construction zone would accurately update the HD map. And the first time could be an issue for the car that knows nothing about it.

Listen to this - for five minutes from the bookmark, and you're learn something about Waymo's approach and why they feel Lidar and maps are needed/useful.

A little info on Tesla vision.

Tesla AI day: Image space to Vector spaceâââthe real breakthrough

The media (both Tesla fans and haters), analysts, and content creators are focused on the Tesla bot and DOJO that were revealed on Teslaâ¦

medium.com

What do think is unique to Tesla here? Voxels or occupancy networks isn't a Tesla "invention" you know. It's research from 2018-2020.

GitHub - autonomousvision/convolutional_occupancy_networks: [ECCV'20] Convolutional Occupancy Networks

[ECCV'20] Convolutional Occupancy Networks. Contribute to autonomousvision/convolutional_occupancy_networks development by creating an account on GitHub.

github.com

Research — LET-3D-AP: Longitudinal Error Tolerant 3D Average Precision for Camera-Only 3D Detection — Waymo

The 3D Average Precision (3D AP) relies on the intersection over union between predictions and ground truth objects. However, camera-only detectors have limited depth accuracy, which may cause otherwise reasonable predictions that suffer from such longitudinal localization errors to be treated...

waymo.com

Waymo has this since long ago and and more and better cameras and 360 coverage and more onboard compute.

I think this will be a case where Waymo works better overall for quite some time due to extra hand holding with HD maps and sensors, until it’s leapfrogged. When/if that happens, remains to be seen. Exciting times regardless.

I find that very unlikely. Waymo has all the bells and whistles in their vision stack that Tesla has in their vision stack, and more.

 

[SO16]

Listen to this - for five minutes, and you're learn something about Waymo's approach and why they feel Lidar and maps are needed/useful.

What do think is unique to Tesla here? Voxels or occupancy networks isn't a Tesla "invention" you know. It's research from 2018-2020.

GitHub - autonomousvision/convolutional_occupancy_networks: [ECCV'20] Convolutional Occupancy Networks

[ECCV'20] Convolutional Occupancy Networks. Contribute to autonomousvision/convolutional_occupancy_networks development by creating an account on GitHub.

github.com

Research — LET-3D-AP: Longitudinal Error Tolerant 3D Average Precision for Camera-Only 3D Detection — Waymo

The 3D Average Precision (3D AP) relies on the intersection over union between predictions and ground truth objects. However, camera-only detectors have limited depth accuracy, which may cause otherwise reasonable predictions that suffer from such longitudinal localization errors to be treated...

waymo.com

Waymo has this since long ago and and more and better cameras and 360 coverage and more onboard compute.


I find that very unlikely. Waymo has all the bells and whistles in their vision stack that Tesla has in their vision stack, and more.

Interesting.

Well we shall see.

 

[bkp_duke]

Listen to this - for five minutes from the bookmark, and you're learn something about Waymo's approach and why they feel Lidar and maps are needed/useful.

What do think is unique to Tesla here? Voxels or occupancy networks isn't a Tesla "invention" you know. It's research from 2018-2020.

GitHub - autonomousvision/convolutional_occupancy_networks: [ECCV'20] Convolutional Occupancy Networks

[ECCV'20] Convolutional Occupancy Networks. Contribute to autonomousvision/convolutional_occupancy_networks development by creating an account on GitHub.

github.com

Research — LET-3D-AP: Longitudinal Error Tolerant 3D Average Precision for Camera-Only 3D Detection — Waymo

The 3D Average Precision (3D AP) relies on the intersection over union between predictions and ground truth objects. However, camera-only detectors have limited depth accuracy, which may cause otherwise reasonable predictions that suffer from such longitudinal localization errors to be treated...

waymo.com

Waymo has this since long ago and and more and better cameras and 360 coverage and more onboard compute.


I find that very unlikely. Waymo has all the bells and whistles in their vision stack that Tesla has in their vision stack, and more.

You asked the question, so I'm going to put forth what I believe are Tesla's 2 largest competitive advantages.

1) Data set. Tesla's dataset is at minimum 2 orders of magnitude larger, and they are spending money to properly label the dataset on which the NN's are trained. Proper data labelling removes the "garbage in, garbage out" scenario.
2) Computational power. Dojo aside, Tesla is crunching data much faster than anyone else, simply because they have invested in the computational power (I'm referring to supercomputer size - not in-car) more than anyone else. This allows them to iterate faster (even with a larger dataset).

That's what makes Tesla's approach unique - SCALE.

 

[Bladerskb]

You asked the question, so I'm going to put forth what I believe are Tesla's 2 largest competitive advantages.

