Waymo

[stopcrazypp]

Does this mean Waymo needs to hand code every curb and telephone pole?
If so, sounds difficult to scale.

Not for the telephone pole, sounds like they just need to classify it properly into a category of hard objects instead of soft objects. This is an example of how lidar doesn't automatically identify objects (as people like to claim), but rather, just like with cameras, you still need to properly classify them.

For the curbs, sounds like sometimes they do need to hand code things into the map. I'm surprised they don't have an automated curb detection system (if not on the vehicles themselves, at least in the mapping software). It's possible however that alleyway was not part of their maps.

Edit: from reading the article, sounds like there isn't actually a raised curb but rather there is just a yellow line to delineate the road boundary, while the road is at the same grade. If they were using lidar based mapping, depending on how that yellow line was laid (a tape or paint; if paint, what type and how thick/reflective vs the road surface) and the lidar used (is it the type that can read road lines) it's possible it can't be detected via automated mapping (at least not with current software, may need to switch to vision based mapping or a hybrid between the two).

 

[diplomat33]

Yikes. I guess this is related to how they deal with soft debris? i.e. tumbleweeds, balloons, plastic bags, etc.

Yes, that is my assumption as well, that Waymo assigns a damage score to objects. Objects like tumbleweeds, balloons, plastic bags, get a low damage score to indicate that it is ok to drive over them whereas objects like poles, cars, pedestrians should get a high score to indicate that the car should avoid colliding with them. In this case, the pole got a low score by mistake because of a software bug, causing the car to think it was safe to drive over and colliding with it.

I wonder if this damage score is heuristic code or part of the NN? It almost sounds like heuristic code to me. If so, I would imagine that Waymo should work on removing that heuristic code and having the NN recognize objects that can be safely driven over or objects to avoid.

You would think it should be tuned to be super conservative...

That's the issue though. If you are too conservative, then the car will phantom brake for objects that it does not need to.

 

[diplomat33]

Does this mean Waymo needs to hand code every curb and telephone pole?

No

 

[diplomat33]

For the curbs, sounds like sometimes they do need to hand code things into the map. I'm surprised they don't have an automated curb detection system (if not on the vehicles themselves, at least in the mapping software). It's possible however that alleyway was not part of their maps.

I would hope that Waymo is not hand coding every curb. Surely, the cameras and lidar can detect curbs. And yeah, this might be an exception where the alleyway was not part of their maps. Normal HD maps likely automatic label curbs.

 

[Max Spaghetti]

Isn't it smothered in Lidar and Radar that apparently Teslas can't drive without?

How could it not identify it as a long vertical obstacle?

Whether it thought it was hard or soft should be irrelevant, why would it even want to hit a long vertical soft obstacle?

I wonder what Tesla's vision-based goo-mode parking would think of that pole.

 

[diplomat33]

Isn't it smothered in Lidar and Radar that apparently Teslas can't drive without?

How could it not identify it as a long vertical obstacle?

Whether it thought it was hard or soft should be irrelevant, why would it even want to hit a long vertical soft obstacle?

I wonder what Tesla's vision-based goo-mode parking would think of that pole.

The Waymo did identify the long vertical obstacle. But there was a bug in the software that told the car to ignore the obstacle.

That is why it is not as simple as if you are camera only, you will never hit objects or if you use lidar and radar, you will never hit objects. You can be vision-only or have lots of lidar and radar, if your NN has a bug, the car might still hit an object.

 

[willow_hiller]

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

As far as I can see, there are no curbs leading to the alley. Both ends of it have curb cuts. Only part of it on the North end is wider and used as the exit from a Starbucks drive-through, so it doesn't really look like it was made for cars (note Google Street View stops after the drive-through). I do not see a "hard road edge in the alleyway" as described by Waymo, unless they're referring to the area immediately around the pole that is marked in yellow paint.

 

[diplomat33]

I do not see a "hard road edge in the alleyway" as described by Waymo, unless they're referring to the area immediately around the pole that is marked in yellow paint.

Correct. It was not a hard curb, it was just an area with yellow painted lines. That is the part that was not labeled in the HD map.

 

[willow_hiller]

Correct. It was not a hard curb, it was just an area with yellow painted lines. That is the part that was not labeled in the HD map.

I guess we'll need to wait to see the exact language that Waymo uses to describe how they've updated their maps when NHTSA publishes the recall, because at the moment that seems to be the Verge's words and not Waymo's.

I would not describe a painted area as a "hard road edge," unless whoever wrote that is also having trouble differentiating hard and soft objects.

 

[enemji]

Does this mean Waymo needs to hand code every curb and telephone pole?
If so, sounds difficult to scale.

