10 Waymo self driving cars join Lyft

[banned-66611]

Waymo expands Phoenix self-driving service by putting 10 cars on Lyft

Waymo has added 10 cars to the Lyft service, opening them up to more users. They still have safety drivers in the front seat, hands off.

 

[diplomat33]

Waymo expands Phoenix self-driving service by putting 10 cars on Lyft

Waymo has added 10 cars to the Lyft service, opening them up to more users. They still have safety drivers in the front seat, hands off.

it's 10 cars "over the next few months" according to the article. Frankly, it's unimpressive. Sure, Waymo has excellent self-driving tech, their cars are "hands off", but their expansion rate is very slow. They are obviously not interested in deploying their self-driving cars on a large scale.

 

[azred]

We live in Chandler (just SE of Phoenix), which is probably where this is “exciting development” is mostly happening. After several years of being in Chandler with many Waymo vehicles, I assume most of the streets here are mapped. What I don’t understand is how this technology will ever be widely deployed nationally or worldwide.

Waymo supposedly has mapped about 80 square miles in metro Phoenix, which presumably includes the 65 square miles of Chandler. Metro Phoenix is over 14,000 square miles so just mapping metro Phoenix seems like it is years away. And yet many of the experts seem to believe Waymo is way ahead of Tesla.

 

[diplomat33]

What I don’t understand is how this technology will ever be widely deployed nationally or worldwide.

At 10 cars every few months, it certainly won't be. My guess is Waymo lacks the ability to mass produce the cars with the hardware so they have to buy the cars and assemble the extra hardware individually and of course, test each car individually too before public deployment, hence why the deployment is so slow. Or maybe, Waymo's tech requires that they calibrate the tech or something for each car individually. Either that or Waymo is being extra conservative because they don't want to deploy too many cars just in case there is a problem since they are still working out the software and their cars lack the ability to do OTA software updates AFAIK. Basically, they can't deploy en masse because if there was a problem, they would have to do a mass recall of all their cars. The only way I see Waymo widely deploying their tech nationally is if they partnered with a big auto maker to mass produce the cars with the hardware. But Waymo seems perfectly content to be a small urban robotaxi service.

Frankly, this illustrates the big advantage that Tesla will have IF they can get the FSD software to work. Tesla is already deploying tens of thousands of cars per quarter with the FSD hardware already integrated. Obviously, Tesla's FSD software is lagging behind right now but if Tesla can solve that part, they will be able to deploy their tech nationwide and even worldwide much faster than anyone else.

 

[acoste]

It is all just resource allocation. Their fleet is limited until next year. Small set of cars are testing customers' experiences. Small set of cars will test the cooperation with Lyft. And all the other cars are chasing edge cases for faster development.

The only way I see Waymo widely deploying their tech nationally is if they partnered with a big auto maker to mass produce the cars with the hardware.

They already have . Waymo picks Detroit factory to build self-driving cars – TechCrunch

their cars lack the ability to do OTA software updates AFAIK

They have 2 car suppliers and both models have OTA. I don't know if they are using them for their software at the moment, but wouldn't take much effort to do that modification if needed.

 

[diplomat33]

It is all just resource allocation. Their fleet is limited until next year. Small set of cars are testing customers' experiences. Small set of cars will test the cooperation with Lyft. And all the other cars are chasing edge cases for faster development.

They already have . Waymo picks Detroit factory to build self-driving cars – TechCrunch

They have 2 car suppliers and both models have OTA. I don't know if they are using them for their software at the moment, but wouldn't take much effort to do that modification if needed.

Good to know. Thanks for sharing. Let's see if they pick up the pace next year then because certainly what we are seeing so far with 10 cars rolled out over months, does not impress me.

Frankly, I don't see how Waymo can ever get to L5 without large fleet data. A small fleet + simulations is not enough to catch all the edge cases in a timely manner. I suspect those cars are going to be chasing edge cases for awhile.

