Waymo

[powertoold]

He shouldn't say "rails", of course. Sure, they can localize very precisely. But it implies every Waymo follows the exact same path, down to the centimeter, and can't deviate from it. That's nonsense and Karpathy knows it.

It's obviously not a train

But the drivable space is predefined like tracks

 

[mycroftxxx]

Accelerometers, radar and gyroscopes are very accurate. I believe they are cm level accurate. And all those methods together would compensate for any "noise" from lidar.

Uh, no.

That said: high frequency “noise” being input into a control effector input (in this case, the front wheels, which are mechanically connected to the steering wheel) is undesirable but is not necessarily a big deal. As long as the noise is of insufficient power to cause premature wear of the system, it may not be worth solving. However, if it were my design, I’d want to fix it.

However, it’s also possible the jittery wheel isn’t ”real”; depending on where the automatic steering actuation force is being input into the steering system, the steering wheel may just be at the end of a not-terribly-torsionally-stiff linkage which is amplifying otherwise-negligible high frequency noise.

Hard to know without more specifics of the design, but my bet is on an unoptimized path control loop that is a nuisance at worst.

 

[bedoig]

I saw one of the Waymo trucks operating north of Houston last Saturday. I don't follow this closely, but seem to recall some kind of announcement regarding testing on I-45 between Houston and Dallas. Kinda cool.

 

[diplomat33]

90 minutes of fully autonomous driving in a Waymo, sped up 8x, from around the Nopa and Cole Valley neighborhoods, and also bop into Golden Gate Park.

 

[Bitdepth]

90 minutes of fully autonomous driving in a Waymo, sped up 8x, from around the Nopa and Cole Valley neighborhoods, and also bop into Golden Gate Park.

It never gets old watching an empty car just driving around a city like San Francisco. When they open it up for public rides i will make a road trip there to test it myself.

 

[diplomat33]

It never gets old watching an empty car just driving around a city like San Francisco. When they open it up for public rides i will make a road trip there to test it myself.

Please do and give us your report.

 

[DrChaos]

Pretty close.

I would not call HD maps a virtual reality world. They are simpler than that. HD maps are a 2D representation of the static driving area. HD maps can also include information like speed limits, one way streets, no parking zones between certain hours etc.. Yes, the Waymo uses real-time scans with cameras, lidar and radar to detect any differences and adjust.

The real-time scans with cameras, lidar and radar also detect moving objects like other vehicles, pedestrians, cyclists etc... The Waymo takes information from the HD maps and the real-time scans to build a complete perception of everything around it, both the static and the dynamic. The Waymo then makes predictions about what other objects will do, like other vehicles, pedestrians etc...

Finally, the Waymo takes the map, the real-time scans and the predictions and makes a decision about what path to drive in order to drive safely. The car then tells the steering wheel and pedals what to do to execute the path.

The distinguishing technical questions are: "is a pre-computed mapping of the environment to ~cm level in 3d necessary for Waymo cars to function? Is a similar resolution direct physical measurement hardware at drive time, interacting with that mapping data, necessary to operate?"

I don't know what it is now, but once upon a time there had to be something like it because it came out of the 2007 DARPA challenge when there wasn't remotely either the computer science algorithmic knowledge or the computational hardware capability to do modern machine learning based vision processing.
I guess sort of similar to TERCOM used by older cruise missiles pre-GPS, correlations vs pre-mapped terrain.

Everything else is required tasks of any ADAS and semantic tags are different from cm level pre-mapping.

Now, Waymo might not need that now and surely does modern ML on vision, but could it operate successfully with low resolution maps of streets only (no buildings/overpasses/terrain mapped) ?

 

[diplomat33]

The distinguishing technical questions are: "is a pre-computed mapping of the environment to ~cm level in 3d necessary for Waymo cars to function? Is a similar resolution direct physical measurement hardware at drive time, interacting with that mapping data, necessary to operate?"

