FSD Beta v11.x

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.

Deep Understanding Tesla FSD Part 1: HydraNet

From Theory to Reality, Analyze the Evolution of Tesla Full Self-Driving

saneryee-studio.medium.com

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).

Twelve Questions about Tesla Autopilot Technologies Exposed at 2021 AI Day

On Aug. 19 2021, Telsa held an AI day , exposing more technologies about Autopilot , from perception, planning, data labeling, simulation…

yuhuang-63908.medium.com

 

[AlanSubie4Life]

I am not a coder, but wonder if 1 setting can have an effect on other things that are unrelated.

Everything matters. It’s a massive system with numerous inputs, and outputs that are not explainable analytically by the programmers (that how this works as I understand it).

I think vehicle, settings, and surroundings all will play into the precise results for a given scenario.

However, in the vast majority of cases, on average, I expect everyone’s experiences to be very similar (while simultaneously of course very different, for the reasons given above). But in the end very similar.
The differences are just perception. Someday someone will produce an app which scores the drives so everyone can have an objective measurement. (Even then of course everyone will get different average scores - for the reasons above.)

For example, San Diego is extremely hilly and has high speed limits, which makes my experience much worse than someone in LA (mostly flat!), and probably San Francisco (super slow), for example. Or the Midwest as a more extreme example. But in the end our experiences are actually very likely the same, when corrected for these factors.

 

[WilliamG]

J

U

 

[FSDtester#1]

U

N

 

[WilliamG]

N

K

 

[WilliamG]

The most ridiculous thing is, - Tesla never reaches out to us, the people who actually bother driving normal roads. No, it's Chuck's Bloody Left Turn (actual name). YAAAS. MUST SOLVE THIS ONE LEFT TURN BECAUSE IT IS SOOOOO FAMOUS ON TEH YOUTUBS. You know FSDb is never coming out of beta when the focus is on one... freaking... left... goddamn... turn.

 

[WilliamG]

The most ridiculous thing is, - Tesla never reaches out to us, the people who actually bother driving normal roads. No, it's Chuck's Bloody Left Turn (actual name). YAAAS. MUST SOLVE THIS ONE LEFT TURN BECAUSE IT IS SOOOOO FAMOUS ON TEH YOUTUBS. You know FSDb is never coming out of beta when the focus is on one... freaking... left... goddamn... turn.

Dude. Calm down already.

 

[FSDtester#1]

Dude. Calm down already.

Maybe they don't see a lot of added benefit of reaching out to people who talk to themselves?

 

[WilliamG]

Maybe they don't see a lot of added benefit of reaching out to people who talk to themselves?

You're probably right.

 

[mtbwalt]

I do find it reassuring that Tesla is actively soliciting causes for forced disengagements.

It suggests that they have enough capacity at this point to actually care about the scenarios that are most problematic, through a systematic rather than anecdotal approach. The refinement stage rather than the core functionality stage.

 

[AlanSubie4Life]

You know FSDb is never coming out of beta when the focus is on one... freaking... left... ********... turn.

Hmmm. I don’t know that it is so much the focus on that turn, as it is that they still can’t even get close to successfully completing that turn. Tons of improvement, for sure. But no success.

Success breeds success.

Think about the fundamentals there:
Detection range.
Speed estimation.
Car position control.
Car speed control.
Car body language.
Intent predictions.
Occlusion detection (occupancy network)
Memory
Etc.

 

[WilliamG]

I do find it reassuring that Tesla is actively soliciting causes for forced disengagements.

It suggests that they have enough capacity at this point to actually care about the scenarios that are most problematic, through a systematic rather than anecdotal approach. The refinement stage rather than the core functionality stage.

The problem is this will take years to collate these, and there will be no end of edge cases. It shows we're nowhere near self driving the way this forum envisions it.

 

[jj3884]

The problem is this will take years to collate these, and there will be no end of edge cases. It shows we're nowhere near self driving the way this forum envisions it.

Depends on how they do that analysis. Natural language processing, combined with the latest contextual AI and then the car telemetry I'd hope they'd be able to rapidly categorize and collate this data.

***EDIT***
Though given that OpenAI hurt Musk's fragile ego they may not have a great way lol.

 

[mtbwalt]

The problem is this will take years to collate these, and there will be no end of edge cases. It shows we're nowhere near self driving the way this forum envisions it.

Perhaps, but just the fact that they are now classifying these disengagements means they have an opportunity to manage the issues.

With likely tens of thousands of reports now likely coming in every day, they should have an intake methodology that will allow them to classify the submitted disengagement issue along several parameter types and file them for statistical purposes and potential analysis.

The driver’s description of the issue is perhaps one of the most critical factors and allows the initial reviewers to quickly classify the problem, presumably augmented by video and telemetry/modeling data.

It’s progress.

 

[EVNow]

Or a gate... apparently

That’s the well known low occlusion object problem. An edge case for NN. Try something else …

It's the TMC app, copy pasta links into Safari and they'll work.

Can’t even copy the address on the app !

 

[valem]

Just a warning, if you're experiencing major problems with 10.69.25, V11 isn't going to solve your issues. I look forward to your experience once you get it.

So far I am actually pretty impressed with V11.
Before we had 10.69 go wide, I was actually enjoying FSDb, but for me, for some reason, with the free for all (everyone gets it), it became so cautious that I really could not stand it.
Since getting V11 yesterday, I have only used it on city streets, within 10 miles of home, but so far I like it. It feels way more natural and definitely smoother.
Still using brakes to stop at lights and stops, which that I can't stand, so I disengage just to properly slow down with regen... other than that, so far so good.

Tomorrow I have a 216mi road trip, with about 80% highway so I will see how it does.

Side note: I would normally hate a road trip with mostly straight interstate driving... Funny how with FSDb V11 I am now excited to test it out

Let's see how V11 does on a long road trip with mostly highway... to be continued...

 

[jtsai1031]

still don't have version 11. still on 2022.44.30.10. should i be concerned?

 

[FSDtester#1]

still don't have version 11. still on 2022.44.30.10. should i be concerned?

Give it 24 more hours, then hit the panic button.
I think you will get it overnight... meaning tonight

 

[Dewg]

still don't have version 11. still on 2022.44.30.10. should i be concerned?

No. It's not done rolling out. Give it another week or two.

 

[jebinc]

The most ridiculous thing is, - Tesla never reaches out to us, the people who actually bother driving normal roads. No, it's Chuck's Bloody Left Turn (actual name). YAAAS. MUST SOLVE THIS ONE LEFT TURN BECAUSE IT IS SOOOOO FAMOUS ON TEH YOUTUBS. You know FSDb is never coming out of beta when the focus is on one... freaking... left... goddamn... turn.

Exactly, but leave the small school children unprotected and in harm’s way (see my school bus post from Friday for details on WTF I’m referring to)…

 

[jebinc]

Dude. Calm down already.

I love this guy; finally breaking out!