 

[CarlK]

If this is what they could do after years of mapping a single city what else they could do to ever take the "safety engineer" off the car or not requiring passengers to sign NDA? If you believe what Elon, Karpathy and Levandowski said recently the answer is not much more they could do and it will never happen.

Early Rider Program – Waymo

Good to know. Thanks for sharing. Let's see if they pick up the pace next year then because certainly what we are seeing so far with 10 cars rolled out over months, does not impress me.

Frankly, I don't see how Waymo can ever get to L5 without large fleet data. A small fleet + simulations is not enough to catch all the edge cases in a timely manner. I suspect those cars are going to be chasing edge cases for awhile.

I don't think these few cars can even be close of getting edge cases. The team just has to put up something to show the management there is progress that's all.

 

[electronblue]

Frankly, this illustrates the big advantage that Tesla will have IF they can get the FSD software to work. Tesla is already deploying tens of thousands of cars per quarter with the FSD hardware already integrated. Obviously, Tesla's FSD software is lagging behind right now but if Tesla can solve that part, they will be able to deploy their tech nationwide and even worldwide much faster than anyone else.

Except unless that current hardware suite turns out not sufficient for Level 5 no geofence or even meeting Waymo at Level 4, then all that benefit would be lost - and indeed become a major liabilty issue instead.

 

[Saghost]

Except unless that current hardware suite turns out not sufficient for Level 5 no geofence or even meeting Waymo at Level 4, then all that benefit would be lost - and indeed become a major liabilty issue instead.

Disagree. Even if it turns out they need more sensors to make FSD a reality, the cars they are building are watching human drivers deal with the edge cases and sending data back to Tesla. They're running in shadow mode and turning up discrepancies between what the NN would do and what the humans did - all without a single mile of Autopilot, let alone FSD. Under Autopilot, they're still learning more, capturing every disengagement and sorting out the important ones and finding more edge cases the system missed there.

That's a benefit the other companies just don't have, and one that won't go away if Tesla decides they need LIDAR (not gonna happen.) If they do add sensors, they won't immediately have the big fleet advantage for training those sensors, but they'll still know a lot more about the edge cases and where the existing systems failed than anyone else - probably enough to reconstruct those edge cases for the new sensors and use the resulting simulations to train the new sensors to a decent operational level, actually.

 

[diplomat33]

Except unless that current hardware suite turns out not sufficient for Level 5 no geofence or even meeting Waymo at Level 4, then all that benefit would be lost - and indeed become a major liability issue instead.

Not sure how the benefit would be lost or be a liability. Even if the hardware is not good enough for L5 or even Waymo level 4, Tesla would still be able to provide a lot of self-driving features with the hardware that would still benefit Tesla owners and add value. And even self-driving that is less than Waymo but available on a lot more cars that the public can buy would be attractive to a lot of folks. I mean if you need a car and your choice is using Waymo's L4 robotaxi (except ooops, it's probably not available where you live) or buying and owning a Tesla that has near L4 self-driving, you are going to buy the Tesla.

Remember that as far as we know, Waymo is not aiming to put the tech on cars that the public can buy at least not anytime soon. So no matter how good Waymo's FSD might become, the best that they can offer right now is robotaxis in very selective location which only benefits people who happen to live in those areas. Tesla will still be in a position to sell cars to the public with some form of self-driving that will benefit the owners everywhere.

 

[electronblue]

@Saghost @diplomat33

Context matters. I agree Tesla’s validation and deployment advantage would not disappear even if current hardware suite would be deemed insufficient for Lever 4 or 5.

What I said was Tesla would not be able to deploy Waymo-like tech to that wide fleet as was suggested. They would have to build a new fleet with a new sensor suite if the current fleet would not be sufficient.

The liability potential of such a move is obvious. They did pre-sell a tall tale and if they don’t deliver those disclaimers will surely be tested in a far bigger class action than thusfar — I’m sure everyone can see why that would be, even if they won’t admit it here.

 

[electronblue]

There are two or more potential avenus to victory here.

