Waymo

[Ben W]

So it's a swiss-cheese kind of failure, where all the systems that did make mistakes had to line up in just the right ways:

It’s a well-worn observation that commercial plane crashes are typically due to about seven coincidental and compounding failures. As AVs become more and more reliable, we will start to see the same thing happening there, as in this case.

As far as radar/lidar, of course they don’t magically instantly solve anything, as they didn’t in this case. But they do enable certain problems to be much more tractably solved, by providing useful information that pure vision can’t. (To be clear, this specific telephone pole scenario should be solvable with pure vision.) Of course the system or NN needs to be properly trained/coded to process and interpret the radar/lidar information. I’m just saying that the overall system should be able to get to more 9’s more quickly with that information than without it, and that this is important for the future of Robotaxi and AVs in general.

 

[Daniel in SD]

If redundancy fails then getting rid of redundancy should eliminate the failures. It’s just basic logic.

 

[enemji]

If redundancy fails then getting rid of redundancy should eliminate the failures. It’s just basic logic.

Absolutely. For every sensor type, you need to account for their ambiguity and failures. That means at least 3x the work, in addition to the work required to build that fusion and handle the failures associated with it.

 

[willow_hiller]

If redundancy fails then getting rid of redundancy should eliminate the failures. It’s just basic logic.

Unironically, yes.

Why did this perception bug lead to a crash before the Waymo team was able to discover it? Their vehicles must pass hundreds if not thousands of poles on a daily basis. They have plenty of data to train a network to identify a pole and correctly label it as a hard object. But the perception network wasn't trained to correctly identify a pole because it never had to before. The HD map precluded all poles from the corpus of objects the network could correctly identify.

 

[Ben W]

Unironically, yes.

Why did this perception bug lead to a crash before the Waymo team was able to discover it? Their vehicles must pass hundreds if not thousands of poles on a daily basis. They have plenty of data to train a network to identify a pole and correctly label it as a hard object. But the perception network wasn't trained to correctly identify a pole because it never had to before. The HD map precluded all poles from the corpus of objects the network could correctly identify.

One of the more significant limitations of the way deep NN’s are trained is that they are fed millions of examples of what to do, but no examples of what not to do. And in training, they are penalized equally strongly for making a harmless mistake as for making a severe mistake; the feedback mechanism only cares that the output didn’t match the expected “correct” output, but it doesn’t care how. If these training systems had a way to integrate “avoid at all costs” examples into their training, that could be enormously helpful.

 

[Doggydogworld]

Q: Are the HD maps made "automatically" by something akin to (or even itself being) the perception stack they use in-vehicle.

I'm sure the HD mapping NNs they run offline are vastly more sophisticated than the in-car perception system. I know they've reduced the amount of human curation, but I doubt their system is good enough to take humans out of the loop altogether.

I think @willow_hiller may be on the right track with the idea that they only trained in-car perception on things in the driveable space. They never trained it on telephone poles because those are never in the road. Well, almost never:

 

[enemji]

View attachment 1056227
Screenshot added for posterity


View attachment 1056228

What does not make sense to me is this:

Waymo has every route HD mapped.
Waymo will only travel on those routes.
This means this back alley was also HD mapped.
How come they mapped this pole like a soft inflatable?

 

[willow_hiller]

What does not make sense to me is this:

Waymo has every route HD mapped.
Waymo will only travel on those routes.
This means this back alley was also HD mapped.
How come they mapped this pole like a soft inflatable?

Two separate failures.

The yellow paint striped area was missing from the HD map (it's flat, so it wouldn't necessarily show up on a LIDAR scan, they would need to use vision to map it and evidently they hadn't).

And then their perception system, encountering an object that it had rarely seen before in driveable space, classified it as something that would not cause damage to hit.

Waymo didn't address whether the pole was in the HD map or not, but I don't think the perception network would be tasked with determining whether an object would cause damage for any objects in the HD map. I can't imagine any permanently installed road objects that would not cause damage if hit, so it must not have been on the HD map, and it must have been classified as a cone or a inflatable-tube-man or something.

