Waymo

[willow_hiller]

Some data on Waymo and other autonomous systems briefly mentioned along with the NHTSA's release of data around ADAS systems:

NHTSA says there have been 130 reported automated driving system crashes from June 2021 to May 15. Waymo, at 62, had the most. It was followed by Transdev Alternative Services at 34, and Cruise at 23 (excluding 16 crashes reported separately by GM).

Along with saying Tesla had the most ADAS accidents, it's a data-crime to give the numerator and skip the denominator (miles driven) for either statistic. But still interesting.

U.S. safety agency says Tesla accounts for most driver-assist crashes, but warns data lacks context

Federal safety officials stressed that the data is not meant to portray whether one advanced driver-assist system is safer than another.

www.cnbc.com

 

[diplomat33]

It's a data-crime to give the numerator and skip the denominator (miles driven) for either statistic.

Yes, I agree. You also need to mention type of ADS, ODD, type of accident etc... They all have an impact on what the data means. It is really meaningless to just throw around numbers like 62 accidents or 273 accidents without any context.

 

[MP3Mike]

Some data on Waymo and other autonomous systems briefly mentioned along with the NHTSA's release of data around ADAS systems:

So Waymo is averaging ~1.5 crashes per week. That seems like a lot for the number of vehicles they have on the road. But as you said the data released is woefully inadequate. No details on actual number of vehicles or miles driven.

 

[Daniel in SD]

So Waymo is averaging ~1.5 crashes per week. That seems like a lot for the number of vehicles they have on the road. But as you said the data released is woefully inadequate. No details on actual number of vehicles or miles driven

A year ago Waymo claimed to be driving 100,000 miles a week in San Francisco. They also report even very minor collisions.
It will be interesting to read the report. Hard to imagine how they’ll be able to reach any conclusions. I’m just curious what the normal rate of hitting emergency vehicles is.

 

[EVNow]

Yes, I agree. You also need to mention type of ADS, ODD, type of accident etc... They all have an impact on what the data means. It is really meaningless to just throw around numbers like 62 accidents or 273 accidents without any context.

What is surprising is NHTSA seems to have not asked for those details ? The report says only the companies have them …

 

[diplomat33]

https://twitter.com/x/status/1537515875179413511

Here are Waymo's presentations this year:

• On June 19, at the LatinX in CV (LXCV) Research workshop, Xinchen Yan will hold a tutorial session on synthetic camera data generation for autonomous driving.
• On June 20, at the Workshop on Autonomous Driving, the Waymo Research team will share the results from this year’s Waymo Open Dataset Challenges and unveil some exciting new additions to this industry leading dataset.
• On June 22, a joint team from Waymo and Google Research, including Vincent Casser, Xinchen Yan, Sabeek Pradhan, and Henrik Kretzschmar, will present BlockNeRF, a novel method for large-scale scene reconstruction based on camera images. Earlier this year, the team used this technique to recreate an entire neighborhood of San Francisco from 2.8 million images—the largest such NeRF-based 3D reconstruction, to date. Waymo also recently released the Waymo Block-NeRF Dataset, one of the datasets used in the experimental evaluation presented in the paper, so other researchers can apply their scene reconstruction methods to the dataset, too.
• On June 24, Waymo research team members will present their work in a poster session on a novel data-driven range image compression algorithm, nicknamed RIDDLE (Range Image Deep Delta Encoding). As lidar become more powerful with increasing resolutions, there is a need for lidar data compression which can lower the costs of data storage and transmission. The proposed method demonstrates a large improvement in the compression quality compared to previous algorithms.

Also, Waymo has a new website with their recent cutting edge research papers:

Autonomous Vehicle Research - Our Latest Publications - Waymo

Explore our autonomous vehicle research and Waymo Open Dataset. Our self-driving car research supports our innovative autonomous technology solutions.

waymo.com

cc @Bladerskb @Bitdepth

 

[diplomat33]

If we look at Waymo's research papers just for this year, I think we get some insight into the kinds of autonomous driving problems that Waymo is working on.

TL;DR Waymo's research seems to be focused on perception, prediction and simulation.

• In Perception, Waymo is looking at ways of compressing high res lidar data, improving camera vision and making ML training & labeling more efficient.
• In Prediction, Waymo is focused on predicting motion of entire scenes and large number of road agents which is key for handling dense urban driving.
• In Simulation, Waymo is focused on building more realistic, large scale simulations with realistic road agents.

