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

 

[Mardak]

V11 still randomly and oddly uses the signal

Would you classify these 11.3.2 turn signal usages as odd? These first 2 are to continue straight through the intersection where a single lane forks into multiple.

Chuck Cook signal left to stay in the middle straight lane:

Black Tesla signal right to stay in the right straight lane:

I think most people typically wouldn't signal for these lanes splitting from one, but the latter video actually praises the turn signal usage here, and it potentially could have confused the oncoming vehicle waiting to make an unprotected left turn thinking the Tesla was making a right.

To be clear, this signaling for forks is different from those to get in/out of turn-only lanes (often confused by inaccurate map data) accompanied by the "Changing lanes to follow route" messages like this for TesLatino where 11.3.2 realizes the left-most lane it's currently in forks to a new leftmost lane and cancels the turn signal:

 

[swedge]

I think what is frustrating for me is that I know how all this stuff works and we would never give customers a release date that doesn't take all those factors into account. It's almost like Elon doesn't understand how traditional software development process works.

Perhaps true. But then most traditional software development programs didn’t have NHTSA, DMV, O’Dowd, WaPo trolling them, nor crowd-sourced video for AI training. Nor were working on real-time AI video situation recognition.

A friend worked on software for insulin pumps, think FDA pre-approval required. That was years ago, and closing the loop with continuous blood glucose monitors is still a work in progress.

I think Tesla is moving damn quick, though not as quick as Elon or we would wish.

 

[cowleggies]

Other than those in early access, your original Safety Score or date of original FSD Beta do not seem relevant to the randomized rollout, which for 11.3.2 so far seems to be around 10%. In fact, those who have had FSD Beta since 10.2 are actually more likely to be in the 90% group still waiting vs long ago back in October 2021 when they/we were the majority of the test fleet getting these updates. But 10% of the 400k fleet eligible for FSD Beta is still 40k vehicles already.

I don't expect I'll be waiting long given what the rollout ramp has looked like in the last 48 hours - I'm just particularly impatient this time around to give v11 a spin

 

[navels]

Okay I can now understand this reticence to "burn" one's daily update

I checked yesterday at 10am so had to wait until 10am today to check again (and did not get it), which swayed me towards working from home instead of going into the office (not a crisis). Now I wish my update clock was reset at 7am instead of 10 . . . the FOMO is real.

 

[Supcom]

Would you classify these 11.3.2 turn signal usages as odd? These first 2 are to continue straight through the intersection where a single lane forks into multiple.

Chuck Cook signal left to stay in the middle straight lane:
View attachment 919561

Black Tesla signal right to stay in the right straight lane:
View attachment 919562

I think most people typically wouldn't signal for these lanes splitting from one, but the latter video actually praises the turn signal usage here, and it potentially could have confused the oncoming vehicle waiting to make an unprotected left turn thinking the Tesla was making a right.

To be clear, this signaling for forks is different from those to get in/out of turn-only lanes (often confused by inaccurate map data) accompanied by the "Changing lanes to follow route" messages like this for TesLatino where 11.3.2 realizes the left-most lane it's currently in forks to a new leftmost lane and cancels the turn signal:
View attachment 919563

FSDb has been doing this for the last couple of releases. It seems to be an overreaction to when it did not signal at all when it would move into a turn lane. Now it tends to overuse the signals. There certainly needs to be a bit of 'polish' applied here. But, I would rather the team work on roundabout and right turn yield issues if there's any choice.

 

[JB47394]

Perhaps true. But then most traditional software development programs didn’t have NHTSA, DMV, O’Dowd, WaPo trolling them, nor crowd-sourced video for AI training. Nor were working on real-time AI video situation recognition.

I think the more important delay comes from knowing that this is a safety-critical system which, if it has a serious flaw for even one version, could end the automation program and potentially cripple the company. I attribute FSDb's conservatism and hesitancy to that. It's okay to get into an accident, but it's not okay to be the cause of an accident. When in doubt, slam on the brakes. The law says that the guy behind you is responsible for the crash.

FSDb has been doing this for the last couple of releases. It seems to be an overreaction to when it did not signal at all when it would move into a turn lane. Now it tends to overuse the signals. There certainly needs to be a bit of 'polish' applied here. But, I would rather the team work on roundabout and right turn yield issues if there's any choice.

As a software engineer, I'm really looking forward to the book that explains all this apparent insanity. There are so many cases where the car can do something perfectly well, then other cases where it does that same thing like a complete moron. Why? Map data? Perception issues? Procedural problems? Data jitter? Interpolation errors?

