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:
So it may be coincidental, but after updating I’ve been unable to reliably connect to my car via the app. Maybe that might be a reason for the delayed rollout. Dunno. Maybe coincidental but I’ve never had the problem before.

I’m very happy the management of FSD on unmarked roads appears improved though.
 
  • Informative
Reactions: VanFriscia
Nope never heard of it. Starting driving a long time ago which I guess explains it. I just merge in as gracefully as possible. It's interesting how many YouTube videos complain about FSD waiting too long till the end to merge.
I only heard about zipper merges a couple years ago. They're fine in most cases, but I know that I don't want to be dumped onto the shoulder if a high speed merge attempt at the end of the merge lane goes bad. Shoulders can have all manner of crap on them, including broken down cars. Or there may not even be a shoulder. The merge that bothers me on my drives ends in a curb, grass and trees. I don't want to go long on that merge, yet FSDb does a high speed run in the merge lane, getting over only at the end.
 
  • Like
Reactions: gsmith123
It's coming - it's a work in progress. It's going to take time to train the occupancy network to recognize a massive set of objects that represent what's drivable and not drivable
Big assumption!

If you’ve seen some of the catastrophically and comically poor results coming out of Midjourney, ChatGPT, Bard, Bing, etc., it makes you realize how far AI has to go. Especially when you realize these are errors those tools make when backed by massive supercomputers. Comprehension of context and the entire scene are a real struggle!

I don’t think it is clear at all that the car (alone) which will not hit objects with safety way better than humans (alone) is going to be realizable with our vehicles. There is just zero margin for error in either direction. It has to be right every time!

Just an incredibly hard problem, without just setting the threshold for action high, and leaving the rest to the human. (L2)

If you are saying it is an L2 system, there is no reason they need to identify these cases, so I don’t think it is reasonable to say confidently they are coming, except in a very general sense (not applicable to our vehicles). It’s not just a training problem!
 
Why? Likely 90% of a long road trip uses AP which you have and FSD will still work. Will not making a left at a light at end of trip give you anxiety?
Because I just want it. I have FSD and I want it. I’m making several stops along my journey will have plenty of chances to use the single stack which looks much improved from what I have seen, even on the highway. Of course now that I’ve seen it - using what I have seems old and ready for an update. Ho hum.
 
For anyone with the new 11.3.2: How does it handle lane changes on freeways while using navigation? Are they automatic — and annoying — like City Streets while navigating? (I seldom use navigation and FSD together on City Streets due to the excessive and bizarre lane changes.)
Seems to be pretty solid, no real issue except when it changes lanes, it does so correctly, but if your passing someone doing 68mph, and you’re doing 71, it takes forever, like it would if you were manually driving. First world problems.
 
It leads me to believe that AutoPark is using vision. I tried it this afternoon, and while parking, no feedback from the USS (mid 2022 Y with USS) was displayed. It parked perfectly, although, I thought I was unble to open the hatchback. Alas, it did, but it was tight. Parked pert near perfectly. A little crooked in the spot, but it worked.
 
Big assumption!

If you’ve seen some of the catastrophically and comically poor results coming out of Midjourney, ChatGPT, Bard, Bing, etc., it makes you realize how far AI has to go. Especially when you realize these are errors those tools make when backed by massive supercomputers. Comprehension of context and the entire scene are a real struggle!

I don’t think it is clear at all that the car (alone) which will not hit objects with safety way better than humans (alone) is going to be realizable with our vehicles. There is just zero margin for error in either direction. It has to be right every time!

Just an incredibly hard problem, without just setting the threshold for action high, and leaving the rest to the human. (L2)

If you are saying it is an L2 system, there is no reason they need to identify these cases, so I don’t think it is reasonable to say confidently they are coming, except in a very general sense (not applicable to our vehicles). It’s not just a training problem!

Teleportation might be easier than HW3 coming anywhere close to solving L2.
 
Where I live, roads are updated frequently, and it's very obvious when FSD has the correct map data for an intersection and when it doesn't. So I'm thinking it's using map data more than people realize.

For example, they recently put up a new stop sign at a nearby intersection, and it reads it at the last possible second. I've had to slam the brakes several times when it didn't pick it up at all (I wanted to see if it would ever read the stop sign and only tested it if no one was around).

On several other nearby intersections, they were converted from a two-way stop to a four-way stop. When driving in the direction that always had a stop sign, it doesn't see that the perpendicular road also has a stop sign and creeps like crazy instead of its usual behavior at four-way stops of going when it's my turn.

On a near by arterial road, it's lane data must be really off because FSD flat out ignores the yellow lane marker and will try to drive on the wrong side of the road. Thankfully the most recent FSD update made it more gracefully handle this failure scenario, but it's still an issue.
Agree, it’s using map data almost exclusively. They just installed a stop sign on a street nearby. It often wants to blow past the sign because it sees it too late. It’s not that it isn’t using vision, it’s that it’s sometimes too slow. If others are stopped at the sign, it stops too and displays the sign on the map.
 
It leads me to believe that AutoPark is using vision. I tried it this afternoon, and while parking, no feedback from the USS (mid 2022 Y with USS) was displayed. It parked perfectly, although, I thought I was unble to open the hatchback. Alas, it did, but it was tight. Parked pert near perfectly. A little crooked in the spot, but it worked.
Autopark systems across the industry are vision-based, hence why they require vehicles on either side or visible lines for the parking spot
 
I only heard about zipper merges a couple years ago. They're fine in most cases, but I know that I don't want to be dumped onto the shoulder if a high speed merge attempt at the end of the merge lane goes bad. Shoulders can have all manner of crap on them, including broken down cars. Or there may not even be a shoulder. The merge that bothers me on my drives ends in a curb, grass and trees. I don't want to go long on that merge, yet FSDb does a high speed run in the merge lane, getting over only at the end.
Zipper only works in/was designed for heavy, slow traffic.

Otherwise, yes, I tend to merge whenever there is an opening.
 
My second tesla - model Y (no fsd on this car) without USS just got update to 2023.6.9 which has autopark support

Downloading and installing...

Very excited.. I missed parking assist the most in my tight garage parking...yay!

View attachment 920832
I have a ‘23 MY with fsd but no uss. I’m still on 10.69.x.

Can someone clarify, does 11.3.2 include the new park assist or are they on separate sw branches?