Here are some more details that I could find from looking at the papers:

Perception

• Compressing data from high res lidar to make storage and transmission better.
• Improving per object depth estimation and tracking of monocular vision
• Using reference objects to improve long range distance estimation
• Doing multi-class 3D object detection with single-class supervision to make labeling cheaper.
• Using gradient based sample weighing to improve long tailed training data.
• Improving semantic segmentation

Prediction

• Improving scene-centric motion predictions.
• NN to predict trajectories and occupancy of hundreds of road agents with reliable accuracy and latency.
• End-to-end deep model for long term trajectory prediction
• Using occupancy flow fields to predict motion of multiple agents
• Developing a unified architecture for predicting multiple agent trajectories.
• Using deep leaning to fuse map and sensor data to predict behavior.

Simulation

• Using ML to generate realistic large scale simulation of a city from images.
• Learning realistic and diverse agents for autonomous driving simulation.
• Creating synthetic sensor data for simulations

General Machine Learning

• Using polynomial expansions to improve loss functions.
• Optimizing deep multitask models

cc: @Bladerskb @Bitdepth

 

[diplomat33]

Driverless ride in the 5th Gen I-Pace:

 

[willow_hiller]

Driverless ride in the 5th Gen I-Pace

Do we know if Gregory works for Waymo or is a regular customer?

It's interesting reading the internal press releases, but I prefer hearing feedback from public persons to make an assessment of how a service is actually doing.

 

[rbt123]

Do we know if Gregory works for Waymo or is a regular customer?

He's the Operations Planning Manager at Waymo (based on a Google search of "Gregory Puckett waymo").

 

[diplomat33]

Do we know if Gregory works for Waymo or is a regular customer?

It's interesting reading the internal press releases, but I prefer hearing feedback from public persons to make an assessment of how a service is actually doing.

Gregory works for Waymo but I am not sure why that makes any difference. This is not an internal press release, it is an unedited driverless ride on the same service that the public will use. Gregory is merely passenger in the backseat, he is not influencing the ride in any way. The Waymo car does not care who the passenger is. The passenger being a Waymo employee or not, does not change the performance of the Waymo vehicle.

 

[willow_hiller]

The passenger being a Waymo employee or not, does not change the performance of the Waymo vehicle.

It doesn't change the performance of the vehicle for that particular ride, but it does change whether or not the user decides to post it. We know it's very easy to cherry-pick good rides into a false sense of good performance.

Honestly, I find it a bit disenguous he's posting rides on YouTube without clearly identifying he works for Waymo.

 

[Supcom]

Gregory works for Waymo but I am not sure why that makes any difference. This is not an internal press release, it is an unedited driverless ride on the same service that the public will use. Gregory is merely passenger in the backseat, he is not influencing the ride in any way. The Waymo car does not care who the passenger is. The passenger being a Waymo employee or not, does not change the performance of the Waymo vehicle.

It makes a huge difference because, as an employee, he would not likely post the video had there been issues with the ride. But, a non-affiliated person would be more likely to post.

 

[diplomat33]

It doesn't change the performance of the vehicle for that particular ride, but it does change whether or not the user decides to post it. We know it's very easy to cherry-pick good rides into a false sense of good performance.

Honestly, I find it a bit disenguous he's posting rides on YouTube without clearly identifying he works for Waymo.

Yes, I agree that it is disingenuous for him not to clearly identify that he works for Waymo.

It makes a huge difference because, as an employee, he would not likely post the video had there been issues with the ride. But, a non-affiliated person would be more likely to post.

True.

And I agree that non-affiliated videos can show more of the good and the bad. But I still think that employee videos have some value since they do show the real world performance on a given route. Obviously, one video will not show you the entire performance of Waymo. But that is not the intent though. IMO, the real purpose of Gregory's video is just to show that driverless testing in central Phoenix is really happening. The fact is that even videos that show fails will not show you the entire performance since the sample is too small. All unedited videos, whether from employees or not, will only show you a small sample of the total performance.

 

[diplomat33]

FYI, Gregory (Waymo employee) does have more clips of driverless rides in the 5th Gen.

