Waymo

[flutas]

The Waymo apparently makes an illegal move but one illegal move does not not necessarily mean that driverless is far from widespread. And Waymo may have already fixed this issue. This video is certainly not emblematic of widespread issues. Waymo does thousands of perfectly good rides every week. Furthermore, Waymo has driverless that is very safe in a growing ODD (multiple cities, all roads including highways, parking lots, airport terminals, 24/7, rain and fog). We need to look at the total picture and not just one video. When we look at the big picture, Waymo's driverless is very capable and scaling very well. I would say broad rollout of driverless is closer than we think. It is certainly not going to be a long while. Lastly, remember that driverless does not need to be perfect to be rolled out broadly. In fact, we will likely see a broad rollout of driverless before it is perfect.

All this could be said exactly for Cruise as well, how's that approach working out? I mean, they only had two "bugs" meet right? It made a judgement error and their classifier couldn't detect what a leg was, as you said yourself.

Cruise perception made error in characterizing the collision. It thought the pedestrian hit the side of the vehicle instead of the front. As a result, it did not identify the pedestrian as going under the vehicle and it did not identify the legs because the classifier was not trained on only identifying legs alone:

Also Jesus, this response to doubt on Waymo REALLY feels like corporate speak.

I'm not saying you're a Waymo shill, well...because I know you aren't, but I feel like you're misguided on how you respond. You instantly jump to downplay (green), misdirect (red), and speculate (yellow). Then flow into what sounds like a PR generated speech (purple).

 

[diplomat33]

All this could be said exactly for Cruise as well, how's that approach working out? I mean, they only had two "bugs" meet right? It made a judgement error and their classifier couldn't detect what a leg was, as you said yourself.

No it cannot be said about Cruise. You are downplaying Cruise. They did not just have two "bugs" while doing great overall. Cruise had many serious issues long before this latest accident with the woman that got dragged. For months prior to the accident, people pointed out serious perception and prediction issues. For example, hitting the Prius at the intersection, hitting the articulated bus, driving too close to pedestrians/not yielding to pedestrians, driving through police tape, several mass stalls, frequent stalls at intersections, hitting fire truck, frequently blocking first responders. So it was way more than just 2 "bugs".

I'm not saying you're a Waymo shill, well...because I know you aren't, but I feel like you're misguided on how you respond. You instantly jump to downplay (green), misdirect (red), and speculate (yellow). Then flow into what sounds like a PR generated speech (purple).

I don't think I am misguided (at least I hope not). My point was to guard against jumping to conclusions. Too often, I see people taking one video and assuming it represents the whole (good or bad). That is a mistake. In this case, taking an example of Waymo making an illegal move (I did not really know it was illegal until someone told me) and assuming that driverless is very far away from being solved as the poster seemed to be suggesting. I don't think we can reach that conclusion just from this one video. And that was my main point.

Do I downplay Waymo's error? Yeah maybe. But it is also a mistake to go the other way and conclude that this error is proof that driverless will not happen for a long time.

Do I speculate about Waymo fixing the issue? Yes. That is why I wrote "may have fixed the issue". The word "may" indicates that I am speculating. I did not hide that I was speculating.

I don't think I misdirect. What I said is true: one video should not be taken as representative of the whole (good or bad). And we should always try to look at the big picture and not focus on just one example of bad or good behavior. And I was trying to point out the "big picture" to provide a broader perspective.

Waymo is scaling to a bigger ODD (ex: adding driverless testing on highways) and scaling more rides. That is a fact. I can see how it might sound like corporate PR. But I am simply stating the information that I have. From what I can see, Waymo seems to be doing quite well in their rides and they are covering more area. I acknowledge that Waymo still has issues to solve. But we cannot ignore the positives either.

 

[diplomat33]

I don’t care about the illegal move (though why would that be programmed to do that around a vehicle that is not stationary?).

It’s the interaction of autonomous vehicles. Seems like a potential problem.

Thanks for clarifying. So you feel having two AVs that don't communicate with each other and might be confused by the other's intentions, is the real issue? Looking at the video, we can see the two Waymos inch forward like they are not sure who is going to yield. It seemed like both Waymos thought the other was going to yield. Could this be an issue with AVs interacting with each other? Maybe. But I still don't think we can conclude that it is proof that driverless has a long way to go. Yes, I am speculating but I think that these are problems that can be fixed. They are not impossible problems to fix.