I thought having LiDAR and RADAR would take care of it. I guess not

/end sarcasm

 

[JulienW]

I guess we'll need to wait to see the exact language that Waymo uses to describe how they've updated their maps when NHTSA publishes the recall, because at the moment that seems to be the Verge's words and not Waymo's.....

HD Maps are updated dynamically as Waymos drive around (or at lest the changes are recorded) and I bet any changes are checked over by a human team and then sent out to the fleet with a quick turnaround. This is more beneficial with pop up construction sites. So it is likely the HD Maps were updated after this incident.

 

[enemji]

Not for the telephone pole, sounds like they just need to classify it properly into a category of hard objects instead of soft objects. This is an example of how lidar doesn't automatically identify objects (as people like to claim), but rather, just like with cameras, you still need to properly classify them.

For the curbs, sounds like sometimes they do need to hand code things into the map. I'm surprised they don't have an automated curb detection system (if not on the vehicles themselves, at least in the mapping software). It's possible however that alleyway was not part of their maps.

Edit: from reading the article, sounds like there isn't actually a raised curb but rather there is just a yellow line to delineate the road boundary, while the road is at the same grade. If they were using lidar based mapping, depending on how that yellow line was laid (a tape or paint; if paint, what type and how thick/reflective vs the road surface) and the lidar used (is it the type that can read road lines) it's possible it can't be detected via automated mapping (at least not with current software, may need to switch to vision based mapping or a hybrid between the two).

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

 

[willow_hiller]

HD Maps are updated dynamically as Waymos drive around and I bet any changes are checked over by a human team and then sent out to the fleet with a quick turnaround. This is more beneficial with pop up construction sites. So it is likely the HD Maps were updated after this incident.

A couple of sources are making it sound like the update to the map was officially a part of the recall, so I want to see if they describe exactly what was added to the HD map once the full text of the recall is public.

The Verge: "The company is filing the recall with the National Highway Traffic Safety Administration (NHTSA) after completing a software update to 672 vehicles — the total number of driverless-capable vehicles in Waymo’s fleet. The update corrects an error in the software that “assigned a low damage score” to the telephone pole, and updates its map to account for the hard road edge in the alleyway that was not previously included. ... Waymo’s recall was deployed by the company’s engineers at the central depot where the vehicles return for regular maintenance and testing. It was not through an over-the-air software update, like some of Tesla’s recent recalls."

TechCrunch: “We went to work immediately and determined that, in certain situations, our vehicles had an insufficient ability to avoid collisions with on-road narrow, permanent objects within the drivable surface,” said Barna. “We have since implemented mapping and software updates.”

 

[Dewg]

The Waymo did identify the long vertical obstacle. But there was a bug in the software that told the car to ignore the obstacle.

That is why it is not as simple as if you are camera only, you will never hit objects or if you use lidar and radar, you will never hit objects. You can be vision-only or have lots of lidar and radar, if your NN has a bug, the car might still hit an object.

This is why I keep jumping up and down when people say vision-only is a failure on Tesla, and they need to add RADAR and LIDAR to their cars. As if RADAR and LIDAR are instant fixes that prevent collisions. It's about perception, and planning. In many cases, it's the planner that's failing.

 

[enemji]

Does this mean Waymo needs to hand code every curb and telephone pole?
If so, sounds difficult to scale.

This is what Waymo thought it was...

 

[Daniel in SD]

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

Try using FSD for millions of miles a year and report back. I suspect you’ll also find bugs.

 

[enemji]

Try using FSD for millions of miles a year and report back. I suspect you’ll also find bugs.

Sure. Meet me in 42 million years

 

[stopcrazypp]

This is what Waymo thought it was...


View attachment 1055774

I actually thought of those things as the only soft object that is long, vertical and of similar thickness. But it still doesn't make sense for the car to run into it. Maybe there needs to be a 3rd category (stuff like empty plastic bags that it's ok to run into, softer stuff that car should still avoid unless absolutely necessary, then hard items).

 

[MP3Mike]

It seems to me like this really should be listed as two separate recalls, as each of them independently should have prevented the collision:

• HD map: They didn't properly define drivable space. Had the map been correct it should have never even tried to take a path that would hit the pole.
• Object classification system: If the object classification system worked correctly, it wouldn't have tried to drive through the pole even if the HD map said it was OK.

But then there is at least one more system that should probably be recalled and fixed as well:

• The vision/perception system: It should have seen the painted lines and determined the proper drivable area, preventing the collision.

or

• The fusion system: Maybe perception got the drivable space correct, but the fusion system saw it was different than the HD map and decided the HD map was more likely to be correct, and followed the wrong one.

So, it seems like this single edge-case exposed many edges/failures in the Waymo driver system.

 

[Daniel in SD]

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?