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:
I lived, smooth as silk drive today. .
Ok, seriously, I noticed some small changes, but I already liked V11 better than 10.69. 1st, for me, Nags are back, about 1 per every minute. Highway is excellent, ego loves to stay in passing lane and did pass a slow driver switching from left lane to right to pass. If it doesn't detect anyone in left lane, it just stays there. City, main complaint is same with everyone else. Need more regen braking. It appears that if nobody is at an upcoming stoplight it slows almost perfectly. However if there are 2 or more cars at a light it waits too long to slow down and applies friction brakes too late, and hard, then releases and finishes with regen. This doesn't make any sense why it isn't fixed. If some jerk cuts you off and it needs to apply the friction brakes, I get it. However why is it smooth when you are coming to a red with no cars, or a stop sign it sees way in advance. It sees the cars at the light, and displays them along with the red light. I am going to start disengaging just to record voice notes for what I just explained. On a plus, on a 60 mile round trip, I had no disengagements, and 1 intervention. 50% city and Highway with medium traffic. The intervention is exactly the same as @Ramphex . There is 1 spot I found that is slows from 55 to 30 for no reason. No cars, shadows, nothing to explain it. I have let it do it's thing in the same place when there aren't any cars behind me multiple times. Same result since about 6 months ago. Fix that map data and friction braking and I will be a happy camper. Otherwise, ULT and URT's are pretty good.
Curious, are you using the setting allowing the car to “use friction brakes selected”? Controls > Pedals and Steering > use friction brakes. I turned that off some time ago when they released it late last year.
 
10.69.25 was definitely better behaved on city streets. But 11.3.x is much better than NOA on either limited access highways or high speed rural two lane highways. Let's hope that 11.4 isn't too far behind.
Oddly, I have a unicorn car, so my freeway AP works great for me. Not sure what more I could want on the freeway.
 
Downloading the update
Yeah! Another 1,000 have been added to TeslaFi cars eligible for 11.3.6 tonight, probably starting around 11:30 tonight, 4/10.

That makes the total pending plus installed around 3,000 out of the ~5,500 FSD cars on TeslaFi, i.e. ~55% in 3 days, with 2,000 added in two bunches today. Cars from 2023, and 2018, and from 10.69 and 11.x. So a full rollout it would seem.

Some folks will delay installing it, so it won't get to 100% any time soon, but I expect it will be available for all the FSD buyers except those on 2023.x versions.

Oh, and my app now says I have an update available, cool, a new version to check out tomorrow. Install 2022.45.15? You betcha! Click.
 
11.3.4 drove about 250 miles for me round trip over the weekend. I've driven this route many times on various FSDb versions, and this was definitely the best trip I've had. There are a few places along the route where I know I will have to disengage, and those aren't much improved. Fortunately, this time I was able to report them with a description. Several of them are cause by bad map data.

I was very impressed with the freeway part of the trip. It made lane changes pretty much exactly as I would have. Most notably, it stayed in the left lane unless/until another car was approaching from behind, then it changed lanes to let the faster car pass. I never saw NOA do that.

The only thing it didn't do was to slow down when it saw a police car watching for speeders. That shouldn't be too hard to implement, should it? :)

Here's one that FSDb almost never gets right. I think it did it correctly once on the last iteration of version 10, but it failed yesterday: Coming north on S Main St., the road curves sharply to the right just as the single lane splits into two. The route calls for a left turn onto Rt.13, but FSDb almost never manages to get into the left turn lane. Yesterday it got stuck in between the two lanes and just stopped.