Going around cement truck:

Driving in SF:

 

[diplomat33]

Sounds like Waymo is in trouble with their $200k sensor suite.

(Moving to the Waymo thread so as not to take the FSD beta thread off topic)

The Waymo sensor suite does not cost $200k. That is a gross exaggeration. Krafcik mentioned a few years back that the total cost of the entire vehicle "costs no more than a moderately equipped Mercedes S-Class". Here is the entire quote dated January 2021:

"If we equip a Chrysler Pacifica Van or a Jaguar I-Pace with our sensors and computers, it costs no more than a moderately equipped Mercedes S-Class. So for the entire package, including the car - today"

Let's do some quick math. A Mercedes S-class starts at around $110k. A fully equipped S-class is around $130k. So split the difference to estimate a "moderately equipped S-class" and you get around $120k. A Jaguar I-Pace starts at around $70k. Take the difference and you get $50k. So the 5th Gen sensor suite probably costs around $50k. Heck, the roof lidar is "only" $7,500. There is no way the rest of the hardware would cost $192,500. Yes, $50k is still a lot more than the Tesla's sensor suite but it is not $200k. And Waymo's sensor cost will continue to go down over time. I think the Waymo Zeekr driverless robotaxi, in development now, will likely to be a lot cheaper since they can design it from the beginning for driverless. They can cut out a lof of the stuff that a luxury consumer car would have, that a driverless car would not need.

Also, it is encouraging that Tesla's FSD beta can do more drives with zero interventions but that is still a far cry from reliable driverless robotaxis. IMO, Waymo would only be in trouble if Tesla is able to remove driver supervision or deploy driverless ride-hailing in more cities than Waymo. But that has not happened yet. Waymo has made great progress towards deploying driverless ride-hailing in light and medium urban environments. Waymo is working hard on "solving" dense urban driving. As long as Waymo can bring down the cost of the hardware even more and get driverless that is good enough for most US cities, they should be fine IMO. They appear to be on track to do that.

Also, Waymo uses 19 cameras on the 4th Gen, increased to 29 cameras on the 5th Gen. So Waymo actually uses a lot more camera vision than Tesla. And Waymo increased the number of cameras on their vehicles, showing an even greater focus on camera vision. And Waymo only uses 5 lidar. So Waymo actually emphasizes camera vision way more than lidar.

So this idea that Waymo focused too much on expensive lidar and HD maps instead of camera vision and now Tesla's vision-only can basically do the same thing for a fraction of the cost, is false on many levels.

 

[diplomat33]

JB Hunt and Waymo add Wayfair to their autonomous trucking testing:

The latest pilot will span six-plus weeks during July and August and take place along the I-45 corridor between Houston and Dallas, the location of J.B. Hunt and Waymo’s original pilot nearly one year ago. It will be the first in-depth transportation of home furnishings retail freight between J.B. Hunt and Waymo Via (the company’s autonomous Class 8 trucking unit powered by the Waymo Driver™ technology).

https://twitter.com/x/status/1541825741838553089

 

[diplomat33]

Waymo co-CEOs took a driverless ride with a couple new Waymo employees. The video is mostly PR fluff, with the employees telling the CEOs how cool Waymo is. We don't see a lot of the actual drive. But the reason I am sharing the video is because at the 50 second mark, we do see one interesting driving case where the Waymo approaches a school zone and has to maneuver around cones with incoming traffic. I don't think we've seen this scenario before by a Waymo. The Waymo handles it really well IMO, slowly down, letting the incoming car pass and then going around the cones, without stopping.

Waymo on Instagram: "Waymo is home to innovative engineers, creative thinkers, and focused problem-solvers—all of whom are dedicated to improving the world’s access to mobility. Today, we’re launching Ride of Your Life, a new series in which we’ll i

300 likes, 7 comments - waymoJune 29, 2022 on : "Waymo is home to innovative engineers, creative thinkers, and focused problem-solvers—all of whom are dedicated to improving the world�..."

www.instagram.com

 

[ThomasD]

Is Wayfair going to use autonomous trucks to transport goods between warehouses? Maybe even customer deliveries. Would they travel from state to state?

 

[diplomat33]

Is Wayfair going to use autonomous trucks to transport goods between warehouses? Maybe even customer deliveries. Would they travel from state to state?

Eventually yes but right now they are just testing the autonomous trucks up and down I-45.