Waymo

[enemji]

I see you are another Waymo doubter who thinks they will fail. Well, good luck with that. The Waymo doubters are like the TSLA shorts who are convinced Tesla will go bankrupt any day now. You keep waiting for Waymo to fail and they keep expanding more and more. If you are betting on Waymo failing, you will lose. Mark my words.

At the same time, don’t mistake me for a Tesla fanboy who thinks Tesla FSD is going to work.

Waymo will work but they will have to chuck the HD maps scenario as they move further along. The HD maps idea is from someone who has no idea of real world.

As the youtube example, on the left is the HD Maps and on the right is the Real-time lidar+Radar+Camera+etc. They are going to have this misalignment all the time.

 

[diplomat33]

Waymo will work but they will have to chuck the HD maps scenario as they move further along.

Well, I don't agree. Waymo says the maps are not the bottleneck. They say HD maps are cheap, scalable and provide safety benefits. So, I don't see any reason why they would ditch HD maps.

 

[enemji]

Well, I don't agree. Waymo says the maps are not the bottleneck. They say HD maps are cheap, scalable and provide safety benefits. So, I don't see any reason why they would ditch HD maps.

Time will tell

That said, I definitely look forward to the evolution of AV. It certainly is not as simple as following traffic and street signs as someone (Elon) thought it would be.

 

[FloridaJohn]

AV needs to perform as good as a human. A human driver scans the environment and makes decisions in real time.

As an actual human driver, I, too, rely on an internal HD map.

I know which of the many available routes to work is the most efficient and reliable. I know which lane to be in for each turn, and importantly, when to get in those lanes. I know what the traffic looks like when things are going smoothly, and also when they aren’t.

I do not figure out my route to work each day. I am also familiar with many routes to other places I frequent. Imagine if your autonomous car had all this “local” information for every route in the country. It would make a great robot chauffeur, I would imagine.

 

[enemji]

As an actual human driver, I, too, rely on an internal HD map.

I know which of the many available routes to work is the most efficient and reliable. I know which lane to be in for each turn, and importantly, when to get in those lanes. I know what the traffic looks like when things are going smoothly, and also when they aren’t.

I do not figure out my route to work each day. I am also familiar with many routes to other places I frequent. Imagine if your autonomous car had all this “local” information for every route in the country. It would make a great robot chauffeur, I would imagine.

No doubt about it. I agree with you and yet I will say that that is not the future of AV. This will definitely be used for a certain period of time as the technology transitions to the “True FSD” ©

 

[diplomat33]

Waymo CEO provided a thread with some updates:

https://twitter.com/x/status/1569402635803066368

https://twitter.com/x/status/1569402999063322624

Interesting tidbit: the recent driverless video with Dolgov was a new software version that shipped a few weeks ago that included their next-gen ML planner:

https://twitter.com/x/status/1569403184770330624

 

[DrChaos]

Waymo builds a HD map first to give the car a prior about the road and traffic rules. For example, HD maps allow the car to know what is up ahead, beyond visual range and see the road when it is occluded. HD maps also allow the car to know about local driving rules. The Waymo can check the HD map but it relies on cameras, radar and lidar to do the actual driving. So Waymo does not need HD maps to drive. It can drive based on just cameras, lidar and radar. But HD maps have helpful information and improve safety.



Your contention is wrong. HD maps improve safety. Without HD maps, the car would have to just figure stuff out again and again and guess what it can't see which would not be reliable. With HD maps, you will have higher reliability which is needed for true FSD, ie driverless.



Localization does happen in real-time. Like I said above, localization uses accelerometers, gyroscopes, wheels speeds, as well as lidar and radar to make real-time measurements of the vehicle's speed and heading. You seem to be confusing localization with mapping. Those are different things.

Semantic maps are a great idea and necessary, but those are not "HD" maps. Those are enriched LD maps.

