FSD Beta v11.x

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.

Deep Understanding Tesla FSD Part 1: HydraNet

From Theory to Reality, Analyze the Evolution of Tesla Full Self-Driving

saneryee-studio.medium.com

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).

Twelve Questions about Tesla Autopilot Technologies Exposed at 2021 AI Day

On Aug. 19 2021, Telsa held an AI day , exposing more technologies about Autopilot , from perception, planning, data labeling, simulation…

yuhuang-63908.medium.com

 

[Yelobird]

Added another Tesla thing to the beta list. Powerwall Beta. My lights flicker and 3 UPS units don’t register powerwall power because of the frequency being set too high. Support Beta comms below.

View attachment 994933

View attachment 994934

View attachment 994935

View attachment 994936

View attachment 994937

And this means what in the FSD thread?

 

[aronth5]

Regarding the placement of the b-pillar camera-
I did some measurements on my Model Y just now. It's definitely shorter than 9 feet from the camera to the nose but that measurement is really irrelevant. What's relevant when comparing to a human driver is the difference in sight lines. When I'm sitting in a normal driving position (for me) there's less than 6" difference between the B pillar camera and my eyes. If I lean forward as I might when trying to peer around a corner it's closer to 18". If I really crane as far forward as I can so my head is hitting the steering wheel it's about 24". Does this make a difference? Yes. Is it significant? Not really, IME.

The bigger issue has to do with processing, perception and lag. On a very frequent basis my car starts creeping forward towards traffic even when the sight lines are totally clear and there's no need to do so. This is disconcerting to the approaching traffic as they think I'm about to pull out. It's also less safe because it's putting the nose of my car closer to the cross traffic unnecessarily.

The other big issue is the lag - very frequently it seems to need to 'think' about the traffic before acting. With blind intersections such as these you need to act decisively. Waffling just makes a bad intersection worse.

Finally, I'll add that while people may be correct that these are poorly designed and dangerous intersections, the fact is that they exist so FSD needs to handle them.

Bottom line for me
When monitoring FSD when you have to lean forward because of an obstructed view and you cannot safely creep any further FSD is compromised. Simple.

For many people this may never occur or occur infrequently. For me it is every day.

 

[aataskin]

I noticed the a Tesla was a 2019. Were the Audi and Ford equally as old?

Point is that a much older Model S that he brings out around minute 20 is just as good as these newer cars.

 

[aronth5]

Saw a post on X confirming someone just received 2023.38.9. (11.4.4). Did a software check and am now downloading the update. Coming from 2023.32.9.

Finally something worth while. FSD now only needs a single pull to initiate or disengage.

 

[kabin]

Added another Tesla thing to the beta list. Powerwall Beta. My lights flicker and 3 UPS units don’t register powerwall power because of the frequency being set too high. Support Beta comms below.

View attachment 994933

View attachment 994934

View attachment 994935

View attachment 994936

View attachment 994937

Gotta wonder what their next higher priority level is? Mind blowing priority?

 

[AlanSubie4Life]

Elon still making wild unsupported claims about safety. FSD v11 makes it most places point to point without intervention…hmmm…not my experience.

https://x.com/chazman/status/1730255569305424011?s=20

I think he also thinks AGI in 3 years or something (so big sigh of relief there - we probably have a decade at least).

 

[Ramphex]

And this means what in the FSD thread?

Sorry, let’s go back to beating the dead horse of a software that hasn’t been updated in like 6 months

 

[jebinc]

Sorry, let’s go back to beating the dead horse of a software that hasn’t been updated in like 6 months

You beat that damn horse!

 

[Ramphex]

Beat that damn horse!

I still get dry wipes and failure to wipe during light rain. Behavior is so unpredictable.

 

[jebinc]

I still get dry wipes and failure to wipe during light rain. Behavior is so unpredictable.

Actually, it's very predicable. When wipers are set to "Auto," you can predict with accuracy that they will not work as intended!

 

[Mardak]

Elon still making wild unsupported claims about safety

"The data is unequivocal that supervised Full Self-Driving is somewhere around 4 times safer (maybe more) than humans just driving themselves."