One is starting small, creating a robust technology set and once finished, rapidly deploying through fleet-build, licensing and whatnot. (Waymo)

The other is going big with the hardware deployment and working on software to make use of that. (Tesla)

Both have their pros and cons. With the latter you can indeed validate and gather data from the fleet. With the former you have the benefit of being able to tweak your suite as you go along and only ”commit” big time once you really are there. It also has the benefit of using a more expensive suite today with the knowledge that it will get cheaper over time and be cheap once your software etc are ready for the big time — Tesla had to commit to a suite that is cost-effective today, even though the software will only come in the future...

Tesla has an incentive to try to make the current suite work even if it would not be the optimal one. Waymo has no such limitations.

There are pros and cons to every approach. If you think the first approach has any merit, 10 Lyfts is not anything to sniff a — it would merely be a precursor to much bigger things.

 

[Daniel in SD]

They are obviously not interested in deploying their self-driving cars on a large scale.

This is a ridiculous assumption to make. Waymo DOES NOT have a working self-driving car yet! Obviously they have no interest in deploying self-driving cars that do not work at a large scale. If they ever do have a working self-driving car then I would expect them to try to deploy it at scale.

they are still working out the software and their cars lack the ability to do OTA software updates AFAIK

Yes, they have near self-driving cars that can be remotely controlled and yet they are stymied by their inability to perform OTA updates.

There are two or more potential avenus to victory here.

Exactly.

 

[CarlK]

Disagree. Even if it turns out they need more sensors to make FSD a reality, the cars they are building are watching human drivers deal with the edge cases and sending data back to Tesla. They're running in shadow mode and turning up discrepancies between what the NN would do and what the humans did - all without a single mile of Autopilot, let alone FSD. Under Autopilot, they're still learning more, capturing every disengagement and sorting out the important ones and finding more edge cases the system missed there.

That's a benefit the other companies just don't have, and one that won't go away if Tesla decides they need LIDAR (not gonna happen.) If they do add sensors, they won't immediately have the big fleet advantage for training those sensors, but they'll still know a lot more about the edge cases and where the existing systems failed than anyone else - probably enough to reconstruct those edge cases for the new sensors and use the resulting simulations to train the new sensors to a decent operational level, actually.

When you add sensor structure you need to retrain the NN from the begining. The thing about NN is even software people who designed it have no idea how it learn things (by modify the neural connections internally each time new data are fed to it). The best thing you might be able to do is to have the new set up running along side with the original NN in the shadow mode while training it. However those Lidar guys did not have much of a NN to begin with it's probabaly better for them to just start from a clean sheet of paper. That means years or additional work even if they could get everything right started at this juncture.

 

[Saghost]

When you add sensor structure you need to retrain the NN from the begining. The thing about NN is even software people who designed it have no idea how it learn things (by modify the neural connections internally each time new data are fed to it). The best thing you might be able to do is to have the new set up running along side with the original NN in the shadow mode while training it. However those Lidar guys did not have much of a NN to begin with it's probabaly better for them to just start from a clean sheet of paper. That means years or additional work even if they could get everything right started at this juncture.

If you trained a single NN to go straight from a video stream to driving outputs as some experts have suggested one should, you'd be correct.

From the investor presentations, it's clear Tesla didn't choose to do that. They clearly showed that they're using the NN to identify objects and driveable space and tag all the different cars and figure out distances.

Then there's a separate system they didn't really discuss (another NN, I assume) that's taking that digested environment data and Navigation requirements and making decisions about steering, acceleration and braking.

That being the case, it seems like you could easily add your putative additional sensor in the middle without starting over on the rest - with it's own set of analysis as required, you'd add it's data to the environment set being passed to the driving NN, or just cross compare the environment as analyzed by the two sets of sensor NNs and add a new dispute resolution logic/NN - and pass the final result back to the driving NN.

 

[acoste]

If you trained a single NN to go straight from a video stream to driving outputs as some experts have suggested one should, you'd be correct.

From the investor presentations, it's clear Tesla didn't choose to do that. They clearly showed that they're using the NN to identify objects and driveable space and tag all the different cars and figure out distances.