Patiently waiting for my turn in the update queue...

 

[Twiglett]

FSDb has been doing this for the last couple of releases. It seems to be an overreaction to when it did not signal at all when it would move into a turn lane. Now it tends to overuse the signals. There certainly needs to be a bit of 'polish' applied here. But, I would rather the team work on roundabout and right turn yield issues if there's any choice.

yesterday my car signaled left 30ft from a right turn it needed to take, so seeing this kind of odd signaling in V11 doesn't give me warm fuzzy feelings for highways.

 

[Todd Burch]

I think what is frustrating for me is that I know how all this stuff works and we would never give customers a release date that doesn't take all those factors into account. It's almost like Elon doesn't understand how traditional software development process works.

He never gave a release date. He'll say something on Twitter very literally..."rolled out"....and people will interpret it the way they want to. It probably rolled out to internal testers who work for Tesla on 11/11 as he said. They encountered problems, had to rewrite stuff, retrain, etc...whatever...and things changed.

That's not a failed promise on his end, nor is it a lie.

 

[uscbucsfan]

He never gave a release date. He'll say something on Twitter very literally..."rolled out"....and people will interpret it the way they want to. It probably rolled out to internal testers who work for Tesla on 11/11 as he said. They encountered problems, had to rewrite stuff, retrain, etc...whatever...and things changed.

That's not a failed promise on his end, nor is it a lie.

There's no way to really spin Elon's timelines. He's missed many more than he's hit, especially regarding FSD V11. February of last year, then "End of Summer", then November, then Christmas, etc. While there definitely is internal testing going on, many of these he spoke of public testing which never happened until recently.

He says it's because he gets excited, but at this point he even makes fun of himself. I can't tell if it's to boost hype or motivate internally or he's still just horrible with timelines.

 

[Acisplat]

He never gave a release date. He'll say something on Twitter very literally..."rolled out"....and people will interpret it the way they want to. It probably rolled out to internal testers who work for Tesla on 11/11 as he said. They encountered problems, had to rewrite stuff, retrain, etc...whatever...and things changed.

That's not a failed promise on his end, nor is it a lie.

https://twitter.com/x/status/1592692576142135302

https://twitter.com/x/status/1595433301363527681

This is a failed promise.

 

[Supcom]

There's no way to really spin Elon's timelines. He's missed many more than he's hit, especially regarding FSD V11. February of last year, then "End of Summer", then November, then Christmas, etc. While there definitely is internal testing going on, many of these he spoke of public testing which never happened until recently.

He says it's because he gets excited, but at this point he even makes fun of himself. I can't tell if it's to boost hype or motivate internally or he's still just horrible with timelines.

When developing new products, it's actually very difficult for people to accurately estimate the time and effort required. Humans have a tendency to minimize the potential roadblocks. Our ability to even identify the roadblocks is generally poor. Even if the task is broken down to small subtasks, we tend to estimate the effort for each subtask assuming a nominal effort for each.

As a result, when issues arise, as the always do, their impacts on the development effort has almost always been underestimated.

 

[LowlyOilBurner]

Overall good day of drives on v11.3.2. A few places where poor lane choices caused me to take over worked perfect, car moved into the correct lane ahead of time, not 100ft from the left or right turn lane it’s supposed to get into. This is surface streets/roads.

Highway lane changes are much faster, only complaint is when merging onto the highway from an on ramp, it doesn’t merge quick enough. It uses up the whole on ramp lane, and moves onto the highway at the last second. Maybe it’s because no other cars were around? I don’t know, not enough highway driving today to really pick up on patterns.

The last update that was this decent was back on 10.69.3.2. The .25 ones were big regressions for me.

 

[zoomer0056]

I don’t know how it works, but I imagine fancy coding…Or are you asking for a relative grading of how v11.3.2 works compared to 10.69?

Here's the answer: Wiki - FSD Beta v11.x

 

[uscbucsfan]

When developing new products, it's actually very difficult for people to accurately estimate the time and effort required. Humans have a tendency to minimize the potential roadblocks. Our ability to even identify the roadblocks is generally poor. Even if the task is broken down to small subtasks, we tend to estimate the effort for each subtask assuming a nominal effort for each.

As a result, when issues arise, as the always do, their impacts on the development effort has almost always been underestimated.

Elon's timelines and the inability to meet them obviously hurt with the public. Add to that he sometimes is outright wrong it exacerbates things; yoke horn from November 2021 and forward, Teslas never needing new brakes over the life of the car, "open" patients, flying roadster, etc.

