Autonomous Car Progress

[Bladerskb]

Xpeng's City NGP - 20 minutes of Door to Door Automated Driving in China (L2)

 

[diplomat33]

From Waymo blog:

We also continue exploring the use of applied machine learning techniques to help reach new frontiers in perception by creating more efficient and accurate detection models for 3D objects at greater distances, utilizing task relations at scale to concurrently and dramatically improve performance and reduce the need for labeled data, and more. These new advances allow us to unlock the full potential of our new sensors, reduce the amount of compute perception uses, and deepen our understanding of other objects on the road.

This section stood out to me. It sounds like Waymo is continuing to explore new ML techniques to make perception even better. They are always looking for ways to make perception even more accurate and also more efficient. Specifically, they are using "task relations" to make computer vision more efficient, reduce the need for labeled data, and improve scene understanding.

Their paper on task relations says this:

Many computer vision tasks address the problem of scene understanding and are naturally interrelated e.g. object classification, detection, scene segmentation, depth estimation, etc. We show that we can leverage the inherent relationships among collections of tasks, as they are trained jointly, supervising each other through their known relationships via consistency losses. Furthermore, explicitly utilizing the relationships between tasks allows improving their performance while dramatically reducing the need for labeled data, and allows training with additional unsupervised or simulated data. We demonstrate a distributed joint training algorithm with task-level parallelism, which affords a high degree of asynchronicity and robustness. This allows learning across multiple tasks, or with large amounts of input data, at scale. We demonstrate our framework on subsets of the following collection of tasks: depth and normal prediction, semantic segmentation, 3D motion and ego-motion estimation, and object tracking and 3D detection in point clouds. We observe improved performance across these tasks, especially in the low-label regime.

Sounds like task relations is where you train on multiple tasks at the same time leveraging the relationship between the tasks? Would an example of task relation be depth perception measures distance to a car in front of you, object detection can also estimate distance, so you can train both at the same?

Does that sound right @Bladerskb ?

 

[Tam]

Stock goes up because of the Hertz order but interestingly "no self-driving options included" in this deal.

Could it be because Tesla FSD is not intuitive enough for an average driver and Hertz does not want that risk?

https://twitter.com/x/status/1453423792592674817

 

[EVNow]

Could it be because Tesla FSD is not intuitive enough for an average driver and Hertz does not want that risk?

Its because Tesla FSD is not yet available. Its also the low end base Model 3.

 

[Tam]

Its because Tesla FSD is not yet available.

Standard widely public release FSD version has been released which has been, for years, functioned as Enhanced Autopilot:

Autopark
Dumb/simple Summon
Auto Lane Change
Navigate on Autopilot

The first 2 features have a history of minor collisions and the current smart summon is not that smart at preventing collisions either. These features are not mature enough for reliability.

Its also the low end base Model 3.

True. But there would be curious rental customers and Uber drivers who might want to pay more just to try out what FSD is like.

If FSD is reliable then the cost is a moot point because the monthly rental is only $199 monthly which is close to nothing for Hertz.

 

[EVNow]

Standard widely public release FSD version has been released which has been, for years, functioned as Enhanced Autopilot for years:

Autopark
Dumb/simple Summon
Auto Lane Change
Navigate on Autopilot

Not the same as FSD (which is City NOA now). AP, Highway NOA and autopark might be useful for renters. I expect those to be offered sometime down the line.

If FSD is reliable then the cost is a moot point because the monthly rental is only $199 monthly which is close to nothing for Hertz.

When true FSD is reliable it won't be $199 a month.

Fairly sure in the future Tesla will offer daily rates to renters.

 

[glide]

Its because Tesla FSD is not yet available. Its also the low end base Model 3.

Really? FSD is not available? Because Tesla sure sells it like it is.

Talk about trying to have your cake and eat it too…

FSD is an available option. Hertz isn’t buying it because it’s overpriced, unproven, and comes with so much liability that it’s not worth it.

 

[EVNow]

Really FSD is not available? Because Tesla sure sells it like it is.

* Available later this next year.

 

[glide]

* Available later this next year.

So available when Hertz takes delivery. Why wouldn’t they want such an amazing piece of technology to add to their fleet? I mean…if it truly is safer than the average human driver, and somehow force-multiplies the value of your purchase, it seems like a no brainer.

Unless Tesla is lying again just to drum up sales. Which they would, of course, NEVER do…

 

[ThomasD]

Uber drivers will get a whole dollar per ride. I'm going to assume it will cost around 400 dollars a week to rent.

 

[diplomat33]

Cruise begins autonomous deliveries for Walmart in Scottsdale, AZ. If you are eligible for a self-driving delivery, you get a notification on your phone app. Walmart loads up the Cruise AV Chevy Bolt with your order, the car drives to your destination and you use the phone app to unlock the car and get your order.

https://twitter.com/x/status/1455203117562167296

 

[EVNow]

Scottsdale, AZ.

What does Scottsdale, AZ have that we don't ?

ps : One thing interesting about robotaxis is that because of operational reasons - they will always be geofenced ...

 

[mark95476]

..CEOs of Hertz and Uber talking about their orders and plans, respectively, is actually Tesla lying about their sales according to TSLAQ trolls..

Unless Tesla is lying again just to drum up sales. Which they would, of course, NEVER do…

 

[glide]

..CEOs of Hertz and Uber talking about their orders and plans, respectively, is actually Tesla lying about their sales according to TSLAQ trolls..

View attachment 728204

Show me where it says Hertz is purchasing the revolutionary FSD package for their fleet.

Because that’s what we were talking about. Way to quote out of context.

 

[diplomat33]

ps : One thing interesting about robotaxis is that because of operational reasons - they will always be geofenced ...

That's not necessarily a bad thing. The purpose of a robotaxi is to pick up passengers and drop them off. People who use a taxi, robotaxi or regular taxi, usually stay pretty local, going to the store, running an errand across town. People don't take a taxi to do a long road trip. So being geofenced fits the purpose of robotaxis. They don't need to be L5. Furthermore, L5 does not exist yet. So L4 (geofenced) is the best we have so far.

Also, keep in mind that geofences are limited right now only because the technology is still a work in progress and AV companies want to limit the risk that the AV drives in an area that it can't handle. But geofences will eventually get bigger and bigger as autonomous driving matures and becomes safer and more reliable.

I know Tesla fans use "geofence" like it is a bad word to imply that the autonomous driving is inferior, but it's not. It just depends on the purpose of your AV. Obviously, geofencing is probably undesirable for a consumer car since consumers generally want to use their car everywhere. You probably would not want to buy a car that can only drive autonomously in say a 50 square mile area. But if the purpose of the AV is to give rides to people who only need to go short distances, then there is nothing bad about being geofenced. I think Tesla fans make a mistake of assuming that consumer cars are a superior form of autonomous driving than geofenced robotaxis. There will be lots of people in cities who will want to use a geofenced robotaxi. And there will be people who want an autonomous consumer car. They are different products to fill different needs. There will be lots of different applications of autonomous driving (ride-hailing, delivery trucking, consumer cars etc) to fill different needs.

 

[EVNow]

I know Tesla fans use "geofence" like it is a bad word to imply that the autonomous driving is inferior, but it's not.

I use it to indicate that FSD has not yet been "solved". Geofencing makes it difficult to ascertain the progress made towards true autonomous driving.

 

[EVNow]

So available when Hertz takes delivery. Why wouldn’t they want such an amazing piece of technology to add to their fleet? I mean…if it truly is safer than the average human driver, and somehow force-multiplies the value of your purchase, it seems like a no brainer.

Unless Tesla is lying again just to drum up sales. Which they would, of course, NEVER do…

Not sure why you are mixing up two different things. "Coming later this year" was an old text in the order page. 100k/150k order was definitely not placed using that order page

As to drumming up sales - well, apparently only 10% or less of Tesla cars even buy FSD. And Tesla has several quarters worth of orders ...

 

[glide]

Not sure why you are mixing up two different things. "Coming later this year" was an old text in the order page. 100k/150k order was definitely not placed using that order page

As to drumming up sales - well, apparently only 10% or less of Tesla cars even buy FSD. And Tesla has several quarters worth of orders ...

Not sure why you are mixing up the fact that I said FSD is con used by Tesla to try and pad sales numbers. Sell something for $10K that does next to nothing it advertises.

The low consumer take-rate and the fact Hertz is not purchasing it for their fleet illustrates people are wising up. Hertz recognizes the investment is not worthwhile and they are not going to leverage it to robo-taxi their customers around as some fanbois here claim.

That’s all I said.

 

[rxlawdude]

Not sure why you are mixing up the fact that I said FSD is con used by Tesla to try and pad sales numbers. Sell something for $10K that does next to nothing it advertises.

The low consumer take-rate and the fact Hertz is not purchasing it for their fleet illustrates people are wising up. Hertz recognizes the investment is not worthwhile and they are not going to leverage it to robo-taxi their customers around as some fanbois here claim.

That’s all I said.

You said FSD is a con. Sounds like a statement of fact, rather than what it really is: some Tesla shortie's absurdly biased opinion.

 

[glide]

You said FSD is a con. Sounds like a statement of fact, rather than what it really is: some Tesla shortie's absurdly biased opinion.

FSD is a con. I absolutely stand by that.

The software doesn’t do half of what is promised today. Let alone what was promised when they originally started selling it. Once again, all I need to back that up is the take rate and a growing number of angry early adopters.