Autonomous Car Progress

[diplomat33]

Simulation City is truly mind blowing.
No game engines here. 100% ML.
Waymo should come out and say that anyone messing with game engines is DOOMED! (lol tesla).
Waymo playing chess, tesla fiddling with checkers.

Yes, it is impressive. But it is not just about creating a photorealistic environment. Anguelov also talks about the need to have realistic sim agents that match the behavior in the real world as closely as possible. If the sim agents don't behave like in the real world, then the sim won't be very effective at training your AV to handle real world scenarios. So, Waymo is working to create sim agents with realistic behavior using imitation learning!

On the last part of the presentation, Anguelov teases some future work that Waymo is doing there:

If I understood him correctly, Waymo is using imitation learning to model their sim agents to behave as closely as possible to real world agents and then using the simulation to train and test their AV model. The AV models can then be tested in the real world, and data can be collected to validate the models. Rinse and repeat until you get the improvement you need.

What do you think the benefits of this approach will be? It seems like it will help make their sim agents behave more realistically and will make the simulation more effective at training their AV models. The net result should be the AV will handle driving scenarios more naturally. Did I get that right?

Also, if Waymo can pull off realistic simulation with realistic sim agents at scale, it should be a true game changer IMO. In theory, it should drastically accelerate the pace towards "solving L5".

 

[rxlawdude]

Yes, it is impressive. But it is not just about creating a photorealistic environment. Anguelov also talks about the need to have realistic sim agents that match the behavior in the real world as closely as possible. If the sim agents don't behave like in the real world, then the sim won't be very effective at training your AV to handle real world scenarios. So, Waymo is working to create sim agents with realistic behavior using imitation learning!

On the last part of the presentation, Anguelov teases some future work that Waymo is doing there:

If I understood him correctly, Waymo is using imitation learning to model their sim agents to behave as closely as possible to real world agents and then using the simulation to train and test their AV model. The AV models can then be tested in the real world, and data can be collected to validate the models. Rinse and repeat until you get the improvement you need.

What do you think the benefits of this approach will be? It seems like it will help make their sim agents behave more realistically and will make the simulation more effective at training their AV models. The net result should be the AV will handle driving scenarios more naturally. Did I get that right?

Also, if Waymo can pull off realistic simulation with realistic sim agents at scale, it should be a true game changer IMO. In theory, it should drastically accelerate the pace towards "solving L5".

But can it avoid traffic cones? "Imitation learning" within a circumscribed area.

You give Waymo every cheer you can, and every raspberry you can to Tesla's FSD (very different beasts). But please carry on as the Waymo Diplomat.

 

[EVNow]

Yes, it is impressive. But it is not just about creating a photorealistic environment. Anguelov also talks about the need to have realistic sim agents that match the behavior in the real world as closely as possible. If the sim agents don't behave like in the real world, then the sim won't be very effective at training your AV to handle real world scenarios.

I thought this was well understood - discussed even 2 or 3 years back here.

The basic argument Musk has been making is to get simulation to really be representative of real life - you need to understand and model real life so well that by then you could have solved FSD, anyway. You can the old Tesla FSD day video.

 

[diplomat33]

But can it avoid traffic cones? "Imitation learning" within a circumscribed area.

You give Waymo every cheer you can, and every raspberry you can to Tesla's FSD (very different beasts). But please carry on as the Waymo Diplomat.

Yes, Waymo can avoid cones. The cone incident was one problem months ago. It is not indicative of Waymo's overall state of FSD or progress. And Waymo is doing imitation learning at scale with data all over the US, not just in one geofenced area. Stop your ridiculous trolling.

The fact is that Waymo is running the only fully autonomous ride-hailing service in the US. Waymo has the most expertise and has demonstrated the best autonomous driving. Put simply, Waymo is the current boss of autonomous driving. I am in the habit of cheering the best and that's Waymo right now.

I give Tesla's FSD Beta raspberries only when it deserves it. When FSD Beta does something well, I am happy to praise it. But sadly, Tesla has not demonstrated reliable autonomous driving yet. I hope they do. And when they do, I will be happy to cheer them.

 

[rxlawdude]

Yes, Waymo can avoid cones. The cone incident was one problem months ago. It is not indicative of Waymo's overall state of FSD or progress. And Waymo is doing imitation learning at scale with data all over the US, not just in one geofenced area. Stop your ridiculous trolling.

The fact is that Waymo is running the only fully autonomous ride-hailing service in the US. Waymo has the most expertise and has demonstrated the best autonomous driving. Put simply, Waymo is the current boss of autonomous driving. I am in the habit of cheering the best and that's Waymo right now.

I give Tesla's FSD Beta raspberries only when it deserves it. When FSD Beta does something well, I am happy to praise it. But sadly, Tesla has not demonstrated reliable autonomous driving yet. I hope they do. And when they do, I will be happy to cheer them.

I assume you are not trying to be a beta tester of FSD?

 

[diplomat33]

I assume you are not trying to be a beta tester of FSD?

Of course I am. I bought a Tesla to have the latest driver assist tech. I paid for FSD. And I love testing the latest tech. And as a tester, I can hopefully send data back that will help Tesla improve FSD. I understand that it is not autonomous. But that does not mean that I hate it or want Tesla's FSD to fail. Being a fan of Waymo does not mean that I am against Tesla or want Tesla's FSD to fail.

 

[diplomat33]

During GM Investor Day, Cruise provided an update on their progress and plans:

At 9:37, there are clips of autonomous driving in SF.

Origin robotaxi is driving autonomously in simulations. Cruise expects it will be validated almost entirely through simulations.

Economics:

Roadmap:

 

[mhan00]

During GM Investor Day, Cruise provided an update on their progress and plans:

At 9:37, there are clips of autonomous driving in SF.