1) Data set. Tesla's dataset is at minimum 2 orders of magnitude larger, and they are spending money to properly label the dataset on which the NN's are trained. Proper data labelling removes the "garbage in, garbage out" scenario.
2) Computational power. Dojo aside, Tesla is crunching data much faster than anyone else, simply because they have invested in the computational power (I'm referring to supercomputer size - not in-car) more than anyone else. This allows them to iterate faster (even with a larger dataset).

That's what makes Tesla's approach unique - SCALE.

 

[EVNow]

In MS we used to have Big Data conferences. I remember years ago a speaker (Tim O'Reilly) talking about how Tesla is likely to crack FSD before all other because of big data advantage. Only when someone cracks FSD will we know whether he was right or not.

BTW, why are you not apologizing for lying about me ?

 

[JB47394]

Two companies race to deploy robotaxis in San Francisco. The city wants them to hit the brakes.

The city’s transportation officials sent letters this week to California regulators asking them to halt or scale back the expansion plans of Cruise and Waymo.

www.nbcnews.com

That's from a quick google search. But something came up in my news feed this week about the City Council trying to put an ordinance in place to either prevent or greatly curtail them.

EDIT - found one more recently dated, with a lot more detail:

‘Incompetent’ driverless cars are wreaking havoc on San Francisco

As Cruise and Waymo wrestle for dominance, local transportation agencies worry the...

www.sfgate.com

I was specifically looking for something about San Francisco passing a ban. Everything I found, and the two links you provided (thank you), are about San Francisco being concerned. They clearly don't want the two services to go hog wild in their city, and are calling for caution and an emphasis on safety. I thought that you had found something that said that San Francisco was trying to eject robotaxis entirely from the city.

 

[Bitdepth]

You asked the question, so I'm going to put forth what I believe are Tesla's 2 largest competitive advantages.

1) Data set. Tesla's dataset is at minimum 2 orders of magnitude larger,

Now put a number to 2 orders of magnitude larger. Let's use Mobileye as they have actually given a number to how large their data set is.

Mobileye has 200PB (petabyte) of Data, equivalent to 16 million 1-minute driving clips spanning 25 years of driving data.

2 orders of magnitude larger would be 100 times that = 200PB x 100 = 20,000PB or 20EB (exabyte)

Entirety of google was estimated to have 10 - 15EB.

Do you see how improbable your statement sounds now?

and they are spending money to properly label the dataset on which the NN's are trained. Proper data labelling removes the "garbage in, garbage out" scenario.

Everyone is spending money to properly label data on which the NN's are trained.

2) Computational power. Dojo aside, Tesla is crunching data much faster than anyone else, simply because they have invested in the computational power (I'm referring to supercomputer size - not in-car) more than anyone else. This allows them to iterate faster (even with a larger dataset).

That's what makes Tesla's approach unique - SCALE.

Tesla has invested in computational power more than Intel, Google, Nvidia. Sometimes I'm not sure if we live in an alternate reality.

 

[EVNow]

Mobileye has 200PB (petabyte) of Data, equivalent to 16 million 1-minute driving clips spanning 25 years of driving data.

MobilEye data is quite different - usually a single camera with zero ability to use triggers to get the needed data / to filter out duplicates. Not really comparable.

 

[spacecoin]

You asked the question, so I'm going to put forth what I believe are Tesla's 2 largest competitive advantages.

1) Data set. Tesla's dataset is at minimum 2 orders of magnitude larger, and they are spending money to properly label the dataset on which the NN's are trained. Proper data labelling removes the "garbage in, garbage out" scenario.

Data set sizes are overrated. What you need is a well curated relevant and representative data set and not 2M copies of the same drive. Sim is great for creating rare scenarios.

2) Computational power. Dojo aside, Tesla is crunching data much faster than anyone else, simply because they have invested in the computational power (I'm referring to supercomputer size - not in-car) more than anyone else. This allows them to iterate faster (even with a larger dataset).

Are you claiming that Tesla has more compute than Alphabet now? Have you heard of the research breakthroughs from Google DeepMind? Most recently AlphaDev today. How do you think they're building and training those models?

System architecture | Cloud TPU | Google Cloud

cloud.google.com

Re Dojo (is it even used yet for anything relevant? Does it exist in any meaningful way? Does it scale? Do they have any software for it? All we know is that Tesla keeps upgrading the NVidia GPU clusters every 6 months)

https://www.hpcwire.com/2022/08/16/tesla-gooses-its-gpu-powered-ai-super-is-dojo-next/

I don't think the word "Dojo" was even in the last two Shareholder Decks. (q4 and q1) and Elon's been less and less confident about it for every months that goes by...

 

[Bitdepth]

Intense fog waymo is driving in.