1681218163937.png
 
Tried FSD 11.3.6 yesterday for a short time. It seems everyone has a problem spot where FSD does not work well. Mine is the 4 way stop with a left turn into my neighborhood. Normally FSD gets confused and takes forever to decide what to do but eventually does the turn. With 11.3.6 it was less bad but still not good. Another big fail was turning from an arterial onto a freeway onramp where FSD for some reason thought the speed limit for 3/4 of the onramp was 25mph instead of getting up to highway speed. Epic fail IMO. Seems like FSD has a poor capacity for context. It needs to recognize that a 25mph speed limit for a freeway onramp is not correct. I REALLY want FSD to work flawlessly but am very pessimistic. I read something recently where there was speculation that the only way FSD is ever going to work is for an AI to do it. If true we might have flawless FSD but zero idea how it works. Also might have real time code modifications.....and of course the AI could purposely crash the car for an unknown reason. Sorry for my pessimism and paranoia:) The question is whether FSD is going to incrementally achieve the final goal, or whether the approach is fundamentally flawed and these updates are just band-aids on an approach that is doomed to never work.
 
Last edited:
  • Informative
Reactions: jebinc
You must be one of the few who has excellent performance on tight freeway interchanges with high curvature! Crazy how cars are so different.
Not too many high curvature interchanges around me. The 55 to 91 is probably the tightest one I take, and it's fine. Speed is automatically lowered on the curve, and the moment the transition is finished, the car resets the limit to 65, so I can just wheel it up if I want. I'm not saying I'd use AP with a full pot of chili in the car, but it works for me.
 
Not too many high curvature interchanges around me. The 55 to 91 is probably the tightest one I take, and it's fine. Speed is automatically lowered on the curve, and the moment the transition is finished, the car resets the limit to 65, so I can just wheel it up if I want. I'm not saying I'd use AP with a full pot of chili in the car, but it works for me.
Yea, I feel like it would make ME feel better if it provided some sort of acknowledgment that it was going to slow down, or maybe slow down before the turn, accelerate after the turn.

...and all depends on the chilli
 
How can we be saying 11 is better than NoA when 11 literally does not merge at all?? It sucks and is unusable for merging until they fix it. Merging (or lack of it) has always sucked in FSD beta. It just “rides it out” 😏🙁
Merging seems no worse than NOA, right ?

BTW, I should say I've not had much problems merging. But then the traffic here is not as maniacal as in some other cities ... if I think there might be a problem, I just signal to merge earlier.

BTW, lane changing with 11 is better than NOA, as others have also commented in detail.
 
Screenshot 2023-04-11 at 9.03.31 AM.png


I am now on 11.3.6 and FSDb continues to put me in the wrong lane (as with previous version). When heading NW on lakeland hills, the plan is to turn Left on 69th. And instead of waiting for the correct left turn lane it always puts me in the opposing turn lane (oncoming). That lane has a hard yellow curb, so once you are in it is nearly impossible get out. Not sure if that is mapping data or just poor planning.
 
When heading NW on lakeland hills, the plan is to turn Left on 69th. And instead of waiting for the correct left turn lane it always puts me in the opposing turn lane (oncoming). That lane has a hard yellow curb, so once you are in it is nearly impossible get out. Not sure if that is mapping data or just poor planning.

Wow. Is that configuration common where you live? In my region, it would definitely be marked as a shared left-hand-turn-lane for the entire section heading SE, and then turn into a dedicated left-hand-turn-lane for the NW drivers closer to the main intersection.
 
I read something recently where there was speculation that the only way FSD is ever going to work is for an AI to do it. If true we might have flawless FSD but zero idea how it works.
We will always know how AI works so long as it is executed on computers. That's because we can examine every facet of their operation; we can collect all the pertinent data. After that, if people cannot figure it out just by staring at that data, and if we cannot use traditional computer tools to examine the data and figure it out, we can use AI to look at the data, figure it out, and explain it to us.

Not sure if that is mapping data or just poor planning.
Yeah, it's not clear what's going on, but getting into a turn lane early is a common problem. My guess is poor planning, and that they need some neural nets that operate on longer time horizons. That sort of thing seems to be what they've just started working on.
 
  • Like
Reactions: FSDtester#1