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:
Awfully cautious rollout for a patch with minimal new release notes, no?
In some sense, 11.4.5 was not good enough to release outside of employees, so 11.4.6 fixing these issues (presumably related to the new AEB for general obstacles) is already a slow rollout of 11.4.5 functionality delayed a couple weeks. And Tesla already knowing 11.4.5 was "bad" might be more cautious with 11.4.6 to confirm it's "good."

On the other hand, 11.4.6 rollout wider than employees also included some of us originally in the Safety Score group, so that's a little bit faster than previous rollouts.

Unclear if "Reduces unnecessary slowdowns" is relative to 11.4.4 or 11.4.5's newly introduced AEB?
 
  • Like
Reactions: willow_hiller
Awfully cautious rollout for a patch with minimal new release notes, no? I wonder if there's more changed than the release notes let on.
Tesla has probably did a lot of "under the hood tuning" to tighten up the bugs to make this a stable release so it can be made available to new HW4 buyers. A lot of engineers likely have their fingers crossed that this is the ONE for 11.4.x and looking closely for any bugs.

Hope to see some videos made by in the morning.
 
Awfully cautious rollout for a patch with minimal new release notes, no? I wonder if there's more changed than the release notes let on.
There are always more changes than the notes let on. I think it is fairly aggressive, they normally limit the first few days of public access to the initial influencers that got it. But people with no social media presence, like me, got this update on the first public day.

And as @Mardak showed Elon mentioned that it is supposed to reduce unnecessary slowdowns.

A little over an hour before I get to take it for a quick spin to see what I think of it.
 
View attachment 959604View attachment 959604This sign defeats FSDb every time. There must be some conflict between vision and map data. If I have nav on it stutters before blowing through, if I just have vision, it blows through full speed ahead.
The car is more reliant on vision than map data...so that's the perfect sign to defeat it. I've had Trucks with stop signs or speed limit signs fool FSDB.
 
  • Like
Reactions: fasteddie7
I don't get 11.4.4. The first few days, at a time when I needed it most (on the way to the hospital behind the ambulance when my father passed away), and for a couple of days thereafter, it was way better than any previous version, with almost zero disengagements.

Then, over the weekend, it was like somebody threw a light switch. Now, it is driving like a drunken chimpanzee, repeatedly trying to change lanes to the right for a left turn, trying to use bike lanes, and generally being a menace to everything on the road, requiring five or six interventions per drive, consistently missing lane changes, etc.

It almost makes me wonder if it does well only on certain roads for which it has lots of training data, and completely falls over on the others, in which case there's probably some serious over-fitting going on.
 
Chuck Cook should be considered an honorary employee, and a few of you guys and gals too.
Everytime there is a point release, we, for the most part, have a subjective feeling that it is mostly better, but hard to quantify.
It's really hard to judge point to point releases, but year to year there is progress.
I'm still waiting for that "holy sh?t!" release.
 
I don't get 11.4.4. The first few days, at a time when I needed it most (on the way to the hospital behind the ambulance when my father passed away), and for a couple of days thereafter, it was way better than any previous version, with almost zero disengagements.

Then, over the weekend, it was like somebody threw a light switch. Now, it is driving like a drunken chimpanzee, repeatedly trying to change lanes to the right for a left turn, trying to use bike lanes, and generally being a menace to everything on the road, requiring five or six interventions per drive, consistently missing lane changes, etc.

It almost makes me wonder if it does well only on certain roads for which it has lots of training data, and completely falls over on the others, in which case there's probably some serious over-fitting going on.
The variability occurs even on the same roads/routes.

Today I noticed that in a place where I disengage to correct for wrong lane/exit it didn't even ask me about why I disengaged. I thought they had disabled the voice memos but later at different disengagement it asked why. On the way home the only interventions I did were speed limit adjustments. In many drives it ends with wheel violently turning left and right as approaches my driveway. Today (as in other times) it just slowed to a stop.

Another trip later in different vehicle but same software built was uneventful except for a dive into left turn lanes when it should have been going straight. I've experienced that same dive before.

The car seems to "flinch" late for things that is sees. By the time it has reacted the thing has long past. This is also not new. Basically still higher latency than me. It also seems to be overly cautions for things that are relatively far away (at least to us humans). I suspect the system tuned very conservatively still.

There are also clear map data induced issues. For example, turning left from east bound Prospect Rd to north bound Lawrence expressway results in a slowdown. I can see on the display that car thinks the speed limit is 35 mph. It doesn't speed up until it gets to the 50 Mph sign.
 
