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:
The wonders of having data from everyone.

I am sure there are lots of problems.

You’d think one priority for them would be making sure to deliver anything to customers by tomorrow that has not already been recognized as revenue. Not sure what that would be at this point though, since we aren’t sure exactly how their revenue recognition for this stuff is handled by the bean counters (plus there is the USS-less stuff for people which is kind of unrelated but may affect timing somehow).

Personally I am still waiting for the reverse Smart Summon I was promised. I imagine there is significant revenue for that that is not recognized yet. But probably not going to get that tomorrow!
And don’t forget about it’s related cousin, “A.S.S.”!
 
The wonders of having data from everyone.

I am sure there are lots of problems.

You’d think one priority for them would be making sure to deliver anything to customers by tomorrow that has not already been recognized as revenue. Not sure what that would be at this point though, since we aren’t sure exactly how their revenue recognition for this stuff is handled by the bean counters (plus there is the USS-less stuff for people which is kind of unrelated but may affect timing somehow).

Personally I am still waiting for the reverse Smart Summon I was promised. I imagine there is significant revenue for that that is not recognized yet. But probably not going to get that tomorrow!
Alan, you definitely should read that green twitter thread.. there is a bunch of stuff in there he found that you have addressed multiple times

Edit: add me to the software stuck category. It says last checked 3 days ago.
Oddly enough, this time it will let me constantly recheck over and over and there is no 24 hour wait.
 
  • Like
Reactions: Ramphex and jebinc
Alan, you definitely should read that green twitter thread.. there is a bunch of stuff in there he found that you have addressed multiple times
Yeah I posted it in the other thread about Elon FSD beta tweets earlier.

Seems like behavior could improve based on this but I wouldn’t expect it to be particularly correlated with a given release. Anyway I commented more elsewhere. I certainly don’t think this has a significant impact on changing drive quality over short time scales.

Brand new streets/traffic control? Sure. But we would kind of expect that anyway.
 
Last edited:
“TeslaImpairedVision” is “Nearsighted,” due to crappy low resolution cameras.
It seems fine to me. I haven’t personally had any issues with phantom braking like some others have had. If anything it seems to do it less now (on the freeway) than it did previously. Obviously on surface streets it is a whole different ballgame. We’ll see how version 11 does! Those left turns look like they have 8-9 seconds of vision now! That is 200 meters!!! Unfortunately it gets target fixated on the oncoming traffic too and consequently stops in the middle of the road.
 
It seems fine to me. I haven’t personally had any issues with phantom braking like some others have had. If anything it seems to do it less now (on the freeway) than it did previously. Obviously on surface streets it is a whole different ballgame. We’ll see how version 11 does! Those left turns look like they have 8-9 seconds of vision now! That is 200 meters!!! Unfortunately it gets target fixated on the oncoming traffic too and consequently stops in the middle of the road.
There’s a difference between vision and clear vision (for tracking and accurate decision making). Why else, then, would Tesla have dialed back to only seven car lengths? Because the radar could see clearly, slightly farther than “ImpairedVision” (for short), can.
 
The NHTSA FSD Beta recall is covered in the 11.3.x update. Oddly the NHTSA window auto rollup recall (which was issued before) is NOT covered. NHTSA states what they want done but NOT how to do it. It is up to Tesla on how to implement the recall and they have done that since NHTSA had also stipulated no more Beta updates until the recall was implemented. You can't send out to ⅓ of the fleet and say the NHTSA recall is ready for them but not the other ⅔.
You absolutely can - sheesh Julien, wake up. The V11.3.0 is ran through simulations, then released to a handful of internal people for preliminary testing. Then V11.3.1 is released to a handful of external people for further validation. They find issues once telemetry shows problems in the real world. They make tweaks to the code to smooth out some of the issues they found. They release V11.3.2 to a wider group, and again find problems once the code is on a wider array of cars in geographically wider areas of deployment. They tweak further and release 3.3 to an even wider range of cars - 1/3 as you indicated. They start getting reports of problems on a wider scale, so they work on the problems and will release 3.4, to the 1/3 that already have 3.3, and perhaps to a wider range. They'll watch telemetry and reports of issues and decide if that's the update that they're comfortable with releasing to the fleet. If not, there will be a 3.5.