Then there's a separate system they didn't really discuss (another NN, I assume) that's taking that digested environment data and Navigation requirements and making decisions about steering, acceleration and braking.

That being the case, it seems like you could easily add your putative additional sensor in the middle without starting over on the rest - with it's own set of analysis as required, you'd add it's data to the environment set being passed to the driving NN, or just cross compare the environment as analyzed by the two sets of sensor NNs and add a new dispute resolution logic/NN - and pass the final result back to the driving NN.

Amnon plans on adding sensing to a working camera system where it is needed to create a redundant system, so it seems possible:

"Once you have a comprehensive solution with only cameras, then you can start pin-pointing where you need additional sensing," Mobileye CEO Amnon Shashua told me. "You are now in a better position to fine-tune exactly where you need a three-sensor modality, and where you need only two... If you fuse from the beginning, for every angle you'll need all three."

So, by adding lidar and radar latter, the other part of the three sensor types Shashua mentions, you can apply them more strategically. "

A lap of Jerusalem in Intel's surprisingly aggressive self-driving car

 

[banned-66611]

it's 10 cars "over the next few months" according to the article. Frankly, it's unimpressive. Sure, Waymo has excellent self-driving tech, their cars are "hands off", but their expansion rate is very slow. They are obviously not interested in deploying their self-driving cars on a large scale.

They are being very careful not to injure or kill anyone.

 

[banned-66611]

Not sure how the benefit would be lost or be a liability.

Simple. They have been selling the feature for years, have built a large number of cars that are supposed to have it, and are betting their future on revenue from the robotaxi service. If it turns out that they need to retrofit additional hardware to make it work, they are going to be facing a huge bill.

That's called a liability.

 

[diplomat33]

There are two or more potential avenues to victory here.

One is starting small, creating a robust technology set and once finished, rapidly deploying through fleet-build, licensing and whatnot. (Waymo)

The other is going big with the hardware deployment and working on software to make use of that. (Tesla)

Both have their pros and cons. With the latter you can indeed validate and gather data from the fleet. With the former you have the benefit of being able to tweak your suite as you go along and only ”commit” big time once you really are there. It also has the benefit of using a more expensive suite today with the knowledge that it will get cheaper over time and be cheap once your software etc are ready for the big time — Tesla had to commit to a suite that is cost-effective today, even though the software will only come in the future...

Tesla has an incentive to try to make the current suite work even if it would not be the optimal one. Waymo has no such limitations.

There are pros and cons to every approach. If you think the first approach has any merit, 10 Lyfts is not anything to sniff a — it would merely be a precursor to much bigger things.

That's actually a pretty good and balanced analysis.

For Waymo, I still think the challenge is in the "once finished" part. It is likely that Waymo will keep encountering edge cases and issues that prevent them from considering their system completely finished and ready for mass deployment. After all, Waymo has an excellent self-driving system now in their little corner of Phoenix and they are still encountering some issues from time to time that require a safety driver. Chances are that Waymo will do a slow deployment in other cities and encounter new unexpected issues that prevent the system from being completely "finished". So I don't think it will be as easy as just finish L4 and them pump out millions of self-driving Waymo cars. Waymo will need to do a relatively slow deployment for some time to come.

For Tesla, the challenge is of course the software. They still have some catching up to do in terms of just getting the basic features out. Also, Tesla is shooting for nothing less than L5 no geofence because they want their system to work everywhere, for every customer. But this makes the software problem much, much more difficult. And if the hardware is insufficient for L4/5 that would be a big problem. It is possible that Tesla could develop a pretty good generalized self-driving system that requires driver supervision but then they encounter some edge case where the cameras or the lack of a rear radar or something proves the hardware suite is totally insufficient.

But, personally, I think the hardware problem is a bit overblown. For one, Tesla engineers most likely considered the HW2 hardware in terms of camera quality, type, placement, radar, etc very carefully when they first decided to roll out AP2 back in Oct 2016. They didn't just slap some hardware on the car haphazardly and hope for the best. I am sure they studied the issue and planned things out very carefully. And ever since Oct 2016, Tesla has insisted that their cars have the necessary hardware for FSD. Tesla is obviously very confident about the hardware in terms of cameras, radar, ultrasonics, and of course the computer can be easily swapped out. They planned ahead of time for future computer upgrades. So I doubt that the issue will be something obvious. If the hardware does prove to be insufficient, it will be something unforeseen.

