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:
Here's a transcript from a Twitter Spaces from today:

Whole Mars Catalog: When is V11 coming. What's your best guess?​
Elon Musk: Well, hopefully, before the end of the year. I mean, we're on V11.2 with a limited beta. So it's probably like V11.3. It's pretty soon. It's not far away. It's measured in single di- sort of in this <inaudible> roughly.​
Whole Mars Catalog: Nice. That's exciting. Well, the existing version's amazing. It's been driving me around with no takeovers most of the time.​
Elon Musk: Yeah. The thing about the V11 is there are a bunch of neural nets that are architecturally much better than 10.69, but they've got much more room to improve. Autopilot has just been going through a series of local maximums. So when you exit a local maximum, you first go down before you go up. So the neural net architecture in V11 is actually much better, but it just takes time to hone the details. It's rare for me to intervene at this point when I drive around. The vast majority of trips that I take on Autopilot are zero intervention. So it bodes well for the future.​
Chuck Cook: First of all, thank you for working on hard problems such as the unprotected left turn. Do you think that many of the hard problems are going to need that much effort? Do you think it's going to generalize across the space? I know V11 is a good example of that.​
I'm just curious if you think that they're going to have to put that much effort towards a lot of these corner cases, or if that was just a good example of how to put engineers to work to create a solution both in the simulator and in real life?​
Elon Musk: Every new neural net architecture is more generalized than the last time, so it actually ends up solving not just, say, extended unprotected lefts, but it ends up solving a whole bunch of things. With each major rev, there's more and more generalization of solving what is effectively the real-world AI problem. Tesla is, as far as I can tell, overwhelmingly the best at real-world AI, and our rate of progress is increasing.​
 
The thing about the V11 is there are a bunch of neural nets that are architecturally much better than 10.69,

Doubtful.

It's rare for me to intervene at this point when I drive around.

💩

The vast majority of trips that I take on Autopilot are zero intervention. So it bodes well for the future.

💩

Every new neural net architecture is more generalized than the last time, so it actually ends up solving not just, say, extended unprotected lefts, but it ends up solving a whole bunch of things

Obviously nonsense.

With each major rev, there's more and more generalization of solving what is effectively the real-world AI problem.

Complete and total nonsense.

Tesla is, as far as I can tell, overwhelmingly the best at real-world AI, and our rate of progress is increasing.

Haha.

He is no longer credible (has not been for some time).

Looking forward to V11 soon but let’s at least be honest about what it is going to be - hot garbage to start with. And that’s fine, I suppose.

Fortunately to all of us, this is now all transparently obvious. We know what to expect, and we will not be disappointed.

Once we get a firm grasp on reality, things will be much better, and might even see some progress.
 
Well, at least it sounds like it'll actually get released within the next 3 months.

It also sounds like it's currently not better than 69.3.3, so it'll take some more time to "hone the details." lol

I wouldn't be surprised if they needed to create and train entirely new neural networks to accurately estimate trajectories at highway speeds. FSD Beta hesitating and braking on city streets leads to being honked at. If it did the same on a highway, you risk being rear-ended at 60 MPH.
 
Oh gotcha. For sure. Lol if anything I'm more attentive when AP is turned on.

🤞 FSD11 comes out before the 26th and isn't stupid with it's lane centering. Be nice to have it on my 8 hour drive back home. 😁
Yea, it's funny I think I noticed that too. I feel like when I am letting it "drive" I am MUCH more aware of everything and generally at a higher state of alert. So basically IF anything is going to happen it is going to happen when we disagree and I intervene!

I am not seeing any "real" version that looks like it has V11 in it yet. The latest version looks like 44.30.x but from what I can tell it is going out to new(ish) cars and not specifically to people with FSDb. While I haven't see anyone publish release notes for it, but I'm pretty skeptical that it will just go out broadly to folks since it sounds like a pretty significant change.
 
Yea, it's funny I think I noticed that too. I feel like when I am letting it "drive" I am MUCH more aware of everything and generally at a higher state of alert. So basically IF anything is going to happen it is going to happen when we disagree and I intervene!

I am not seeing any "real" version that looks like it has V11 in it yet. The latest version looks like 44.30.x but from what I can tell it is going out to new(ish) cars and not specifically to people with FSDb. While I haven't see anyone publish release notes for it, but I'm pretty skeptical that it will just go out broadly to folks since it sounds like a pretty significant change.
V11 is on a very limited group of cars, likely only employees, at this time. As a result, you don't see the version appearing on teslafi or teslascope. Once Tesla produces a version that they feel is good enough to push to public testers, they will do so.

I believe Tesalascope had a set of earlier release notes for V11 that were leaked to them.

I expect that Tesla will use a similar strategy that has been used during the FSDb test program. They've learned that it is best to do staged rollouts of new versions, starting with a small number of testers, probably around 1000 or so. They can examine the telemetry from these cars to detect and fix any big issues. Then they rollout to larger groups, examining telemetry after each wave. V11 will, as Musk alludes, have some significant issues when it first comes out. It certainly sounds like it is much more than simply adding some specific highway NNs to the existing 10.x FSDb. And, if anything, NNs seem to be particularly susceptible to 'unintended consequences'. That being the case, we should expect overall performance to degrade at first.

Hopefully, with improved simulation capability, Telsa can minimize this degradation. But it will be there, for sure.
 
It certainly sounds like it is much more than simply adding some specific highway NNs to the existing 10.x FSDb. And, if anything, NNs seem to be particularly susceptible to 'unintended consequences'.
Tesla's approach has been to find general solutions that solve multiple issues at the same time, so it doesn't seem likely they'll add a new neural network specifically for highways but instead probably re-architect some neural networks to handle both highways and city streets. An example of re-architecting is the two-stage conversion initially for object kinematics then applied to VRU and NonVRU, and this probably was to reduce prediction latency especially for highway driving speeds but also helps city street driving too.

An example of general solution probably led to the Occupancy network solving issues related to unidentified objects and visibility/occlusion needed for safe creeping and road surface understanding of hill crests. Each of these could have been addressed with specialized code/NNs, but a general solution that understands things more holistically probably takes more effort to train correctly but also "shares knowledge" that helps other parts perform better too.

But your overall point of highlighting that new neural network architectures are needed for single stack does show that "just" having the city streets driving behavior handle highways wasn't enough. New general architectures would affect all of FSD Beta including city streets behaviors, so it does make sense for additional testing and polishing to reduce regressions with a slower rollout. Because it's now not "just" highway polish, this could be more like 10.69's initial release to the early access group then about a month before wide release.
 
Do you mean the CENTERING "jerk" or alignment to the road markings that occurs?
Both? In unmarked residential streets it's more noticeable. Super wide streets where I'm at. Either way, no traffic no movement, so jerky. Noticing that in general. Stops short, jerky acceleration, whatever.

Definitely not as good as YouTubers like to make it out to be. Had a hard brake due to a bicycle and it would refuse to stop decelerating when I pushed throttle forward for a good block or so. Anyways don't want to get too off-topic.

Hopefully FSD 11 general problem solving is generally better.
 
  • Funny
Reactions: AlanSubie4Life
Both? In unmarked residential streets it's more noticeable. Super wide streets where I'm at. Either way, no traffic no movement, so jerky. Noticing that in general. Stops short, jerky acceleration, whatever.

Definitely not as good as YouTubers like to make it out to be. Had a hard brake due to a bicycle and it would refuse to stop decelerating when I pushed throttle forward for a good block or so. Anyways don't want to get too off-topic.

Hopefully FSD 11 general problem solving is generally better.
That is weird, I've never had the car decelerate while pushing the juice peddle. In fact it warms you it won't brake, while you are pressing the throttle
 
  • Like
Reactions: powertoold