Ultimately, I guess it depends on if you view AVs as a glass half empty or glass half full. In other words, do you focus on the problems still unsolved or do you focus on all the problems that have been solved so far. I am a glass half full kind of guy. I believe the remaining problems left to solve in autonomous driving are less than the problems we have already solved so far. Having said that, the remaining problems might be fewer but they are likely harder to solve. But we will get there.

 

[flutas]

No it cannot be said about Cruise. You are downplaying Cruise. They did not just have two "bugs" while doing great overall. Cruise had many serious issues long before this latest accident with the woman that got dragged. For months prior to the accident, people pointed out serious perception and prediction issues. For example, hitting the Prius at the intersection, hitting the articulated bus, driving too close to pedestrians/not yielding to pedestrians, driving through police tape, several mass stalls, frequent stalls at intersections, hitting fire truck, frequently blocking first responders. So it was way more than just 2 "bugs".

Yes, they did have two bugs. One was their perception of not understanding where the impact came from, the second not recognizing human legs as an obstacle, per your own words:

Cruise perception made error in characterizing the collision. It thought the pedestrian hit the side of the vehicle instead of the front. As a result, it did not identify the pedestrian as going under the vehicle and it did not identify the legs because the classifier was not trained on only identifying legs alone:

Now to get on to why I don't trust anything you say anymore and the meat of "who here is downplaying" red highlights in your quote above for some fun....

in 2021 you claimed in fact that Cruise has "very sophisticated prediction and planning".

I thought the Cruise Under the Hood video was very impressive. Cruise does have very sophisticated prediction and planning. In particular, they showed off how their autonomous driving is programmed to handle uncertainty which is key for safe driving.

(...)

More recently you claimed (in the context of stalls, which is directly related to prediction and planning)...

Waymo and Cruise are constantly getting better over time with software updates as the engineers add more data and train their neural nets.

So their prediction and planning must be even better than in 2021 when it was "very sophisticated"...yet isn't sophisticated enough to figure out where an impact is from (accelerometer anyone?) and that a human leg (or any object for that matter that magically dissapears) sticking out from under the car is an obstacle to avoid.

It's like a parrot with ADHD that can only quote the latest PR speak given to them.

Also a fun little quote I dug up from you regarding prediction and planning failures of Cruise, I guess you're technically correct that they "haven't killed anyone" just badly mangled them.

I am not dismissing anything. But I hate when people say "Can we end this experiment now or does someone need to be killed first?" It is hyperbolic fearmongering from people who want to ban robotaxis. It basically implies that robotaxis are completely unsafe and untested tech that companies are just putting out there to see what happens and that robotaxis are deadly and will kill people. And that is not what is happening. AV companies, like Waymo, have actually been very cautious in how they deploy their robotaxis. They geofence them, limit their ODD, monitor them, And they spent years testing them with safety drivers and only removed the safety drivers when they were ready to do so, and only in a limited ODD. And Cruise and Waymo have published data that shows they are very safe in terms of reducing injuries and deaths on the road. Do they still have issues? Of course. Every new tech has issues. In fact, maybe you could make a case that Cruise or Waymo should keep safety drivers a bit longer. But Cruise and Waymo are fixing these issues. But none of these stalls killed anyone and they won't kill anyone since the cars are stopped. Also, how do these stalls compare to human drivers? People act like it is only robotaxis that are causing these problems when the reality is that human drivers are far worse. Human drivers block traffic too. Human driven cars also sometimes break down or get a flat tire or something. We don't freak out when a human driven car breaks down so why do we freak out when a robotaxi does? And human drivers actually cause accidents that injure and kill people. So maybe we should remove the human experiment from the road since it is so deadly? But no, they want to completely ban robotaxis because on some rare occasions, they block traffic for a few minutes. How silly!

Don't get me wrong. I think the stalls are bad. I've been critical of the stalls, especially the Cruise ones that are really bad. But Waymo and Cruise are addressing the problem. Why doesn't the article talk about the good safety of robotaxis? Instead, these articles just quote a bunch of people who are annoyed by robotaxis and bad mouth them and push the narrative that robotaxis are dangerous. I am just tired of seeing these articles that only push one side of the story and push fearmongering.

 

[diplomat33]

Yes, they did have two bugs. One was their perception of not understanding where the impact came from, the second not recognizing human legs as an obstacle, per your own words:

No. I am talking about Cruise's failures before the accident. They had way more problems than just the two problems in this specific accident.

Now to get on to why I don't trust anything you say anymore and the meat of "who here is downplaying" red highlights in your quote above for some fun....

in 2021 you claimed in fact that Cruise has "very sophisticated prediction and planning".

