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:
Some guy on Twitter seems to be downloading 11.4.4 under software version 2023.7.20. I’ve been stuck on 11.3.6 for a month now, and no updates since privacy settings on 2023.12.11
You are on a segregated Beta branch. You could say that 11.4.x is the BETA of FSD Beta and 11.3.6 is the final release candidate (to the "general public") of FSD Beta 11.3 version.

Your software 23.12.x is higher than 23.7.x so you could never get 23.7.20/11.4.4. Likely (but no one knows/conjecture) you will have to wait until 11.5.x is released and then your branch will likely be updated to the final 11.4.x version. But again no one knows for sure what the HELL Tesla will do.
 
Last edited:
Looks like it has new release notes. That’s promising.

Here's the full extract of 11.4.4 release notes, it's firmware version 2023.7.20:

- Improved short-deadline lane changes, to avoid going off-route, through better modeling of target lane vehicles to improve gap selection assertiveness.

- Improved offset consistency when controlling for static obstacles. Also improved smoothness when changing offset direction by adjusting speed more comfortably.

- Improved handling of oncoming cars on narrow unmarked roads by improving prediction of oncoming car's trajectory and leaving enough room for them to pass before re-centering.

- Improved Occupancy Flow prediction from the Occupancy Network for arbitrary moving obstacles by 8%.

- Expanded usage of the new object ground truth autolabeler for the NonVRU detection model, improving distant vehicle recall and geometry precision for semi-trucks, trailers, and exotic vehicles.

- Improved VRU control by expanding planning scope to control gently for low-confidence detections that may interfere with ego's path.

- Improved handling for VRUs near crosswalks by predicting their future intent more accurately. This was done by leveraging more kinematic data to improve association between crosswalks and VRUs.

- Improved ego's behavior near VRUs by tuning their assumed kinematic properties and utilizing available semantic information to classify more accurately their probability of intersecting ego's path.

- Improved Automatic Emergency Braking recall in response to cut-in vehicles and vehicles behind ego while reversing.
 
"- Improved short-deadline lane changes, to avoid going off-route, through better modeling of target lane vehicles to improve gap selection assertiveness."

I hope this translates to fewer instances of my car darting into a turning lane on the highway when the path it should take is clearly continuing on a straight path. That's been the biggest problem I've had in 11.4.3!
 
"- Improved short-deadline lane changes, to avoid going off-route, through better modeling of target lane vehicles to improve gap selection assertiveness."

I hope this translates to fewer instances of my car darting into a turning lane on the highway when the path it should take is clearly continuing on a straight path. That's been the biggest problem I've had in 11.4.3!
Sounds like it's more related to improved lane selection when target vehicles are nearing things like 2 lanes merging into one, onramps, offramps, ...

This might make even more hasty lane changing. I was hoping to see improvements for fewer needless turn signals and lane changes.

I agree with fixing those last second abrupt lane changes into bus lanes etc. Maybe next time?
 
Last edited:
  • Like
Reactions: FSDtester#1
"- Improved short-deadline lane changes, to avoid going off-route, through better modeling of target lane vehicles to improve gap selection assertiveness."

I hope this translates to fewer instances of my car darting into a turning lane on the highway when the path it should take is clearly continuing on a straight path. That's been the biggest problem I've had in 11.4.3!
It is sort of the opposite of that. It isn't about having fewer lane changes, it is about not missing a lane change that it needs, or at least thinks it needs. (Or probably more accurately fewer aborted lane changes, which is something I've had quite a few of recently.)
 
  • Like
Reactions: GSP
Sounds like 11.4.4 will try to ignore estimated paths for low confidence VRUs. 11.4.3 severely over reacted to those with crazy lane offsets (straddle lane lines) and slowing speed. Nothing like overreacting to a slow ambulating senior citizen calmly walking on a sidewalk parallel to traffic flow with a bike lane separation. o_O
 
"- Improved short-deadline lane changes, to avoid going off-route, through better modeling of target lane vehicles to improve gap selection assertiveness."

I hope this translates to fewer instances of my car darting into a turning lane on the highway when the path it should take is clearly continuing on a straight path. That's been the biggest problem I've had in 11.4.3!
That is unfortunately not what it means.
 