Origin robotaxi is driving autonomously in simulations. Cruise expects it will be validated almost entirely through simulations.

Economics:

Roadmap:

Didn’t watch the vid yet, but I’d like to know where they’re getting 4 bucks per mile for drivers. I drove Uber for a few months a few years ago for fun, and the rates were no where close to that in Los Angeles. Base rate was sub-1 dollar per mile, and that was only for time when passengers were actually inside the car, there was no compensation for the time waiting or the miles driven to reach passengers. I have a hard time believing they’d charging be nearly 4x in SF. I’d be interested to know what other expenses drivers were incurring to ride share apps since they don’t provide any benefits. Is it all just bonuses to try to attract new drivers to replace people dropping out because they realize they make almost nothing?

 

[diplomat33]

Didn’t watch the vid yet, but I’d like to know where they’re getting 4 bucks per mile for drivers. I drove Uber for a few months a few years ago for fun, and the rates were no where close to that in Los Angeles. Base rate was sub-1 dollar per mile, and that was only for time when passengers were actually inside the car, there was no compensation for the time waiting or the miles driven to reach passengers. I have a hard time believing they’d charging be nearly 4x in SF. I’d be interested to know what other expenses drivers were incurring to ride share apps since they don’t provide any benefits. Is it all just bonuses to try to attract new drivers to replace people dropping out because they realize they make almost nothing?

It is an estimate. He says a person pays on average $5/mile in SF, "probably more". He then says "I bet most of it goes to the driver", estimates that the driver takes about 75-80% of the price for themselves and rest goes to the company. So the $4/mi to the driver is an estimate based on assuming that the driver takes about 75-80% of a $5/mi ride.

 

[Daniel in SD]

Didn’t watch the vid yet, but I’d like to know where they’re getting 4 bucks per mile for drivers. I drove Uber for a few months a few years ago for fun, and the rates were no where close to that in Los Angeles. Base rate was sub-1 dollar per mile, and that was only for time when passengers were actually inside the car, there was no compensation for the time waiting or the miles driven to reach passengers. I have a hard time believing they’d charging be nearly 4x in SF. I’d be interested to know what other expenses drivers were incurring to ride share apps since they don’t provide any benefits. Is it all just bonuses to try to attract new drivers to replace people dropping out because they realize they make almost nothing?

This is only a 6 mile drive (that takes 25 minutes!). Most of the cost is per-minute which works out to $23 per hour.

 

[Bitdepth]

Keeping with the topic, some videos of Waymo's 5th gen Lidar from yesterday's keynote by Drago Anguelov

 

[diplomat33]

Keeping with the topic, some videos of Waymo's 5th gen Lidar from yesterday's keynote by the Dmitri Dolgov

The keynote was by Anguelov, not Dolgov. But great clips.

 

[Bitdepth]

The keynote was by Anguelov, not Dolgov. But great clips.

Darn, you caught me right as i was fixing my error.

 

[EVNow]

Economics:

Wonder what is the volume needed to achieve those economics. I guess he didn't mention it ?

 

[powertoold]

Just food for thought: if Tesla spent 3 months focusing on SF alone (collect tons of SF fleet data, building their entire road / lane geometry NNs based on SF data, fine tune lane semantics and intersection types, etc.), would Tesla's FSD be better than Waymo and Cruise in SF?

You obviously know my answer.

Waymo, Cruise, you name it, are wayy behind Tesla. The difference is shocking. Many don't see it.

 

[Daniel in SD]

Just food for thought: if Tesla spent 3 months focusing on SF alone (collect tons of SF fleet data, building their entire road / lane geometry NNs based on SF data, fine tune lane semantics and intersection types, etc.), would Tesla's FSD be better than Waymo and Cruise in SF?

You obviously know my answer.

Waymo, Cruise, you name it, are wayy behind Tesla. The difference is shocking. Many don't see it.

Maybe they could be on par...
if they mapped it after evacuating the city so there would be no moving vehicles, pedestrians, cyclists, or scooters. They'd probably have to restrict it to a time of day when the cameras weren't blinded too...

 

[powertoold]

Maybe they could be on par...
if they mapped it after evacuating the city so there would be no moving vehicles, pedestrians, cyclists, or scooters. They'd probably have to restrict it to a time of day when the cameras weren't blinded too...

What really bugs me about the other fsd developers is that they hide their true colors and distort public perception. In the recent Waymo and Cruise videos, they show little highlights here and there. If they really wanted to show videos, why not post full length rides, showing us all the times remote assistance took over or augmented the fsd, all the disengagements and/or interventions, etc.

We have to take a step back and appreciate that Tesla is the ONLY company showing their true colors when it comes to fsd.

 

[FloridaJohn]

What really bugs me about the other fsd developers is that they hide their true colors and distort public perception. In the recent Waymo and Cruise videos, they show little highlights here and there. If they really wanted to show videos, why not post full length rides, showing us all the times remote assistance took over or augmented the fsd, all the disengagements and/or interventions, etc.

We have to take a step back and appreciate that Tesla is the ONLY company showing their true colors when it comes to fsd.

You mean videos that you would immediately dismiss out of hand as “marketing” videos?

 

[EVNow]

Maybe they could be on par...
if they mapped it after evacuating the city so there would be no moving vehicles, pedestrians, cyclists, or scooters. They'd probably have to restrict it to a time of day when the cameras weren't blinded too...

They have to just write special planning rules for S.F. and oversample S.F. videos. Esp. that small area Waymo is covering now.

 

[EVNow]

You mean videos that you would immediately dismiss out of hand as “marketing” videos?

Not if they are not heavily "curated".