I don't know what it is now, but once upon a time there had to be something like it because it came out of the 2007 DARPA challenge when there wasn't remotely either the computer science algorithmic knowledge or the computational hardware capability to do modern machine learning based vision processing.
I guess sort of similar to TERCOM used by older cruise missiles pre-GPS, correlations vs pre-mapped terrain.

Everything else is required tasks of any ADAS and semantic tags are different from cm level pre-mapping.

Now, Waymo might not need that now and surely does modern ML on vision, but could it operate successfully with low resolution maps of streets only (no buildings/overpasses/terrain mapped) ?

Short version: Waymo does not require HD maps to operate. But HD maps are cheap, easy and add safety. So Waymo sees no reason not to use HD maps.

https://twitter.com/x/status/1561506448340910082

Anguelov has explained that Waymo uses HD maps "as a prior, not as an immutable truth." (direct quote). Waymo relies primarily on their rich suite of cameras, lidar and radar to drive. But he also explained that HD maps add safety because they help the car anticipate parts of the road that are hidden from view. Additionally, Waymo cars can share map changes with each other, sometimes in real-time, to help the fleet anticipate something it may encounter later in the drive. Those are the two main safety benefits of HD maps according to Anguelov.

 

[AlanSubie4Life]

90 minutes of fully autonomous driving in a Waymo, sped up 8x, from around the Nopa and Cole Valley neighborhoods, and also bop into Golden Gate Park.

Are they trying to win a competition for worst camera angle?

 

[diplomat33]

Are they trying to win a competition for worst camera angle?

Yeah the camera angle is bad. The person who set up the camera did not lower it enough.

 

[diplomat33]

In Waymo news, Waymo hired a new CFO.

https://twitter.com/x/status/1573406009242140690

 

[diplomat33]

According to Bloomberg Intelligence, Waymo could generate about $5B in revenue by 2025:

 

[diplomat33]

Waymo published some new data that shows Waymo Driver outperforms attentive human drivers in avoiding fatal collisions:

https://twitter.com/x/status/1575516589427941376

https://twitter.com/x/status/1575516594330927104

https://twitter.com/x/status/1575516598563127300

https://twitter.com/x/status/1575516688186978304

 

[powertoold]

Great, we all love simulated results

 

[diplomat33]

I would note that Waymo is not claiming the simulation results prove safety greater than humans. In the blog, Waymo says that the purpose of the research is to offer a benchmark to help measure AV safety. So, other AV companies could compare their AV with the NIEON model to see how good their AV is at taking evasive action to avoid or mitigate crashes. Waymo uses other methodologies for measuring safety but this NIEON model will be very useful as a way to measure one important safety metric. NIEON is a great model because a non-impaired, attentive human driver is a good standard to compare your AV to.

I think Waymo's simulation results are very encouraging. But simulation is not perfect. You still need to compare the simulation results with the real world. But if the Waymo simulation results are confirmed by real world data, then it would be a very good indicator of Waymo's safety.

 

[ThomasD]

Are Waymos programed to respond to Horns and Train lights so these kinds of collisions will never happen?

 

[diplomat33]

Are Waymos programed to respond to Horns and Train lights so these kinds of collisions will never happen?

Yes, Waymo is able to respond to horns and train lights. Waymo has external microphones that are able to detect sounds like horns and emergency vehicle sirens and the Waymo can respond. Waymo can also detect trains with cameras, lidar and radar and would avoid a collision.

 

[FloridaJohn]

According to Bloomberg Intelligence, Waymo could generate about $5B in revenue by 2025:

I remember when 2025 was way off in the future.

 

[ThomasD]

I would just like to ride in one. Wonder how many years it will be until they are in every city in every state?

 

[Bitdepth]

Great, we all love simulated results

Yes? Of course, is there any tech industry that does not rely on simulation? Space sector relies on simulation heavily; Automotive sector relies on simulation heavily for aerodynamics. In manufacturing, the systems are simulated before it is deployed. In the medical sector, drug interactions are simulated and of course in autonomy the cars a trained in a simulator before they are deployed on the road.