Second, there are many ways that Tesla could solve a problem before resorting to new hardware. Some issues can be solved with clever software. Some issues could be solved with specific driving rules. Some issues could solved with just a computer upgrade which is easy to do. Basically, there are a lot of things that Tesla engineers could do to overcome any weaknesses in the hardware without resorting to replacing the hardware. So I think it is likely that Tesla engineers would figure something out, rather than admit defeat.

But I do think that FSD is like a super giant puzzle with like a million pieces. Waymo's approach is to start in one corner of the puzzle and slowly connect adjacent pieces, and work their way out. So they have an entire corner completely done already. Tesla's approach is to have a large number of people (Tesla fleet) add pieces all over the puzzle at once. At first, it looks like a lot of disjointed pieces but when it is done, the entire puzzle will be done. The question is which approach will finish the entire puzzle first?

 

[acoste]

I think the hardware problem is a bit overblown. For one, Tesla engineers most likely considered the HW2 hardware in terms of camera quality, type, placement, radar, etc very carefully when they first decided to roll out AP2 back in Oct 2016. They didn't just slap some hardware on the car haphazardly and hope for the best. I am sure they studied the issue and planned things out very carefully. And ever since Oct 2016, Tesla has insisted that their cars have the necessary hardware for FSD. Tesla is obviously very confident about the hardware in terms of cameras, radar, ultrasonics, and of course the computer can be easily swapped out. They planned ahead of time for future computer upgrades. So I doubt that the issue will be something obvious. If the hardware does prove to be insufficient, it will be something unforeseen.

I have a different opinion on this. I think their approach is based on business decisions and not on technicals.

Situation:
- There is a car company who wants to increase its cars sales
- The company's marketing strategy is based on (instead of advertising) getting into the news every day
- Its CEO, to make the product more appealing, decides to deploy autonomous driving in its development stage.

Problem:
- the cost of a full set of autonomous tools would make the car very expensive, these parts will become cheaper over years
- even the high end cars (S and X) can't have the better hardware because the same parts should go into the cheaper cars (Model 3) which have smaller budget.
- developing an autonomous system in the background and deploying it once it reaches level 4 takes too long, and just a promise is not enough to bring in customers today

Plan:
- Start with minimum hardware (low res cams, 1 radar, computer), low cost.
- Develop the product on the fly and once you run into limitations, upgrade the affected hardware
- these hardware upgrades will generate new car sales with existing customers every 3 years. So that's great.
- owners will be ok with the hardware changes even if that's not what was promised because every hardware update will make people excited that this is a better (and final) product. (Very few new car buyers want to keep their cars beyond 3 years anyway)
- keep talking sh about competitors' products to support the strategy. Give them simple expressions: fool cell (more and more companies are developing it by the way); crutches (most other companies think lidar is crucial); simulators are not important (they are) and so on. These are all the products of the marketing department to help sales. Creating enemies is the best way of building a strong fan base.
- some accidents with the development hardware are expected but staying on Level 2 protects the company

Actions and my predictions:
- 2016. Sell cars with minimum hardware and the promise of full self driving to increase media presence and sales numbers
- 2019. Computer has reached its limits so lets update it
- 2021 NN becomes trained enough to drive reasonably well in cities in good weather on Level 3. But with the current hardware set, corner cases can't be solved. Let's deploy Tesla Network in very few selected areas with drivers (Tesla employees) just for the show. And let's update the resolution of the cameras
- 2024 The car drives well but there are still some corner cases that has to be solved so let's add more sensors (lidar, high res radar or some future sensors) for redundancy
- 2025 Reach geofenced, 3 seasons level 4 (no harsh weather). Deploy driverless Tesla Network in selected cities, like Phoenix