 

[EVNow]

San Francisco Officials Make Last-Ditch Effort to Block Robotaxi Deployment

State commissioners will decide whether to allow 24/7 Cruise and Waymo service this month.

sfstandard.com

“Waymo driverless AVs have committed numerous violations that would preclude any teenager from getting a California Driver’s License,” San Francisco officials wrote. They listed intrusions by Waymo vehicles into crime scenes and construction zones, driverless cars blocking emergency vehicles and instances of vehicles parked at the curb being trapped by a shut down Waymo.​

...​

Using available data, San Francisco officials said they’ve come to the conclusion that automated driving is not yet safer than human driving and that an unlimited approval “abrogates Commission responsibility to protect public safety.”​

​

@Bladerskb @diplomat33

 

[EVNow]

Intense fog waymo is driving in.

I think they need to figure out how to pass a simple CA driving test, first.

 

[Dewg]

It's a tough balancing act. There's been a time in the evolution of everything that works when it didn't work. On one side, we need to protect the public. On the other side, the technology can only evolve so far in a simulation, and at some point needs to be tested on real roads with real people in real scenarios.

If we stopped the space program when a rocket exploded on the launch pad, or a shuttle exploded on take-off, we'd be nowhere near where we are today with technology.

We have to look to the future, and the potential benefit to society. A little pain today could result in a paradigm shift on transportation for the next generation. That sentiment is in direct opposition to today's "this better not impact me negatively in any way" mentality.

 

[diplomat33]

San Francisco Officials Make Last-Ditch Effort to Block Robotaxi Deployment

State commissioners will decide whether to allow 24/7 Cruise and Waymo service this month.

sfstandard.com

“Waymo driverless AVs have committed numerous violations that would preclude any teenager from getting a California Driver’s License,” San Francisco officials wrote. They listed intrusions by Waymo vehicles into crime scenes and construction zones, driverless cars blocking emergency vehicles and instances of vehicles parked at the curb being trapped by a shut down Waymo.​

...​

Using available data, San Francisco officials said they’ve come to the conclusion that automated driving is not yet safer than human driving and that an unlimited approval “abrogates Commission responsibility to protect public safety.”​

​

@Bladerskb @diplomat33

The data is real. The question is how do you interpret it. If you look at the data, it shows that the incidents they mention are pretty rare, only a handful in 1M driverless miles. In fact, none those violations caused an accident. So they were not safety issues. SF officials are redefining minor violations or inconveniences as safety issues to try to argue Waymo is unsafe. If you look at Waymo's 1M miles driverless safety report, Waymo showed good safety. But SF officials are choosing to interpret the data to mean that Waymo is unsafe because they have a clear anti-Waymo agenda. In fact, they have been trying to shut down Waymo for months now to no avail. This is just another desperate attempt.

 

[Bitdepth]

It's a tough balancing act. There's been a time in the evolution of everything that works when it didn't work. On one side, we need to protect the public. On the other side, the technology can only evolve so far in a simulation, and at some point needs to be tested on real roads with real people in real scenarios.

If we stopped the space program when a rocket exploded on the launch pad, or a shuttle exploded on take-off, we'd be nowhere near where we are today with technology.

We have to look to the future, and the potential benefit to society. A little pain today could result in a paradigm shift on transportation for the next generation. That sentiment is in direct opposition to today's "this better not impact me negatively in any way" mentality.

While i agree with the general sentiment you are making, it is a tricky balancing act. You can't release a "beta" software out in public, especially one controlling a 2+ ton vehicle without making sure the software you are releasing won't cause injuries to people (Uber). Lots of safety mechanisms need to be put in place to ensure you are doing your tests in the real world safely.

I agree, we must allow the technology to be used in the real world in order for it to get better because the benefits at least long term outweigh the current minor inconvenience it poses now like blocking traffic.

 

[diplomat33]

I think they need to figure out how to pass a simple CA driving test, first.

I'd like to see FSD beta pass a simple CA driving test first. It would fail within a mile. LOL.

The fact is that Waymo would easily pass a CA driving test. SF Officials are just biased against Waymo. Naturally you believe them without question because it confirms your anti-Waymo bias.

 

[EVNow]

I'd like to see FSD beta pass a simple CA driving test first. It would fail within a mile. LOL.

Yes - hilarious, considering how Tesla has driverless permit.

Unpaid (one hopes) Waymo PR people don't want to face reality.

 

[diplomat33]

Yes - hilarious, considering how Tesla has driverless permit.

No they do not! They have a CA DMV testing permit with a safety driver. The permit does not allow them to do driverless.

Here is the list of everyone that has a driverless permit. As you can see, Tesla is not on the list.

Permit Holders (Driverless Testing)
As of November 19, 2021, DMV has issued Autonomous Vehicle Driverless Testing Permits to the following entities:

• APOLLO AUTONOMOUS DRIVING USA LLC
• AUTOX TECHNOLOGIES INC
• CRUISE LLC
• NURO, INC
• WAYMO LLC
• WERIDE CORP
• ZOOX, INC

Source: Autonomous Vehicle Testing Permit Holders - California DMV