The difference between this roll out and all the other ones is that this one is going slower than others because this one is in response to regulators and the world is staring at Tesla to see how they handle it. If Tesla rushes out the update because users are impatient "I WANT IT NOW!" children, it could backfire in the media and regulators could come down even harder.

Disclaimer: I don't have V11, and I'm totally fine waiting for whatever the final version ends up being. I want Tesla to get it right, and get regulators off their back.
 
Why else, then, would Tesla have dialed back to only seven car lengths? Because the radar could see clearly, slightly farther than “ImpairedVision” (for short), can.
They didn’t change this, did they? Or am I misremembering?

I thought it went from 7->1 to 7->2.

And it isn’t car lengths of course.

I don’t ever remember sufficient following distance (time) being available but maybe my memory is poor.
 
They didn’t change this, did they? Or am I misremembering?

I thought it went from 7->1 to 7->2

And it isn’t car lengths of course.
I don’t recall, other than the Vision version is “less than” what the capabilities of radar was. So, Tesla felt at the time, and still does, that Vision needs to remain “less than” (capable) …
 
  • Like
Reactions: Ramphex
I don’t recall, other than the Vision version is “less than” what the capabilities of radar was. So, Tesla felt at the time, and still does, that Vision needs to remain “less than” (capable) …
Sure. The issue seemed to be that they don’t trust the vision in higher speed closer following scenarios. (Hence <=85mph and >= 2.) But that is not what I was originally referring to of course. Now we have no idea how the settings will work (plus I want aggressive/assertive behavior with maximum following distance which now seems hopeless since these things have been merged!!!).

I do not know if anyone has compared the following time at a given speed (80mph would be good) on old setting 7 vs. the three new settings. Seems like a very easy thing to test and document.

But undoubtedly we won’t get 3-4 seconds as a result, sadly.
 
  • Like
Reactions: Ramphex and jebinc
Sure. The issue seemed to be that they don’t trust the vision in higher speed closer following scenarios. But that is not what I was originally referring to of course. Now we have no idea how the settings will work (plus I want aggressive/assertive behavior with maximum following distance!!!).
Agreed. I had it reversed (my poor memory). The premise remains, however, correct - Currently, Telsa feels Vision is “less than” what it replaced, for the time being.
 
  • Like
Reactions: Ramphex
Can anyone report whether FSD v11 has figured out how not to set up camp and pitch a tent in people’s blind spots on the freeway? Any evidence of some sort of optimization here? I thought maybe since it is now truck aware it might play some “tricks” (this is not very complicated!) and minimize blind spot time.

Unfortunately my guess is that it is so hopelessly unable to adjust speed without jerk that this would be a non-starter for the team.
 
The recall is a done deal. You can't "test" a recall on some customers and then implement it. NHTSA was clear and no new FSD Beta installed software until the recall was meet. There may be other reasons but the recall CAN'T be one of them.

Why is it so hard for you to grasp the concept? Yes, V11.3.X will address the items in the recall, but Tesla is not satisfied with the performance of V11.3.X right now and is making changes to ensure it's performing as best as it can. You want the entire fleet to have V11.3.3. That version has some issues as reported by several here, and reddit, and FB, etc. And some of those issues are in reference to the recall items. It may be only a small percentage of users experiencing those problems, but it's enough to warrant further tweaking of the code and a release of V11.3.4. Once Telsa has telemetry showing that the cars are performing within an acceptable margin of error with regards to the recall items, they will release to the entire fleet and notify NHTSA that they have satisfied the recall and are resuming FSD Beta rollout. NHTSA will likely monitor carefully and hopefully reports are light and a new recall isn't issued. That could be V11.3.4, or V11.3.5 - we just don't know yet.

Just because your car didn't get the update, and you weren't one of the 1/3 of cars that got it is not a reason to jump up and down and rend your clothes. You should be worried about the big picture - the safety of others, and the potential damage that could occur to the program if this rollout isn't managed carefully while regulators are staring at Tesla. You will eventually get the update, you just have to wait and trust that the company is trying to be responsible about the rollout to minimize issues.
 
