Cybertruck reportedly left out of Unsupervised FSD

[Sjohnson20]

June is a month away I highly doubt anyone other than internal Tesla employees will have unsupervised FSD Cybertruck access. Maybe June 2026 customers will have it.

Sponsored

 

[SCTesla]

I would think that 99% of the FSD data needed would be applicable for all Tesla passenger vehicles, with a small amount of size referencing for each vehicle for how wide to turn etc.
I think we will obviously benefit from the millions of miles driven by M3 MY, etc

It sure doesn't seem that way. We are around 17 months in the CT cycle and it doesn't have the full FSD. As Wes stated they made the cameras as similar to the MY as possible to get a working of FSD. We've been waiting for a full version since.

 

[REM]

I would think that 99% of the FSD data needed would be applicable for all Tesla passenger vehicles, with a small amount of size referencing for each vehicle for how wide to turn etc.
I think we will obviously benefit from the millions of miles driven by M3 MY, etc

The stacks do share some data, sure. But not all of it. And for the model to be the safest and most accurate, there are a tremendous amount of data points that have to be solidly in place.

Mass, acceleration, braking power, traction, how the vehicle fairs against crosswinds (which is partly the reason why the CT pathing is so terrible, I think), suspension profiles, regent strenth, etc.

 

[riggster]

They need to allow everyone to have free FSD. Not gonna pay $99 for something that doesn't autopark safely, and the fsd is what my model y was 2 years ago.

Your cybertruck FSD experience must not be nearly as good as mine. I use it to drive me to and from work every day and rarely have to intervene. Way better than model y FSD from 2 years ago, which was FSD 11. I switch back and forth between a model Y with hw4 and the cybertruck. While not as smooth as the model Y 13.2.8, the cybertruck is way better than anything running on FSD 11.

 

[riggster]

I would think that 99% of the FSD data needed would be applicable for all Tesla passenger vehicles, with a small amount of size referencing for each vehicle for how wide to turn etc.
I think we will obviously benefit from the millions of miles driven by M3 MY, etc

I think you may be onto something here. While having enough data to do a native build for the cybertruck would be best, DNNs can be tuned (also called fine-tuned) with a different dataset. This is commonly done in machine learning tasks where a pre-trained DNN model, trained on a large dataset, is adapted to a new, smaller, or different dataset relevant to a specific task. I’ve personally done it with images, but it works the same way with video, just takes a crapton more processing power. I’d be shocked if they weren’t taking this approach at Tesla. The great thing is that the more data they get, the better our trucks will drive. I personally think we’ll have uFSD not too long after it comes out for the y, depending on your definition of not too long.

 

[Gigahorse]

I think you may be onto something here. While having enough data to do a native build for the cybertruck would be best, DNNs can be tuned (also called fine-tuned) with a different dataset. This is commonly done in machine learning tasks where a pre-trained DNN model, trained on a large dataset, is adapted to a new, smaller, or different dataset relevant to a specific task. I’ve personally done it with images, but it works the same way with video, just takes a crapton more processing power. I’d be shocked if they weren’t taking this approach at Tesla. The great thing is that the more data they get, the better our trucks will drive. I personally think we’ll have uFSD not too long after it comes out for the y, depending on your definition of not too long.

Cybertruck, Cybercab, and whatever vehicle Tesla has coming next starting at 0 or not 99% for that matter seems like it would be crazy to me.
Sure the vehicles are all a little different size, have a bit different turn radius, etc, but for them not to use the billions of hours of data from other vehicles that in the big picture are almost the same would seem crazy.

 

[SCTesla]

Cybertruck, Cybercab, and whatever vehicle Tesla has coming next starting at 0 or not 99% for that matter seems like it would be crazy to me.
Sure the vehicles are all a little different size, have a bit different turn radius, etc, but for them not to use the billions of hours of data from other vehicles that in the big picture are almost the same would seem crazy.

They have a bunch of M3s mimicking the Cybercab. So it won't be at 0. It's probably more like a 50% start. As we see with the CT, it can drive with a modded version from the Model Y, but it's not complete.

 

[Alto]

I think you may be onto something here. While having enough data to do a native build for the cybertruck would be best, DNNs can be tuned (also called fine-tuned) with a different dataset. This is commonly done in machine learning tasks where a pre-trained DNN model, trained on a large dataset, is adapted to a new, smaller, or different dataset relevant to a specific task. I’ve personally done it with images, but it works the same way with video, just takes a crapton more processing power. I’d be shocked if they weren’t taking this approach at Tesla. The great thing is that the more data they get, the better our trucks will drive. I personally think we’ll have uFSD not too long after it comes out for the y, depending on your definition of not too long.

Just my supposition, but resources are limited. Cyber truck sales are low. The consequential business decision is why cyber truck lags in development of its software, even though it is the most recently released vehicle from Tesla.

 

[riggster]

Just my supposition, but resources are limited. Cyber truck sales are low. The consequential business decision is why cyber truck lags in development of its software, even though it is the most recently released vehicle from Tesla.

As a programmer who has programmed many deep neural networks, I can tell you that it takes almost nothing to reconfigure your program to go from a model Y to a Cybertruck. Just data. Now fine tuning a model Y DNN to work for a cybertruck is another story. Much more nuanced. If there were already 2 million cybertrucks on the road, we’d be where the model Y is right now, or very close to it. In another year they'll have more than twice as much data as they have now. Will that be enough? I have no idea but i sure hope so

 

[rovert43]

I don’t think there’s any reason whatsoever to suppose that we’re not getting unsupervised. We’re clearly going to get it after everything else. Last night he said that the model Y is going to get it first. It seems like the Cybertruck will get it last. But you still call the guy who graduated last in his med school class “doctor!”

As others have said, this is probably a numbers problem that Tesla didn’t expect. With so few CT’s on the road there’s less driver data to rely on. Dojo will do a lot but even so I think this is our biggest problem.

I think this is directly related to the expansion of the early beta cohort.

I think if the CT gets FSD and ASS they're going to sell a ton of them for those two reasons alone. Most people are intimidated by the power and capabilities of the CT and they're scared of it due to the 6 figure price tag. You make it autonomous all that goes out the window.

Sponsored