I think the Waymo style HD maps were in fact used for positioning & localization---whether or not that's true now we don't know. The company came out of the 2007 DARPA challenge with lots of lidar, and well before computational capacity to do on-board neural network image recognition was feasible.

I think semantic maps, i.e. including information beyond observable geometry about what drivers actually do is very important. Tesla conceivably could be acquiring this, at least in well traveled areas, using their fleet driving in human mode. I worry it's an Elonist problem who decreed "no maps ever" that's holding it back, as it's an obvious approach. It would take a 'big data' analysis by averaging where human cars go when they want to get to some future state, and continually updating that to the fleet, whether in bulk or on-the-fly, and a major nnet retrain where this data becomes part of the driving policy, and even perception (easy to identify lanes & driveable surfaces with fleet hints, though that's not generally a problem now).

If they had more of a ML based driving policy they could train that against safe human behavior to become more natural. Nothing is easy, it's a few years of work no doubt.

 

[diplomat33]

Semantic maps are a great idea and necessary, but those are not "HD" maps. Those are enriched LD maps.

I think the Waymo style HD maps were in fact used for positioning & localization---whether or not that's true now we don't know. The company came out of the 2007 DARPA challenge with lots of lidar, and well before computational capacity to do on-board neural network image recognition was feasible.

I think semantic maps, i.e. including information beyond observable geometry about what drivers actually do is very important. Tesla conceivably could be acquiring this, at least in well traveled areas, using their fleet driving in human mode. I worry it's an Elonist problem who decreed "no maps ever" that's holding it back, as it's an obvious approach. It would take a 'big data' analysis by averaging where human cars go when they want to get to some future state, and continually updating that to the fleet, whether in bulk or on-the-fly, and a major nnet retrain where this data becomes part of the driving policy, and even perception (easy to identify lanes & driveable surfaces with fleet hints, though that's not generally a problem now).

If they had more of a ML based driving policy they could train that against safe human behavior to become more natural. Nothing is easy, it's a few years of work no doubt.

Waymo HD maps include rich semantic info, yes. They are built with lidar so they are very accurate (cm level accuracy) but they also carry rich semantic information as well.

 

[enemji]

Waymo HD maps include rich semantic info, yes. They are built with lidar so they are very accurate (cm level accuracy) but they also carry rich semantic information as well.

How does the localization work? Is it GPS based or based on the HD maps dimensional information of the infrastructure as scanned?

 

[diplomat33]

How does the localization work? Is it GPS based or based on the HD maps dimensional information of the infrastructure as scanned?

I explained this before. Waymo uses a combination of different measurements. Specifically, Waymo uses accelerometers, gyroscopes with input from GPS, maps, wheel speeds and lidar and radar to measure the precise position, velocity and heading of the vehicle in real-time:

"This module uses accelerometers and gyroscopes with input from GPS, maps, wheels speeds, and laser and radar measurements to provide highly accurate position, velocity, and heading information to the vehicle." Safety Report, page 14

 

[diplomat33]

38 mn driverless Waymo ride at normal 1x speed:

Not perfect. At 5:15 mn, the car seems to "wobble" a bit when moving to the left turn lane. But overall, it is a safe and reliable drive IMO. It handles a lot of different scenarios really well IMO.

 

[Doggydogworld]

38 mn driverless Waymo ride at normal 1x speed:

Not perfect. At 5:15 mn, the car seems to "wobble" a bit when moving to the left turn lane. But overall, it is a safe and reliable drive IMO. It handles a lot of different scenarios really well IMO.

That was a bit odd, but I was more bothered by the wheel twitching as he got into the car. It slowed down at 2:30 due to mist from the guy washing trash bins. I was a little surprised it kept driving when he unbuckled his seat belt at 3:50, though maybe it's legal to be unbelted in the back seat of a taxi.

Not sure I'll watch much of this -- Waymo rides are too boring

 

[diplomat33]

That was a bit odd, but I was more bothered by the wheel twitching as he got into the car.

Maybe I should ask Waymo but I wonder if maybe the car "wobbling" at 5:15 was on purpose to give the sensors a quick peek around the car to see incoming traffic before making the turn? Otherwise, yeah, it is odd. Waymo knows where the lanes are so there is no reason it would be uncertain about which lane to be in.

It slowed down at 2:30 due to mist from the guy washing trash bins.

That was interesting.

I was a little surprised it kept driving when he unbuckled his seat belt at 3:50, though maybe it's legal to be unbelted in the back seat of a taxi.

Maybe because they are a passenger, they can unbuckle. But I did not think it was legal.

Not sure I'll watch much of this -- Waymo rides are too boring

Well, they say the metric of good autonomous driving is when it gets boring.

 

[MP3Mike]

FSD Beta 10.69.1.1 in light to moderate rain.

is this the comedy you speak of? what happened to trillions of miles of rain data?

FSD Beta 10.69.2 seems to do just fine in heavy rain, it wasn't until he go on the highway and it switched back to NoA that he had to take over.

 

[diplomat33]

FSD Beta 10.69.2 seems to do just fine in heavy rain, it wasn't until he go on the highway and it switched back to NoA that he had to take over.

He gets the "FSD may be degraded" alert almost immediately. So the system detected that FSD reliability was degraded.

 

[Supcom]

He gets the "FSD is degraded" alert almost immediately. So FSD performance was degraded.

Wrong, Dip. It says it "may" be degraded. Read the image you posted. It's quite clear. I saw nothing in the video that indicated any degredation. If you did, please let us know. Mr. Cook certainly did not mention any beta issues.

I've driven FSD beta many times in the rain and have seen this notice a lot, but always wonder what is actually degraded as I have never noticed any difference in how it drives.

 

[diplomat33]

Wrong, Dip. It says it "may" be degraded. Read the image you posted. It's quite clear. I saw nothing in the video that indicated any degredation. If you did, please let us know. Mr. Cook certainly did not mention any beta issues.

I've driven FSD beta many times in the rain and have seen this notice a lot, but always wonder what is actually degraded as I have never noticed any difference in how it drives.

Yes, it says "may be degraded". And no, I did not notice any actual degradation but the alert is letting the driver know that degradation is a possibility. Just because there was no apparent degradation in the video, does not mean there wasn't a drop in reliability in the system or that degradation can't happen. I always disengage when I get that message out of caution because you cannot trust that degradation won't happen.

 

[powertoold]

I don't think Waymo will ever be able to fix the jittery wheel problem. It's a consequence of noise from lidar localization. We've seen it 2 years ago, and now it's still there with the 5g cars.

The wheel is already jittery from localizing during clear weather: imagine the noise from any amount of rain on the sensors.

 

[diplomat33]

I don't think Waymo will ever be able to fix the jittery wheel problem.

That's pretty arrogant. Lots of problems seem difficult and they eventually get solved. Never say never.

And the "jittery wheel" clearly does not impact the driving at all so it is not a big problem IMO.

It's a consequence of noise from lidar localization. We've seen it 2 years ago, and now it's still there with the 5g cars.

We know Waymo does not use lidar-only for localization. Waymo uses several other methods for localization, specifically accelerometers, GPS, maps, gyroscopes and wheel speeds. None of those other methods would be affected by any "lidar noise". So that cannot be the answer. And you have no proof the "jittery wheel" is even related to localization. You are just making stuff up as always.

 

[EVNow]

Yes, it says "may be degraded". And no, I did not notice any actual degradation but the alert is letting the driver know that degradation is a possibility. Just because there was no apparent degradation in the video, does not mean there wasn't a drop in reliability in the system or that degradation can't happen. I always disengage when I get that message out of caution because you cannot trust that degradation won't happen.

Happens often here - but usually continues to work. When it can no longer work, switches to AP. No need to disengage.