FWIW the legacy highway NoA was about the same amount of close to perfect as the new V11 one so far for me.

Works fine generally.

Is kinda garbage in merge situations since it waits to last possible second.... even for things like a double-on-ramp where there's arrows showing one lane is ending, nobody is in the other lane, and the car still waits till the lane going away narrows/ends to get out of it.

So I mean I'm glad it wasn't any worse given the major change of code... but haven't seen any evidence of it being better either.

(arguably going from 6 specific distance settings to only 3 vague ones is worse I guess?)
Here are the specific improvements "for me"
  • Smoother lane changes. My wife who hasn't liked FSD didn't believe I was using FSD they were that much better.
  • With NoA I would confirm lane changes because NoA would frequently try and move into the faster lane when a car was fast approaching. I didn't trust the lane changes then but today every change was safe.
  • On some exit ramps I often had to reduce the exit speed or even disengage when it was jerky. Today they were all fine.
  • Stays further away then NoA used to when trucks are along side. Frankly I was fine with the previous behavior so a wash on this feature.
  • Handled wide lanes better. There are 3 specific wide lanes that used to drive me crazy. All 3 were hands off today.
Other
  • Merge- I was lucky here since the traffic was light today. Even though I now understand how Tesla is trying to do zip merges they aren't done very well. Need a lot of work.
 
other useful information like lane info
This lane info can contain connectivity from/via/to probably similar purpose and usage as OpenStreetMap: https://wiki.openstreetmap.org/wiki/Relation:connectivity

Here's some examples from Tesla's data where the numbers are lane positions with 1 being left-most:
JSON:
// left lane splits increasing from 2 to 3 lanes
          "from_lanes": "1|1|2",
          "to_lanes": "1|2|3",

// right lane splits increasing from 2 to 3 lanes
          "from_lanes": "1|2|2",
          "to_lanes": "1|2|3",

// right lane splits increasing from 3 to 4 lanes
          "from_lanes": "1|2|3|3",
          "to_lanes": "1|2|3|4",

// right lanes merge decreasing from 3 to 2 lanes
          "from_lanes": "1|2|3",
          "to_lanes": "1|2|2",

// right-2 lanes continue while left lane ends (probably left turn lane at intersection)
          "from_lanes": "2|3",
          "to_lanes": "1|2",
          "vias": 1,

// 3 lanes continue as right-most 3 of 5 lanes (e.g., interstate merge)
          "from_lanes": "1|2|3",
          "to_lanes": "3|4|5",

FSD Beta 11.x might have required new navigation lanes data types to be added or maybe focused on increasing the quality especially for highway data. Most likely lane counts and lane connectivity data was understood by 10.x, so if Tesla generally improved the quality of this data in preparation for rolling out 11.x, people could notice better lane selection and longer-term planning while still on the older FSD Beta.
 
Here are the specific improvements "for me"
  • Smoother lane changes. My wife who hasn't liked FSD didn't believe I was using FSD they were that much better.
  • With NoA I would confirm lane changes because NoA would frequently try and move into the faster lane when a car was fast approaching. I didn't trust the lane changes then but today every change was safe.
  • On some exit ramps I often had to reduce the exit speed or even disengage when it was jerky. Today they were all fine.
  • Stays further away then NoA used to when trucks are along side. Frankly I was fine with the previous behavior so a wash on this feature.
  • Handled wide lanes better. There are 3 specific wide lanes that used to drive me crazy. All 3 were hands off today.
Other
  • Merge- I was lucky here since the traffic was light today. Even though I now understand how Tesla is trying to do zip merges they aren't done very well. Need a lot of work.

How is it doing with blind spots when driving with traffic at a similar speed? I guess probably not a big enough freeway to properly test - two lanes it is usually less of an issue (you are either passing or following). At 4 or 5 lanes it starts to be the norm that you are driving close to the same speed as adjacent vehicles.

This is a key usability issue for in-city travel (less of an issue for long-distance travel due to above) so hopefully they have addressed it, since it is an easy one.
 
Last edited: