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 is it random then? Like will some people just get missed for the version that has the Beta? Or does it force everyone through them in sequential order? Or do you just have to make sure to let it install when it’s trying to push the version with FSD and if you miss it you’re screwed for a while?
So. Say you sit in the car. Open up the Software menu on the touchscreen. Do a, “Check for Update” press.

You’ll get a notification that there’s a SW update available if there is one, and what version number that update is.

The car will normally download said update, Willy-nilly. But then, to install, the user has to OK the install.

If there’s no OK, there’s no install. And that’s how people can ignore an update. For example: Outside of FSD-b, there’s been owners who’ve ignored updates for years; like, for example, there was that big GUI change that some took exception to.

As far as when and what sequence an update gets offered: that’s pretty much up to Tesla. They don’t talk much, if at all, about their reasoning and procedures. What we all know about when and how has been based nearly completely upon observation.

Observation has noted that a new version gets pushed out, first, to Tesla employees; second, to various influencers; then, to a subset of the general public; then, the rest of the public. The whole process seems, typically, to take a couple of weeks. But we’ve seen Tesla shorten that up considerably at times. And it’s not unusual that, with a given point release, Tesla may suddenly halt downloads. That is suspected to happen if some bug or other is found; or maybe because Tesla finds some improvement in some internal release and figures they may as well halt the current load while they work on a better.

As I said, there’s very little communication from Tesla on their thought processes on the upgrades. Once in a while there’ll be a tweet from Musk or, even rarer, some internal development type might say something. Mostly people just wait and look at the version tracker web sites to look for new and updated changes on the subset of cars that have manually subscribed their cars to that service.

Finally, in the rare cases where some owner gets truly left behind from everyone else, it’s possible to contact Tesla and get them to force a download. Especially if a download truly failed, which happens from time to time. But they get pretty durn snippy if it’s somebody just trying to get first in queue.

It used to be possible to drive one’s car to a parking lot across from a Tesla Service Center; there were reports that a car so placed would automatically get connected to the SC’s wi-fi and would autoload the latest and greatest, ahead of the rest of the crowd. But the last I heard that avenue was closed a few years ago.

Personally, I’ve sometimes got a new load ahead of nearly anyone else; sometimes I’ve brought up the rear of the pack; but mostly it’s been in the middle, a couple-three days after the GA load gets released.

My advice: take a Tums and relax. The load will get there when it gets there.
 
So is it random then? Like will some people just get missed for the version that has the Beta? Or does it force everyone through them in sequential order? Or do you just have to make sure to let it install when it’s trying to push the version with FSD and if you miss it you’re screwed for a while?
Tesla has algos that spread OTA versions out as they test changes. They don't send everyone the latest and 'greatest' as it could contain minor errors/bugs or they could be testing some new external or internal feature they added. Look at a couple of TeslaFI screens where you can see it spread around long term and then the past several days. As well you can see pending downloads. TeslaFi.com Firmware Tracker

For those that do not have FSD but bought it or got a subscription they may be on a different 'forked' branch of the software. You can see that FSD 11.4.1 is corresponds to standard version 2023.7.5 numbers. If your non-FSD version is > 2023.7 then you are 'ahead' of FSD.

xOqzAHr.jpg


xQbpWsb.jpg
 
$15k is a lot - but so is $100k for a car.
I disagree. $115,000 is chump change. I give that much to the homeless for soup! I usually buy multiple versions of FSDp just to see which one works best or fails first. I'm buying Twitter next week to see if I can drive it off a cliff. So stop all your bellyaching and start playing again in reality.........
 
I know, brother! All those people who fall down? Actors!
Ha. If you ask physicists, they’ll tell you that the model is not the reality; the models are the models, and the models are designed to match observations. With good luck and refinement, a decent model can be used to predict things, like the motion of planets and such. But it’s not a law; and a better model that does a better job may always be around the corner. Even Newton said so, with respect to the underlying reality. Why things obey the rules of motion, he had no clue and said so. That his model worked, and better than the crystal spheres approach, there was no doubt.

But even he would have had a bit of a headache, I bet, if informed of the Eisensteinian approach that has everything moving in straight lines: but space is warped, so, to our senses, it looks like things are taking curved paths 😁 .
 
  • Like
Reactions: GSP and FSDtester#1
Ha. If you ask physicists, they’ll tell you that the model is not the reality; the models are the models, and the models are designed to match observations. With good luck and refinement, a decent model can be used to predict things, like the motion of planets and such. But it’s not a law; and a better model that does a better job may always be around the corner. Even Newton said so, with respect to the underlying reality. Why things obey the rules of motion, he had no clue and said so. That his model worked, and better than the crystal spheres approach, there was no doubt.

But even he would have had a bit of a headache, I bet, if informed of the Eisensteinian approach that has everything moving in straight lines: but space is warped, so, to our senses, it looks like things are taking curved paths 😁 .
Wait, you're saying there is a law that models have to fall down? That doesn't sound very professional.......
 
I suspect many passengers that are next to the big trucks appreciate it moving over. I *know* my spouse appreciates it.
The problem I have is the ping pong affect in rush hour traffic with all the trucks. Plus it moves over for smaller trucks when that really isn't necessary. What was wrong with the previous behavior which moved over just not as far?
 
It's the Goldilocks principle.

No reaction to a big truck was bad, scary to the Tesla driver.
Overreaction with jerk is bad, scary to everyone.
Maybe 11.4.x will have the Just Right reaction!
On that vein, I had a case where FSDb started to veer into a turn lane, recognized it fairly early and corrected back into the driving lane. All good except they tweaked the algorithm to make it less jerky meaning it didn’t straighten out quickly enough after correcting and overcorrected across the yellow line.
 
  • Like
Reactions: JHCCAZ
The problem I have is the ping pong affect in rush hour traffic with all the trucks. Plus it moves over for smaller trucks when that really isn't necessary. What was wrong with the previous behavior which moved over just not as far?
We were happy with the previous behavior/distance. I don't have rush hour traffic experience with it but will on this summer's road trip on some truck-heavy highways.
It is really too bad that Tesla doesn't give you an advanced screen to tweak/customize some of these settings (none|min|max). One really nice thing on FSD highway seems to be how smooth (gentle) it is when changing lanes or moving over within a lane. It appears to 'take its time' to be 'calm' about it.
 
  • Like
Reactions: pilotSteve
Some updates from my HW4 11.4.1 FSD experiences. From the perspective of someone new to FSD, when I first started using 11.4.1, it was indeed more work than driving myself. Fast forward to today, and driving with HW4 and FSD 11.4.1 is definitely less work than driving without it. The major difference is that I can now predict where there might be issues and be hyper vigilant under those circumstances, and relax under all other circumstances. So… what conditions make me hyper vigilant?:

1) Lanes that are ending, forcing a merge right or left. For some reason, most of the time, 11.4.1 merges very late, sometimes running into the emergency section of the lane that is ending.

2) Freeway exits. 11.4.1 has been in the high speed lane of the freeway, and cut across two lanes to try and make an exit. I now force the car into the correct lane using the turn signal so that it exits smoothly if it is in the incorrect lane. I do not give it much freedom under these circumstances.

3) Roundabouts. Twice, I have had 11.4.1 exit a roundabout on the wrong exit, even though the predicted path is displayed correctly, and I have had it become confused and hesitate when there are other cars in the roundabout. One time, it pulled into the roundabout and then balked even though it was fully in the roundabout. I could see no reason why it balked, but I ALWAYS cover the accelerator pedal when entering a roundabout so that I can “motivate” it to hesitate less. By the way, my FSD is set to “Average” aggressiveness.

4) Narrow two lane roads and streets, especially around corners. On numerous occasions, I have had 11.4.1 apply the brakes because it thought that the vehicle in the opposite lane of traffic is in my lane. It is disconcerting to have the brakes applied at 60 mph when they’re people behind me. It does not do this all the time, but it does it enough that I am hyper vigilant on narrow streets.

5). Stop signs. FSD 11.4.1 moves more quickly than I would like towards stop signs, waiting, it seems, until the last minute to brake. This makes me uncomfortable because I do not entirely trust FSD yet.

6) Driving in areas where there is no map. I live in a rural area, and FSD has troubles if it does not have a map to guide it. For example, my driveway is about .5 miles long, and FSD feels “apprehensive” when driving on my driveway… going extremely slowly, balking, etc. On one road there is a new roundabout which is not on the map, and FSD 11.4.1 balks at it every time, requiring me to use the accelerator to “force” it through.

The VAST majority of the time, FSD 11.4.1 on HW4 works flawlessly… so much so that I would not want to live without it. The $15k I spent which many suggested was as smart as “burning cash” turned out to be (so far, after one month) a great decision on my part.

I will report more findings as I have them, but so far, FSD 11.4.1 on HW4 has been an amazing experience.

Joe
 
4) Narrow two lane roads and streets, especially around corners. On numerous occasions, I have had 11.4.1 apply the brakes because it thought that the vehicle in the opposite lane of traffic is in my lane. It is disconcerting to have the brakes applied at 60 mph when they’re people behind me. It does not do this all the time, but it does it enough that I am hyper vigilant on narrow streets.
Joe, nice write up.

FYI, related to #4 ... there were several YouTubers that pointed out that problem and is fairly common it seems. It seemingly was referred to as a fix going into 11.4.2 that Elon tweeted:
According to Musk’s tweet, the new update “addresses some easy to fix bugs” in the current version, 11.4.1, that’s been giving testers a bit of trouble with “excess conservatism with narrow roads & with lane changes in heavy traffic.”
 
  • Like
Reactions: FSDtester#1
Since you're new to the beta just trying to give you some context on the issues you raise


1) Lanes that are ending, forcing a merge right or left. For some reason, most of the time, 11.4.1 merges very late, sometimes running into the emergency section of the lane that is ending.

This has been broken for pretty much ever.

Some folks hand-waive it away mumbling about proper zipper merging--which would be fine if it only did it when there was a bunch of traffic-- but it does it ALWAYS, even when there's literally nobody in adjacent lanes at all preventing merging over well before you're about to run off the road.