  • Like
Reactions: Jeff N
Ok, here is my report on 11.4.6, which seems better than 11.4.4 based on the ~13 mile drive today:
  • It seems to respond to what it thinks an object is going to do more than it used to:
    • It slowed for a pedestrian that looked like they were going to walk across the road, but as soon as they stopped it resumed normal speed.
    • It slowed slightly for a car that slightly moved turning right in front of me, but didn't resume normal speed when they stopped until it was almost passed them.
    • It responded to a car that started to turn left out of a parking lot in front of it. The car in the lane to my right didn't stop so the person in the parking lot stopped, but just before they stopped FSDb "emergency stopped". (If I was driving I'm not sure if I would have stopped or kept going.)
  • It only made one unnecessary lane change. (On the route it drove today there would have normally been 4-6 unnecessary lane changes.)
  • My insurance safety score app dinged it for three events:
    • Two hard braking events. (Obviously for the "emergency stop", but the other one it showed I don't recall at all, so it must have been just over their threshold.)
    • One excessive acceleration event, just after making a right turn on red.
  • I will say that the acceleration seemed a little more digital. (Meaning quicker acceleration when it wanted to speed up, I think it was smoother on 11.4.4, but it could have just been the vehicle we were following was accelerating quicker than what we normally follow and FSDb was just matching that.)
  • Zero interventions.
  • One disengagement: But that was only because I had my hand on the steering wheel during the "emergency stop" and the stopping force threw me forward enough that the torque on the wheel kicked it out of FSD.
My passenger didn't know FSDb was driving. (Normally during a hard stop like that they ask why it did that, but they saw the car starting to turn in front of us and understood why it/I made the stop.)
 
Last edited:
I don't get 11.4.4. The first few days, at a time when I needed it most (on the way to the hospital behind the ambulance when my father passed away), and for a couple of days thereafter, it was way better than any previous version, with almost zero disengagements.

Then, over the weekend, it was like somebody threw a light switch. Now, it is driving like a drunken chimpanzee, repeatedly trying to change lanes to the right for a left turn, trying to use bike lanes, and generally being a menace to everything on the road, requiring five or six interventions per drive, consistently missing lane changes, etc.

It almost makes me wonder if it does well only on certain roads for which it has lots of training data, and completely falls over on the others, in which case there's probably some serious over-fitting going on.
I got a lot of the drunken chimp behavior today. Much of it on the same roads I drive on every day. Different behavior today for some reason. Something is causing it to change its behavior. The software hasn't changed, the lighting hasn't changed, the temperature is about the same. Are there background map updates we don't know about? I can't figure out what is causing it to act differently all of a sudden….
Edit: it just occurred to me what it could be. Several years ago Greentheonly said that specific cars would have instructions to send back particular types of data. So even though they all have the same software they could have different situations to flag. If this is so, when Tesla finds a car that regularly encounters a situation they need data on they could send it instructions on what to do. So the software could be slightly modified by Tesla between system updates and that might cause slightly different behaviors.
 
Last edited:
Another video. Version 11.4.6 looks like major regression.
FSD 11.4.6

It looks like this driver would prefer the vehicle to stop and wait at intersections for the traffic to fully clear. I don't think his risk tolerance really reflects on FSD Beta's performance.

Here's the first frame where he first disengaged:

1690377329144.png


And then it's another 56 frames until the truck is even visible on the edge of the visualization:

1690377480024.png


And a total of 92 frames until it would have overlapped with his vehicle:

1690377563851.png


At 30 frames per second and 17 MPH max, that's a total of ~75 feet his vehicle could have moved within that time period. And his disengagement report is "Unsafe unprotected left. Almost got hit by a truck." If anything, he's the one that made an unsafe disengagement.
 
Chuck Cook should be considered an honorary employee, and a few of you guys and gals too.
Everytime there is a point release, we, for the most part, have a subjective feeling that it is mostly better, but hard to quantify.
It's really hard to judge point to point releases, but year to year there is progress.
I'm still waiting for that "holy sh?t!" release.
Looking at FSD Beta history, not sure there will be a holy *sugar* release. Every release has had some improvements and some regressions. When I look back at when I started testing at 10.3/10.3.1, in some aspects it has come a long way and other aspects it's just as inconsistent. So far 11.3.6 was the most consistent as far as disengagements, but they continuously are having to move targets and redesign things as they develop which creates regressions, some very dangerous. From my perspective, which I do love FSD Beta, it's a few years out from a very consistent level 2 ADAS. I don't know that I see a pathway, currently, for true level 5.
 
Looking at FSD Beta history, not sure there will be a holy *sugar* release. Every release has had some improvements and some regressions. When I look back at when I started testing at 10.3/10.3.1, in some aspects it has come a long way and other aspects it's just as inconsistent. So far 11.3.6 was the most consistent as far as disengagements, but they continuously are having to move targets and redesign things as they develop which creates regressions, some very dangerous. From my perspective, which I do love FSD Beta, it's a few years out from a very consistent level 2 ADAS. I don't know that I see a pathway, currently, for true level 5.
Agree, 100%
 
  • Like
Reactions: Pdubs and spacecoin