Waymo

[MP3Mike]

It doesn’t seem like not allowing the car to drive there would work well. What if another car is approaching and that space is needed to allow it to pass?

Stand your ground!

 

[AlanSubie4Life]

FSD works awesome in alleys with no lane markings. Even better if there are markings. I use it everyday.

Just does not work as well with trains! (That report seems pretty legit.)

Every system has gaps. Lots of them! I assume at some point Waymo will run into something even more significant than a telephone pole. Just the way things go with difficult problems. We’ll see. Of course, Waymo knows this and is trying very hard to avoid that event.

 

[enemji]

Just does not work as well with trains! (That report seems pretty legit.)

Every system has gaps. Lots of them! I assume at some point Waymo will run into something even more significant than a telephone pole. Just the way things go with difficult problems. We’ll see. Of course, Waymo knows this and is trying very hard to avoid that event.

I have tried the train scenario numerous times. It works.

 

[AlanSubie4Life]

I have tried the train scenario numerous times. It works.

I believe you. It just doesn’t always work.

I expect train crossings to work when I use FSD, but I also expect them to fail. It’s a combination of working and also not working.

 

[enemji]

I believe you. It just doesn’t always work.

I can agree on that

 

[willow_hiller]

I believe you. It just doesn’t always work.

I expect train crossings to work when I use FSD, but I also expect them to fail. It’s a combination of working and also not working.

Given the fellow that reported FSD not yielding for a train barrier was going over 60 MPH on a rural highway, I would assume that it was operating in V11 mode instead of V12 mode. So I would be curious to see whether V12 would have failed in the same situation.

In my area, and imagine many others, train crossings are on low-speed roads instead of highways, which is why we observe very different behavior.

 

[AlanSubie4Life]

Given the fellow that reported FSD not yielding for a train barrier was going over 60 MPH on a rural highway, I would assume that it was operating in V11 mode instead of V12 mode.

It looked like a road type where v12 would be active. Not aware of any speed dependence of v11 activation and certainly not at fairly low speed of 60mph (I routinely use v12 at those pedestrian speeds). Anyway it is off topic here. Can discuss in that thread if you want.

 

[kabin]

I just noticed Waymo expanded their Phoenix coverage again. Not sure if they operate on freeways but I wouldn't be surprised if they did.

 

[Doggydogworld]

It doesn’t seem like not allowing the car to drive there would work well. What if another car is approaching and that space is needed to allow it to pass?

I think Fleet Response always handled such cases in the past. In fact, Waymo's Fleet Response blog article has this scenario -- FR directs a car to go off the mapped driveable space onto a private driveway to help unclog a road.

PoleGate shows multiple errors. First, the map was wrong. The striped area should be off limits. (The pole should be on the map, too, but that's a whole 'nother story.) Map errors happen, though. The car should notice the stripes when building its real-time map. Maybe notify the mother ship of the mismatch. Maybe even ask FR if it's OK to use the striped area for pickup.

Did Perception simply miss the stripes? That plus missing the giant pole would be two colossal failures. If Waymo perception sucks that bad after all these years they're pretty much doomed.

I think something else is going on. I think FR calls were costing too much and Ruth Porat lowered the boom. Reddit dude with very specific info claims they moved FR to the Philippines in April, btw. IMHO they simultaneously reduced FR call frequency by granting the car more autonomy to override boundaries. Suddenly we see a rash of unprecedented incidents -- PoleGate, extended jaunts down the wrong side of the road and crazy swerving behind TrailerTree.

Under my theory, Perception saw the stripes but Planning overrode it without calling FR cuz Ruth said no. This might even be a special PuDo Planner with more leeway to violate boundaries than the normal En Route Planner.

But why hit the pole? That's the whole "point" of lidar (haha). Sure, you can do fancy scene analysis with lidar NNs, sensor fusion and other toys. But you have a #%#$^ point cloud. If a bunch of highly correlated stationary points are right in front of you STOP THE CAR. You don't need to know what it is, you don't need fancy NNs and you sholdn't need a "damage score". Just stop. Maybe you get a false positive every 50k miles, but so what?

 

[flutas]

For anyone wanting the PDF of the recall, I've linked it at the bottom of the post.

It provides a (tiny) bit more than regurgitated bites from the news.

Prior to the Waymo ADS receiving the remedy described in this report, a
collision could occur if the Waymo ADS encountered a pole or pole-like
permanent object and all of the following were true: 1) the object was within
the the boundaries of the road and the map did not include a hard road edge
between the object and the driveable surface; 2) the Waymo ADS’s perception
system assigned a low damage score to the object; 3) the object was located
within the Waymo ADS’s intended path (e.g. when executing a pullover near
the object); and 4) there were no other objects near the pole that the ADS
would react to and avoid.