2) Freeway exits. 11.4.1 has been in the high speed lane of the freeway, and cut across two lanes to try and make an exit. I now force the car into the correct lane using the turn signal so that it exits smoothly if it is in the incorrect lane. I do not give it much freedom under these circumstances.


This has varied a bit over time, there's been some releases people complained it got over too early and they were stuck too long in slow traffic... Now it's the other way... overall it seems to be a lack of overall context/understanding of where it's likely to make the most sense in the future to merge over based on where the car is now and what's coming up ahead.



3) Roundabouts. Twice, I have had 11.4.1 exit a roundabout on the wrong exit, even though the predicted path is displayed correctly, and I have had it become confused and hesitate when there are other cars in the roundabout. One time, it pulled into the roundabout and then balked even though it was fully in the roundabout. I could see no reason why it balked, but I ALWAYS cover the accelerator pedal when entering a roundabout so that I can “motivate” it to hesitate less. By the way, my FSD is set to “Average” aggressiveness.

I've never had it take a wrong exit- but I've seen all the other behaviors.... often seen all of em at different times on the same roundabout on the same version so apart from generally the roundabout code not being great it seems heavily impacted by other conditional things (traffic, lighting, weather, etc).

I'd say the more recent versions handles roundabouts GENERALLY better than it did say a year ago but it's still shocking inconsistent and very much a situation to be vigilant on (but again, a year ago it was a situation to just disengage and do it yourself, so improvement at least)


4) Narrow two lane roads and streets, especially around corners. On numerous occasions, I have had 11.4.1 apply the brakes because it thought that the vehicle in the opposite lane of traffic is in my lane. It is disconcerting to have the brakes applied at 60 mph when they’re people behind me. It does not do this all the time, but it does it enough that I am hyper vigilant on narrow streets.

As others have mentioned this seems to be a now-known regression and will be improved in next point release


5). Stop signs. FSD 11.4.1 moves more quickly than I would like towards stop signs, waiting, it seems, until the last minute to brake. This makes me uncomfortable because I do not entirely trust FSD yet.

This seems another one where some people like it one way and others like it another and Tesla keeps adjusting so someone's always going to be unhappy.


6) Driving in areas where there is no map. I live in a rural area, and FSD has troubles if it does not have a map to guide it. For example, my driveway is about .5 miles long, and FSD feels “apprehensive” when driving on my driveway… going extremely slowly, balking, etc. On one road there is a new roundabout which is not on the map, and FSD 11.4.1 balks at it every time, requiring me to use the accelerator to “force” it through.

I found similar on older versions when I had a long unmapped driveway... Some allege you can improve this (and some parking lot behavior) by making edits to OSM (open street maps) but YMMV... (I did actually get good results doing this on a work parking lot to make smart summon a lot more useful- but not so much trying to add my driveway)

I think ultimately it's going to be over cautious because it can't "know" what's coming up beyond its limited vision without any mapping data and don't know there's really any "fix" for that other than using fleet data to augment the maps.


The VAST majority of the time, FSD 11.4.1 on HW4 works flawlessly… so much so that I would not want to live without it. The $15k I spent which many suggested was as smart as “burning cash” turned out to be (so far, after one month) a great decision on my part.

I will report more findings as I have them, but so far, FSD 11.4.1 on HW4 has been an amazing experience.

Joe

Very solid feedback BTW, look forward to more of your posts.
 
Once again, I don't think fsdb is using any downloaded road geometry.


Didn't we already go over this?

Full thread on this linked below- but he specifically cites parking lot outlines among the data it grabs per-route, as well as showing part of the output from the car including shape and edge data including, specifically "updating the properties of existing edges" for roads.

 
Narrow unmarked roads is routine where I drive and very common in New England so not really an edge case. FSD has improved a lot though with V11.
Yes - I was thinking of Boston when it came to narrow unmarked roads - but don't remember you posting that the car stops for every passing vehicle. So, I'm assuming its something peculiar to where these happen or perhaps a bit too narrow.
 
Once again, I don't think fsdb is using any downloaded road geometry. It's probably trained on many densely labeled parking lot geometries and can infer based on visual patterns:


That's interesting, thanks for posting. I was thinking of Summon today as I had to park my car a couple of blocks from home due emergency vehicles blocking access to our parking garage. I thought, "wouldn't it be nice to have the car come back on its own and park itself after the path is cleared?"

Then I was watching the video and thought, "I was thinking about summon earlier today, what a coinci..... Oh heck! I have to walk back and move my car home!"

So your post helped me not lose the car!
 
Didn't we already go over this?

Full thread on this linked below- but he specifically cites parking lot outlines among the data it grabs per-route, as well as showing part of the output from the car including shape and edge data including, specifically "updating the properties of existing edges" for roads.


That has nothing to do with the actual geometries within the parking lot. The example data he presents is basic metadata.

You can tell the parking lot geometry predictions on the visualization aren't part of a static map because of the fuzzy factor and how the fuzziness updates as you drive around the parking lot. It's the same sort of fuzzy predictions that Karpathy showed in his talks about birds eye view road geometry predictions.