You are on a segregated Beta branch. You could say that 11.4.x is the BETA of FSD Beta and 11.3.6 is the final release candidate (to the "general public") of FSD Beta 11.3 version.
It is curious that Tesla has new release notes with 11.4.4 instead of just calling it 11.5. Maybe it is refining in preparation for a final 11.4.x release candidate for use in the main software branch similar to how 11.3.x got up to .6.

A lot of the release notes seems to focus on comfort and safety refinements as opposed to new types of functionality that might be saved for 11.5 version bump.
 
  • Like
Reactions: VanFriscia
At least those release notes mention narrow roads. I'll report back on whether it solves anything on marked roads, since that's been my issue. A bit concerning it only mentions unmarked roads.
Yep. It's a bummer the team prioritizes Chuck's complaints. His ridiculous residential streets should be low on the priority fix list given many more important issues as well as how the team overfits his solutions until the masses get artifact effects.

I suspect once they attempted addressing his unmarked narrow road issues funky results started on marked narrow roads.
 
  • Like
Reactions: dl_jehu
It is curious that Tesla has new release notes with 11.4.4 instead of just calling it 11.5. Maybe it is refining in preparation for a final 11.4.x release candidate for use in the main software branch similar to how 11.3.x got up to .6.

A lot of the release notes seems to focus on comfort and safety refinements as opposed to new types of functionality that might be saved for 11.5 version bump.
Well I sure WISH they would test 11.4.4 with 23.20.x on use. 🤔 🤣 🤣
 
I hope they improve this so as to not cause confusion on oncoming cars when Tesla is centered and moves over in a more timely manner. With 11.4.3, it is very uncomfortable when Tesla waits too long to move over.

Improved handling of oncoming cars on narrow unmarked roads by improving prediction of oncoming car's trajectory and leaving enough room for them to pass before re-centering.
 
Yep. It's a bummer the team prioritizes Chuck's complaints. His ridiculous residential streets should be low on the priority fix list given many more important issues as well as how the team overfits his solutions until the masses get artifact effects.

I suspect once they attempted addressing his unmarked narrow road issues funky results started on marked narrow roads.
I mean it's not just Chuck's complaints... Some of us were complaining about these things, too. The way FSDj handles narrower roads lately is absolutely horrific, and that's a significant part of getting FSDj to be... well... FSD.
 
I mean it's not just Chuck's complaints... Some of us were complaining about these things, too. The way FSDj handles narrower roads lately is absolutely horrific, and that's a significant part of getting FSDj to be... well... FSD.

I haven't tested those road types but suspect it's a design dilemma of needing vehicle separation margin for FSDj to safely respond to a potential head-on versus FSDj punting control back to the driver. Dollars to donuts the team removes the last minimum safety margin in narrow road scenarios since the driver is ultimately responsible and key to overall safety.
 
I haven't tested those road types but suspect it's a design dilemma of needing vehicle separation margin for FSDj to safely respond to a potential head-on versus FSDj punting control back to the driver. Dollars to donuts the team removes the last minimum safety margin in narrow road scenarios since the driver is ultimately responsible and key to overall safety.
They have to reduce the margin. You simply can't drive around many areas (where I live in Seattle, for example) without the car coming to a complete stop (0mph) over and over when another vehicle is coming toward you. This requires some kind of "Trust Neural Network" (TNN, I coined it)! I mean, humans trust each other not to swerve at the last second into you, on a daily basis. So you know, it needs to happen.
 
Last edited:
I was on a work trip over the weekend and took a few Uber rides...
It's interesting when you ride in a car with other drivers and see what their style of driving is.
FSDb is smoother and safer than any of the drivers I had.
I think that we are so used to our own driving styles that it's really difficult to adjust to others and that includes FSDb, and when we are essentially monitoring rather than driving, anything we perceive as an anomaly is heightened.
 
Yep. It's a bummer the team prioritizes Chuck's complaints. His ridiculous residential streets should be low on the priority fix list given many more important issues as well as how the team overfits his solutions until the masses get artifact effects.

I suspect once they attempted addressing his unmarked narrow road issues funky results started on marked narrow roads.
Not a bummer. This is a significant problem every day for me and I'm sure others too. Elon has said safety factors in on which problems get priority and suddenly stopping with a car behind is certainly a safety issue.