More recently you claimed (in the context of stalls, which is directly related to prediction and planning)...


So their prediction and planning must be even better than in 2021 when it was "very sophisticated"...yet isn't sophisticated enough to figure out where an impact is from (accelerometer anyone?) and that a human leg (or any object for that matter that magically dissapears) sticking out from under the car is an obstacle to avoid.

It's like a parrot with ADHD that can only quote the latest PR speak given to them.

Also a fun little quote I dug up from you regarding prediction and planning failures of Cruise, I guess you're technically correct that they "haven't killed anyone" just badly mangled them.

Interesting that you go all the back to 2021. Back in 2021, what Cruise showed us, looked very good. We did not know then what we know now. You cherry pick and conveniently ignore all my more recent posts where I slammed Cruise for being terrible when the problems came to light. But that is inconvenient to your narrative that somehow I am parroting PR.

But if you want to ignore my posts because you think they are just Waymo PR, be my guest. Nobody is forcing you to read my posts.

And attacking me, trying to attack my credibility, by drudging up past posts of mine, to accuse me of parroting PR, is frankly weak and not constructive to this thread. Let's keep things on topic and avoid personal attacks.

 

[Doggydogworld]

If the cost of running Waymo in '22 was -2-3B in total, most of that would be from operations and not R&D. Perhaps 20% is R&D.

If that's true they might as well just shut down. Take 20% R&D out of that 2b and that leaves 1.6b in operating costs. They did ~800k rides in 2023, so that's $2000 per ride (!!!) in direct costs -- remote ops, roadside assist, customer support, etc. plus vehicle depreciation, maintenance and fuel.

That 2b has to be R&D. It's the only way the numbers make any sense.

 

[spacecoin]

If that's true they might as well just shut down. Take 20% R&D out of that 2b and that leaves 1.6b in operating costs. They did ~800k rides in 2023, so that's $2000 per ride (!!!) in direct costs -- remote ops, roadside assist, customer support, etc. plus vehicle depreciation, maintenance and fuel.

That 2b has to be R&D. It's the only way the numbers make any sense.

We really have no idea. We don’t know what their revenue is. We don’t know what the capex/opex mix nor what the total expenses nor r&d is.

I just know it drives well and that they will scale out faster and faster from here.

 

[flutas]

No. I am talking about Cruise's failures before the accident. They had way more problems than just the two problems in this specific accident.



Interesting that you go all the back to 2021. Back in 2021, what Cruise showed us, looked very good. We did not know then what we know now. You cherry pick and conveniently ignore all my more recent posts where I slammed Cruise for being terrible when the problems came to light. But that is inconvenient to your narrative that somehow I am parroting PR.

But if you want to ignore my posts because you think they are just Waymo PR, be my guest. Nobody is forcing you to read my posts.

And attacking me, trying to attack my credibility, by drudging up past posts of mine, to accuse me of parroting PR, is frankly weak and not constructive to this thread. Let's keep things on topic and avoid personal attacks.

You mean like the post from August 9th 2023 (not even 2 months before the drag incident, mind you) that I quoted in my previous post, where you wanted to focus on the safety of the cars rather than their planning and perception defects...

I am not dismissing anything. (...) Cruise and Waymo are fixing these issues. But none of these stalls killed anyone and they won't kill anyone since the cars are stopped. Also, how do these stalls compare to human drivers? (...)

Don't get me wrong. I think the stalls are bad. I've been critical of the stalls, especially the Cruise ones that are really bad. But Waymo and Cruise are addressing the problem. Why doesn't the article talk about the good safety of robotaxis? (...)

So instead of just discussing an issue in perception and planning (stalls) you instantly side step to "BUT THEY'RE SAFE WHY NOT TALK ABOUT THAT."

Any other remarks you want to make? Since you seem to focus on the date there and nothing else, even though I quoted more recent examples in my previous posts?

Mmhmmm.... Gotta love the "why are you using my words against me" argument as well, clearly this is all personal attacks and not judging you for exactly what you say. Cruise and Waymo of today must be better than 2020 Cruise right, I mean after all they "are addressing the problems" as you said.

After all it's not like Waymo can't even have their cars stop properly for rider support in 2021 or 2024...[0], [1], let alone an actual emergency.

 

[diplomat33]

You mean like the post from August 9th 2023 (not even 2 months before the drag incident, mind you) that I quoted in my previous post, where you wanted to focus on the safety of the cars rather than their planning and perception defects...

