Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register
This site may earn commission on affiliate links.
The next big milestone for FSD is 11. It is a significant upgrade and fundamental changes to several parts of the FSD stack including totally new way to train the perception NN.

From AI day and Lex Fridman interview we have a good sense of what might be included.

- Object permanence both temporal and spatial
- Moving from “bag of points” to objects in NN
- Creating a 3D vector representation of the environment all in NN
- Planner optimization using NN / Monte Carlo Tree Search (MCTS)
- Change from processed images to “photon count” / raw image
- Change from single image perception to surround video
- Merging of city, highway and parking lot stacks a.k.a. Single Stack

Lex Fridman Interview of Elon. Starting with FSD related topics.


Here is a detailed explanation of Beta 11 in "layman's language" by James Douma, interview done after Lex Podcast.


Here is the AI Day explanation by in 4 parts.


screenshot-teslamotorsclub.com-2022.01.26-21_30_17.png


Here is a useful blog post asking a few questions to Tesla about AI day. The useful part comes in comparison of Tesla's methods with Waymo and others (detailed papers linked).

 
Last edited:
Elon says 11 was planned for this year but was *surprise* delayed. It is in ”alpha” now.

Question is what is alpha … rough with a lot of bugs ? But that is what Beta looks like !

To me alpha sounds like a POC - proof of concept. Significant dev work remains, apart from fixing bugs. Hope we get 11 by this time next year.
 
I think alpha is just a roughly shelled out version that mostly touches on each of the critical components with at least some of them not working ideally or possibly even as intended but far enough along to be tested internally and allowing for rapid iteration in preparation for beta releases.

I am most interested in how object persistence winds up improving things because I think that is a big gap with today’s modeling.

Once we see a few spins of this model, we might gain some insight as to whether the current suite of hardware is enough or if additional revisions might be necessary.

Looking forward to seeing what this generational improvement brings.
 
I think alpha is just a roughly shelled out version that mostly touches on each of the critical components with at least some of them not working ideally or possibly even as intended but far enough along to be tested internally and allowing for rapid iteration in preparation for beta releases.
That would be a good definition - though in the industry I think people use it to indicate that it’s still not good enough to be beta, mostly for external communication. We never used alpha internally, using various other milestones instead. You wouldn’t find Windows builds named alpha, for eg.

I wonder - with so many items slated for 11, whether any of them come to some 10.x version. Also, Tesla will probably put most effort into 11 and not spend a lot of effort on 10.x code that is possibly short lived/throwaway.
 
  • Like
Reactions: edseloh and impastu
Yes and Musk even reiterated Lex Friedman Podcase a couple of days ago. However I bet that we get Highway first with 11.0/11.1 with Summon/parking lots added later in 11 (like 11.6).
Auto parking already looks ready to be part of FSD.

Summon is partly done in the cloud … not sure that part will change soon. But once navigation is set in the cloud, actual driving can easily be FSD, it’s just like driving unmarked roads.
 
  • Like
Reactions: impastu
Small clarification The 13ms is for all 8 cameras combined, something like ~1.5ms per cam FYI. Elon makes it sound like per camera at first but his next couple of sentences sort of clarify it. I'm also not entirely swallowing his great night vision statement, because he's not accounting for sensor noise. It's not like the sensors produce nothing if photons are hitting it. I'm not saying its not better than expected, just he doesn't fill in the details.

This raw image / photon count / night vision thing brings up many questions …

- Would raw images be different enough between different camera models that if they train with HW3, they have to retrain with HW4, maintain different NNs.
- I wonder whether in their tests raw image at night is doing better than processed. Ofcourse it is possible they were aggressively reducing noise when processing images that might account for this.
- There are NNs that can greatly improve night vision. I wonder whether they are planning on using those.


421EF7B6-5EDD-4739-8451-87B36A3628EE.png


 
  • Like
Reactions: FSDtester#1
Auto parking already looks ready to be part of FSD.
In what way? Is there a "Go find parking" feature? Or do you mean it's on the FSD roadmap? That will be great! I do expect to see a bunch of Teslas in the news for parking on top of things that are not parking spaces when that feature is first released.

Another thing that is going to need to be better explored is FSD v11 vs v11 Holiday release, perhaps a better naming strategy when talking about FSD to eliminate confusion.
 
In what way? Is there a "Go find parking" feature? Or do you mean it's on the FSD roadmap? That will be great! I do expect to see a bunch of Teslas in the news for parking on top of things that are not parking spaces when that feature is first released.
No - just normal vision based perpendicular auto park.

BTW, go find parking is just a combination of auto park and navigating through parking lots. Once they have proper parking lot navigation, go find parking should be fairly easy.
 
  • Like
Reactions: israndy and impastu
Wonder if we will get a 10.9 first?
I hope so. I don't expect 11 for several months ... and I don't want to be stuck with 10.8.

But I've to say ... the FSD team is probably interested in working on 11 rather than making small improvements to 10.x. Afterall they keep making improvements to 10.x and a lot of us think there isn't much of an improvement or there are a lot of regressions.

These are always tough decisions for the engineering team ...
 
I am happy if they don't bother taking on more betas, *now that I got mine*, until v11. Focus on the new paradigm, don't bother with making pretty changes.

Now as far as the other v11 (Is there a NAME for those two sw releases that is different??) that everyone is upset about the look of, feel free to keep making changes to that and releasing it's updates until the rabble have been sated.
 
  • Like
Reactions: melsee and Sporty
I am happy if they don't bother taking on more betas, *now that I got mine*, until v11. Focus on the new paradigm, don't bother with making pretty changes.

Now as far as the other v11 (Is there a NAME for those two sw releases that is different??) that everyone is upset about the look of, feel free to keep making changes to that and releasing it's updates until the rabble have been sated.
v11 is the UI (user interface) update for almost all Tesla MCU's and 10.8 is the current FSD Beta. Soon we will have FSD Beta 11.x and at the same time almost everyone will also have the v11.x UI. Just a pure coincidence of numbering and nothing else. By later this year the FSD Beta numbers (like 13.2) will be well past the UI version number of v11.x (the UI will be v11.x for at least year or so since the UI was on v10.x for 2 years) so that they will no longer be conflated.
 
  • Like
Reactions: LastGas
- Planner optimization using Monte Carlo simulation

Pretty sure FSD beta already uses Monte Carlo tree search for its planner / decision making. Monte Carlo tree search itself doesn't involve NNs, although it might use the NN outputs.

The example used during AI Day involved planning for lane changes based on predicted futures, and FSD beta already does this. Sometimes it would speed up vs slow down for route-based lane changes.
 
  • Like
Reactions: rob1987
Pretty sure FSD beta already uses Monte Carlo tree search for its planner / decision making. Monte Carlo tree search itself doesn't involve NNs, although it might use the NN outputs.

The example used during AI Day involved planning for lane changes based on predicted futures, and FSD beta already does this. Sometimes it would speed up vs slow down for route-based lane changes.
No they use a hybrid approach now.

I've to watch again - but it sounded like they were using Monte Carlo so to reduce the number of simulations / searches they had to do. May be that was WIP.

Either way - the core idea doesn't change. They are not "hard coding" the path for every situation.

ps : After a bit of reading - looks like what they were/are using is a hybrid approach. MuZero based NN for planning is probably in development.


0984FC71-42AB-429E-85FC-8A744BA26AA6.png