The most recent data from Tesla's Vehicle Safety Report Q4 2022 says US average crash every 652k miles (on any roads) while Autopilot (mostly highway) is 4.85M miles for a 7.4x "safer." Given that Tesla is publishing this data for Autopilot in general, they might be using these same metrics for FSD Beta.

Does 4 times safer seem like a reasonable "drop" from 7.4x? Of course, FSD Beta usage is biased to where people actually activate it especially with 11.x allowing usage on freeways, so it's still probably not that useful to compare against US average for all miles driven.

If we say FSD Beta 10.x averaged around 10M miles/month while 11.x was around 60M, can we say anything meaningful that the additional 50M is mostly from "easier" highway driving? And if so, can we estimate what city streets driving safety would be by removing highway? Hopefully we'll get updated safety numbers for 2023 soon with more details.

 

[AlanSubie4Life]

will the pounding never stop

Not with Elon around.

The most recent data from Tesla's Vehicle Safety Report Q4 2022 says US average crash every 652k miles (on any roads) while Autopilot (mostly highway) is 4.85M miles for a 7.4x "safer." Given that Tesla is publishing this data for Autopilot in general, they might be using these same metrics for FSD Beta.

Hopefully we'll get updated safety numbers for 2023 soon with more details.

Yeah pretty useless, wouldn’t count on useful data unfortunately.

 

[sleepydoc]

Bottom line for me
When monitoring FSD when you have to lean forward because of an obstructed view and you cannot safely creep any further FSD is compromised. Simple.

For many people this may never occur or occur infrequently. For me it is every day.

I dont' doubt that the intersection is an issue - just that the problem is the position of the camera.

 

[ArchHamster]

I'm not sure if it's been posted yet, but apparently the new cybertrucks are shipping with 'v12'

I assume that's UI v12, or none-fsd beta v12, but one can dream.

 

[jebinc]

I'm not sure if it's been posted yet, but apparently the new cybertrucks are shipping with 'v12'

View attachment 995244

I assume that's UI v12, or none-fsd beta v12, but one can dream.

It’s possible as CT deliveries are reported on Tesla’s order page as either “in 2025,” or “in 2024,” depending on variant ordered.

 

[Supcom]

I'm not sure if it's been posted yet, but apparently the new cybertrucks are shipping with 'v12'

View attachment 995244

I assume that's UI v12, or none-fsd beta v12, but one can dream.

Yes, that is the UI, not FSD version.

 

[kabin]

Elon still making wild unsupported claims about safety. FSD v11 makes it most places point to point without intervention…hmmm…not my experience.

No interventions? Yea, maybe for those with a death wish or an FSD youtube channel but even some of the latter are starting to come to their senses.

Elon will continue the FSD fake it until you make it plan.

 

[Mardak]

Anybody else any improvements with NA-2023.44-14828? This morning I noticed FSD Beta 11.4.4 not slowing down anymore for a bridge that is 35 mph where previously it would think it's 30 mph when passing the cross street below for about 100 feet before seeing another posted speed limit sign. Pretty sure I've never reported this with voice notes, but I manually adjust the set speed back up pretty much every day except today. Unclear if Tesla is detecting the set speed changes or generally from the fleet reporting back vision-detected speed limits or could just be upstream map data provider corrected the issue.

From other quick testing on normal routes this morning, 11.4.4 even with the map update unnecessarily switches out of lanes, but maybe this isn't surprising as if it was fixed, it would probably be included in the per-trip navigation route data and not needing to wait for the large regional map update.

 

[JulienW]

Did get another hint at 11.4.8.69 last night. Maybe it will make it in time for the holiday update.

 

[ChaoticNeutral]

Saw a post on X confirming someone just received 2023.38.9. (11.4.4). Did a software check and am now downloading the update. Coming from 2023.32.9.

My car won’t even check for new software. It last checked on Nov 20th… when did they break the every 24hr check? (I’ve tried rebooting to no avail)