PDF

 

[MP3Mike]

The official recall is out for PoleGate: Recalls | NHTSA

Details of the recall: https://static.nhtsa.gov/odi/rcl/2024/RCLRPT-24E049-1733.PDF

On May 22, 2024, Waymo’s Field Safety Committee began overseeing the analysis of the event and the development and evaluation of two mitigations: a fully-validated software update to improve the ADS response to pole or pole-like permanent objects, and robust mapping updates and improvements to ensure that the map meets Waymo specifications near pole or pole-like permanent objects by including a hard road edge between the object and the driveable surface.

 

[enemji]

Given the fellow that reported FSD not yielding for a train barrier was going over 60 MPH on a rural highway, I would assume that it was operating in V11 mode instead of V12 mode. So I would be curious to see whether V12 would have failed in the same situation.

True. The only train crossings I have had were in the city, so city speed limits apply.

 

[enemji]

The official recall is out for PoleGate: Recalls | NHTSA

Details of the recall: https://static.nhtsa.gov/odi/rcl/2024/RCLRPT-24E049-1733.PDF

Only Poles. That is faf

 

[enemji]

It doesn’t seem like not allowing the car to drive there would work well. What if another car is approaching and that space is needed to allow it to pass?

I ran into that situation yesterday. Had to takeover and turn into the driveway on the left out for the other to pass

 

[enemji]

I've found the exact location of the crash, and it probably raises more questions than it answers. Local news reports said the intersection of 7th and Roosevelt, and here is the alley near that intersection that matches the photographs. I put the map marker directly on top of the pole: 33°27'29.2"N 112°04'55.1"W · Downtown Phoenix, Phoenix, AZ 85003

Screenshot added for posterity

 

[willow_hiller]

For anyone wanting the PDF of the recall, I've linked it at the bottom of the post.

It provides a (tiny) bit more than regurgitated bites from the news.



PDF

Interesting, so "hard road edge" is actually the language Waymo used to describe the yellow-paint-striped area. I guess they're marking it on their map as a kind of "virtual curb" so the vehicle considers it to be a hard barrier that it should never cross under any circumstances.

Ideally, the Waymo driver would be able to make the contextual decision with an onboard network as to what paint on the road indicates.

 

[MP3Mike]

Interesting, so "hard road edge" is actually the language Waymo used to describe the yellow-paint-striped area. I guess they're marking it on their map as a kind of "virtual curb" so the vehicle considers it to be a hard barrier that it should never cross under any circumstances.

I'm not sure they are referring to the yellow striped area:

robust mapping updates and improvements to ensure that the map meets Waymo specifications near pole or pole-like permanent objects by including a hard road edge between the object and the driveable surface.

It sounds to me like they are putting a "virtual curb" around every single pole in a drivable surface even if it doesn't have paint markings near it.

 

[willow_hiller]

I'm not sure they are referring to the yellow striped area:


It sounds to me like they are putting a "virtual curb" around every single pole in a drivable surface even if it doesn't have paint markings near it.

I understand the point that Waymo achieves reliability by overlapping multiple systems (HD map, LIDAR, camera), but it really sounds like some of those systems are very fragile and have weaknesses that align rather than fully back each other up.

The perception system was reliant on the HD map to avoid encountering objects that don't typically appear in drive-able space, so a failure of the HD map immediately became a failure of the perception system.

 

[Daniel in SD]

I understand the point that Waymo achieves reliability by overlapping multiple systems (HD map, LIDAR, camera), but it really sounds like some of those systems are very fragile and have weaknesses that align rather than fully back each other up.

The perception system was reliant on the HD map to avoid encountering objects that don't typically appear in drive-able space, so a failure of the HD map immediately became a failure of the perception system.

That’s the definition of redundancy. Two failures were required for the system to fail.
We never find out when there is a failure in perception or HD maps that doesn’t cause a problem because the redundancy worked.

 

[flutas]

That’s the definition of redundancy. Two failures were required for the system to fail.
We never find out when there is a failure in perception or HD maps that doesn’t cause a problem because the redundancy worked.

I think it's more than just a yes or no, but whether I would agree that their weaknesses align or not depends on quite a few questions I'll likely never have the answers to.

An example and the flow:

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

If yes
- Are the maps then reviewed?
If Yes
- Their failures may line up. (Much further questioning to go here for a real answer)
If no
- Their failures line up.
If no
- What is the process for creating the maps. (Their failures may or may not line up, depending on how the maps are created.)

The fact that Waymo has touted being able to automatically update their HD maps with the cars (and that their mapping cars are the same as the production cars) makes me think that their failures do line up and this was an example of that.