I rightfully criticized the safety issues that Waymo and Cruise were having while also pointing out the truth that Waymo and Cruise were working to address those issues. I think that is fair.

After all it's not like Waymo can't even have their cars stop properly for rider support in 2021 or 2024...[0], [1], let alone an actual emergency.

Yes, Waymo still has some issues to resolve. I have never denied that. But there are many issues that Waymo used to have that they don't have anymore. Waymo is working to address these issues. But they still have issues they have not solved yet. Congrats. You just proved the super obvious that Waymo has not solved autonomous driving yet.

 

[flutas]

I rightfully criticized the safety issues that Waymo and Cruise were having while also pointing out the truth that Waymo and Cruise were working to address those issues. I think that is fair.

I didn't know saying "I'm not dismissing anything" followed by "why aren't we talking about how safe they are" is "pointing out the truth that Waymo and Cruise were working to address those issues."

But I guess maybe we're using different versions of English?

Yes, Waymo still has some issues to resolve. I have never denied that. But there are many issues that Waymo used to have that they don't have anymore. Waymo is working to address these issues. But they still have issues they have not solved yet. Congrats. You just proved the super obvious that Waymo has not solved autonomous driving yet.

So apparently a simple "don't move" flag that CS (sorry "rider support") can apply to a car has taken 3+ years so far?

What this tells me is that one hand of Waymo doesn't know what the other is doing, much like Cruise. Which in an autonomous driving context, yeah... that doesn't exactly inspire confidence.

 

[flutas]

Apologies for the double post, can't edit the previous one anymore but I wanted to add:

My whole reason for starting this dialogue was because I'm tired of any negative discussion around Waymo, Cruise, Comma AI, Et al. being instantly shut down in this or other threads. Instead of discussing any issues the reply is always instantly "Ok, but it doesn't matter and anyways your wrong. Here's how good they are!"

As an example post and reply to a video of two Wayno vehicles struggling with each other:

Makes one realize how far away broad rollout of driverless is. Going to be a while!

Why would you say this? The Waymo apparently makes an illegal move but one illegal move does not not necessarily mean that driverless is far from widespread. And Waymo may have already fixed this issue. This video is certainly not emblematic of widespread issues. Waymo does thousands of perfectly good rides every week. Furthermore, Waymo has driverless that is very safe in a growing ODD (multiple cities, all roads including highways, parking lots, airport terminals, 24/7, rain and fog). We need to look at the total picture and not just one video. When we look at the big picture, Waymo's driverless is very capable and scaling very well. I would say broad rollout of driverless is closer than we think. It is certainly not going to be a long while. Lastly, remember that driverless does not need to be perfect to be rolled out broadly. In fact, we will likely see a broad rollout of driverless before it is perfect.

Instead of being able to discuss two Waymo vehicles interactions with each other, or what the poster meant, the reply was instantly a personal attack to start "Why would you say this?" that isn't a discussion about the content.

Then "yeah it made an illegal move, but who cares, and you're wrong because I'm sure they fixed it"

Followed by "here's why Waymo is jeebus confirmed, and PS you're still wrong."

 

[diplomat33]

Apologies for the double post, can't edit the previous one anymore but I wanted to add:

My whole reason for starting this dialogue was because I'm tired of any negative discussion around Waymo, Cruise, Comma AI, Et al. being instantly shut down in this or other threads. Instead of discussing any issues the reply is always instantly "Ok, but it doesn't matter and anyways your wrong. Here's how good they are!"

As an example post and reply to a video of two Wayno vehicles struggling with each other:



Instead of being able to discuss two Waymo vehicles interactions with each other, or what the poster meant, the reply was instantly a personal attack to start "Why would you say this?" that isn't a discussion about the content.

Then "yeah it made an illegal move, but who cares, and you're wrong because I'm sure they fixed it"

Followed by "here's why Waymo is jeebus confirmed, and PS you're still wrong."

I think you are misinterpreting my words. I did not launch any personal attacks. And I never said that Waymo is perfect and we are not allowed to discuss or criticize Waymo. You are misinterpreting my defense of Waymo as shutting it down. I was not doing that. I am perfectly willing to discuss or critic the Waymo issues if you want. If I point out that Waymo is generally safe in the overall big picture, I am not shutting down discussion. I am just trying to provide balance because too often I see people just want to make fun of Waymo's mistakes and nothing else. I am tired of people always being negative and trashing Waymo that yeah, I tend to push back against that. Maybe I get too defensive too. I am perfectly open to a balanced discussion.