Some of these things were never true and maybe he didn't know better, but there are very few other CEOs publicly making false statements and proposing timelines that are really a joke, but you take the good with the bad and with him, Tesla wouldn't have been anything near what it is...if it even got off the ground.

 

[JulienW]

This is a failed promise.

....but Thanksgiving is still 8 months away. Looks like you owe Elon an apology.

 

[JHCCAZ]

...But seriously: the 12 principles of agile software development as shown here 12 Principles Behind the Agile Manifesto |
I suspect most modern programmers-- excuse, software engineers-- would do better to read a good book on patterns and anti-patterns. Then again: old fart.

My career included one major acquisition by a much larger company, with all the attendant difficulties and culture battles that we were trying to work through. At one point, my new boss gifted me a book called
Rradical Innovation.

The title intrigued me but the book sat there in my office for a week or two while I was thinking about when I could sit down to read through it.

Then, an email came down to us from the same executive boss, announcing that any and all side idea investigations should be deferred, with no hours spent until the innovative idea could be written up, reviewed and approved by the new Innovation Council committee at company headquarters.

Fortunately this new committee didn't last long, but after that I never opened that book on Radical Innovation. I wonder if it was a good book, and I wonder if it recommended establishment of a Radical Innovation ldea Review Council.

 

[swedge]

Okay I can now understand this reticence to "burn" one's daily update

I checked yesterday at 10am so had to wait until 10am today to check again (and did not get it), which swayed me towards working from home instead of going into the office (not a crisis). Now I wish my update clock was reset at 7am instead of 10 . . . the FOMO is real.

Hmmm.... I'm trying to decide when to check again. My last time was 3pm yesterday, so I could do it now. But TeslaFi shows around 605 total downloads and pending, and that number hasn't grown much since morning. So, perhaps today's wave is over, and I should wait till tomorrow morning before checking again.

I'm not worried about Missing Out. It is more that I'm tired of AutoPilot doing the same annoying (sometimes scary) swerve maneuver each time we pass an on-ramp lane merge. That and the several places where I need to intervene every time I leave the house. I, and many others, reported these particular problems months ago with only some improvement. So the long delay since the last release in Dec wore thin long ago.

So, I'm hoping for a download tomorrow... Fingers crossed, still.

 

[Acisplat]

Hmmm.... I'm trying to decide when to check again. My last time was 3pm yesterday, so I could do it now. But TeslaFi shows around 605 total downloads and pending, and that number hasn't grown much since morning. So, perhaps today's wave is over, and I should wait till tomorrow morning before checking again.

I'm not worried about Missing Out. It is more that I'm tired of AutoPilot doing the same annoying (sometimes scary) swerve maneuver each time we pass an on-ramp lane merge. That and the several places where I need to intervene every time I leave the house. I, and many others, reported these particular problems months ago with only some improvement. So the long delay since the last release in Dec wore thin long ago.

So, I'm hoping for a download tomorrow... Fingers crossed, still.

I would say wait until there's at least a 300-500 pending car bump
It's anyone's guess when more rollouts will be happening

 

[zoomer0056]

Pretty sure Tesla FSD team ain't agile, then. Agile principle #1: "Our highest priority is to satisfy the customer through early and continuous delivery of valuable software."

But seriously: the 12 principles of agile software development as shown here 12 Principles Behind the Agile Manifesto | Agile Alliance

...semm kinda, I dunno, nebulous? Take Agile principle #7: "Working software is the primary measure of progress." I mean, I started my professional career in the early 1980s and that was the metric then, too-- and this was procedural programming in Pascal and 68K assembler.

I suspect most modern programmers-- excuse, software engineers-- would do better to read a good book on patterns and anti-patterns. Then again: old fart.

My Service Center has a window that you can see their whole service floor. I was looking and all the techs gathered in a circle and spoke one at a time. I had nightmares that night of never ending SCRUM meetings. Those days are over for me, whew!

 

[archae86]

First drive on 2022.45.11/FSDb 11.3.2:

Overall--it was better for me. Smoother in steering, throttle, and braking, especially at negotiating intersections.

I had zero freeway content, a mix of major arterial streets and slightly less major streets.

Big plus (for me) the deceleration in response to my moving the max down by scroll wheel was much more than in the past. Now it is enough to be useful. Yay!

Not a big plus--the dwell time at stop signs was appreciably up, and, yes, I routinely saw 0 mph displayed.

Worst minus--as seen in some of the influencer videos, I had it make a turn timed well to fit in traffic and with good initial throttle choice, only to suddenly go anemic just round the turn. Easy to adjust with a throttle nudge, but very annoying, and potentially dangerous to the inattentive.