 

[enemji]

Waymo didn't address whether the pole was in the HD map or not,

I remember the first day this was reported, a waymo engineer commented it was given a weight of very low damage and hence this happened.

 

[Dewg]

Remind me to never dress up like a tree trunk for Halloween

 

[DanCar]

Waymo returns to minivan format for 6th gen, using Geely.

https://twitter.com/x/status/1802732940507373996

Reply / continuation tweet:
Waymo’s 6th-generation Driver will build on the unparalleled capabilities of our current-generation hardware with a simplified and cost-effective design that can autonomously navigate colder cities and help us further scale.

 

[Doggydogworld]

Waymo returns to minivan format for 6th gen, using Geely.

https://twitter.com/x/status/1802732940507373996

Reply / continuation tweet:
Waymo’s 6th-generation Driver will build on the unparalleled capabilities of our current-generation hardware with a simplified and cost-effective design that can autonomously navigate colder cities and help us further scale.

It carries it's own telephone pole around with it to do battle with those delinquent ones that hang out in alleys....

 

[DanCar]

Just hit button C, 30 secs into the video.

 

[diplomat33]

Anybody have any idea what this thing is on the back of the Waymo?

 

[Dewg]

Anybody have any idea what this thing is on the back of the Waymo?

Looks similar to a wide angle radar system. Seeing them go up around my town, used for traffic flow.

 

[diplomat33]

Looks similar to a wide angle radar system. Seeing them go up around my town, used for traffic flow.

View attachment 1057404

Thanks. That makes sense. I can definitely see the benefit of a wide angle radar in the rear corners of the car to accurately detect rear cross traffic. I just noticed that the sensors seems quite big and it sticks out a lot from the car.

 

[Yelobird]

Anybody have any idea what this thing is on the back of the Waymo?

My guesses would be an additional 30k to the car cost, a passenger head thumper, or external Sonos speakers to scream for help when stuck.

 

[diplomat33]

Waymo improves overall safety:

New data shows that the Waymo Driver continues to make roads safer. Over 14.8M rider-only miles driven through the end of March, it was up to 3.5x better in avoiding crashes that cause injuries and 2x better in avoiding police-reported crashes than human drivers in SF & Phoenix.

https://twitter.com/x/status/1803101356250849693

The collision avoidance clip is quite impressive.

 

[AlanSubie4Life]

Waymo improves overall safety:

https://twitter.com/x/status/1803101356250849693

The collision avoidance clip is quite impressive.

Why the awful video quality frame rate, and is it running at real time? Bizarre.

If you want to demonstrate your safety, at least have a decent video.

I have no idea if this was impressive. Why did the Waymo not slow down significantly? (Over the relevant period, it slowed from 34mph to 21mph. It took 1.5-2 seconds from the vehicle becoming visible to get this 13mph speed reduction, just 0.4g!). This was also a classic dangerous situation that warranted easing off and preparing to brake even before any car became visible - you don't just charge past stopped traffic like this.

It does seem clear that a collision was avoided because the other driver stopped. Not ideal.

 

[diplomat33]

Why the awful video quality frame rate, and is it running at real time? Bizarre.

The frame rate seemed ok to me.

I have no idea if this was impressive. Why did the Waymo not slow down significantly? (Over the relevant period, it slowed from 34mph to 21mph. It took 1.5-2 seconds from the vehicle becoming visible to get this 13mph speed reduction, just 0.4g!). This was also a classic dangerous situation that warranted easing off and preparing to brake even before any car became visible - you don't just charge past stopped traffic like this.

The Waymo had a green light. The other car was occluded by the white car in the adjacent lane, so it is only visible when the Waymo is already at the intersection. By the time you see the other car making the turn into the Waymo, there was less than 1 second to avoid a collision. The Waymo reacted very quickly. Very impressive reaction time by the Waymo Driver.

It does seem clear that a collision was avoided because the other driver stopped. Not ideal.

No. The other car brakes but not enough to avoid a collision. If the Waymo had not swerved, there would have been a collision. That was all on the Waymo that there was no collision.