If you read my initial reply to the other poster, you saw that I did talk about the issue in the video. I am happy to elaborate. I think the two Waymos were uncertain about who was going to yield. We see that in the behavior of the Waymos, both inching forward like they are waiting for the other to yield. This issue could be caused by the fact that both Waymos use the same prediction module so both Waymos are making the same prediction that the other vehicle is moving so it should yield. I said in my post that this could be an issue but one that I would speculate Waymo will probably work on. Finally, we see one Waymo take the lead and the other yielding. But since the Waymo that took the lead in going was in the left lane and making a right turn, it cuts in front of the other Waymo, in an illegal move. This is issue #2. Waymo should work on making sure the Waymo in left lane does not cut in front of a car to the right.

As you can see I am perfectly willing to discuss the Waymo issue.

 

[powertoold]

Here's enough room to squeeze here but it can't because the driveable space is predefined:

 

[diplomat33]

Here's enough room to squeeze here but it can't because the driveable space is predefined:

This is false. The drivable space is not pre-defined. We know this because Waymo encounters construction zones where the drivable space is different from the map and it is able to drive it just fine. There have been other cases where Waymo is able to drive through tight squeezes with no problem. We've seen it in videos. I am not making this up. Waymo engineers have said that the car can define its own drivable space. Waymo defines the drivable space based on the cameras, lidar and radar. If the Waymo did not move in this case, it is for a different reason.

 

[powertoold]

This is false. The drivable space is not pre-defined. We know this because Waymo encounters construction zones where the drivable space is different from the map and it is able to drive it just fine. There have been other cases where Waymo is able to drive through tight squeezes with no problem.

I know, but this is show me you don't know what drivable space means without telling me

 

[diplomat33]

I know, but this is show me you don't know what drivable space means without telling me

I know what drivable space is. It is the space that the car's autonomous driving system says is ok to drive on. Autonomous cars will define space that is drivable like roads and space that is not drivable like a sidewalk. In this case, the reason the Waymo did not move is not because the lane was pre-defined as not drivable as you claimed.

Also, if you want to speculate as to why the Waymo did not move, that's fine. We do a lot of speculation/discussion about autonomous driving on this forum. But please do not state your opinion/speculation as fact.

 

[Daniel in SD]

Here's enough room to squeeze here but it can't because the driveable space is predefined:

It could just be programmed not to drive across private driveways. What difference does it make if it determines what is a private driveway in realtime or if it's predefined? It doesn't look like the drivable space in the alley has changed recently. Is your hypothesis that a human mapped out the drivable space in the alley?

 

[powertoold]

It could just be programmed not to drive across private driveways. What difference does it make if it determines what is a private driveway in realtime or if it's predefined? It doesn't look like the drivable space in the alley has changed recently. Is your hypothesis that a human mapped out the drivable space in the alley?

I'm basically just repeating what the passenger says, pre mapped, there's enough room, no way to change it on the fly, stuck because of that

That little piece missing is indicative of some human going over and polishing it up imo, it's difficult for me to see how a NN would be able to differentiate a tiny piece of "private" driveway in a tiny alley like that

 

[Bitdepth]

I'm basically just repeating what the passenger says, pre mapped, there's enough room, no way to change it on the fly, stuck because of that

You are basically making up what the passenger says.

Kevin Chen: "so I think the thing is, if you look at here, it's like even though we could probably kind of cut into the driveway on the left a bit, uh its drawn as a road boundary so I wonder if that's uh related why we are not willing to um, the car isn't willing to just kind of squeeze a bit, um it does seem like there might be enough room".

To rephrase, Kevin is speculating that they have drawn boundary to prevent the car from going into people's driveway even though it does seem there is space for the car to squeeze through.

That little piece missing is indicative of some human going over and polishing it up imo, it's difficult for me to see how a NN would be able to differentiate a tiny piece of "private" driveway in a tiny alley like that

Road boundary estimation is not a hard problem to solve. Look at your FSD beta visualization, it does not mark driveways as drivable space. Why do people still believe mapping is hard when your FSD software does it in real-time with limited compute vs offline HD Map generation.

 

[diplomat33]

Waymo I-Pace in Buffalo spotted with different sensor rack on roof.

https://twitter.com/x/status/1752025270142267797

That sensor rack looks like a temporary retrofit, not a final version. Waymo is doing winter testing in Buffalo. I wonder if Waymo is testing new sensors designed to do better in winter weather? Maybe sensors for mapping? Or maybe Waymo is testing the sensors that will go on the Geely vehicle although they look too big compared to the Geely prototypes we've seen?