FSD Beta v11.x

jebinc · Dec 27, 2022

Supcom said:
The encouraging thing about this vehicle getting V11.x is that it does not appear to have received the earlier V11, indicating that the new version is being rolled out a little wider. Hopefully, we will see a version released to public testers in the near furture.

Soon, probably, next week…

JulienW · Dec 27, 2022

jebinc said:
Soon, probably, next week…

...or so.

I suspect no rollout until mid or near the end of Jan and updated to a 2023 branch.

gsmith123 · Dec 27, 2022

Anyone want to speculate on what's going on with V11 and what sort of issues they're running in to?

What we know today is:
- They've been testing it pretty heavily internally, with 3+ iterations
- They're being more secretive than in the past
- Based on Elon's tweets they've thought it would be ready for a wider release "soon" quite a few times, and it hasn't

Some possible motivations for this:

- it's having a lot of issues on highways where we use production AP today
- it's having a few issues on highways, but they really want to completely get to parity with production AP before a public release
- it's having a few issues on highways, but Tesla views interventions on highways as substantially higher risk due to higher speeds
- for risk mitigation and/or PR reasons they want to scale back public testing of alpha-stage FSD features
- they want to big-bang release V11 single-stack to the whole fleet (not just the beta group), as the release to officially release city streets driving + single-stack on highway, so it needs to be perfect (doesn't make any sense to me for risk mitigation, but it's possible)

Early on in the FSD Beta program it was really rough, compared to today (yet they still released it to a limited set of beta testers). I wonder if they're risk appetite has changed? Perhaps due to the massive increase in scale of the FSD Beta program, or perhaps more perceived attention.

I wonder if they'll potentially take any of these paths:

- keep "single stack" feature flagged and not available externally, but otherwise keep releasing the latest builds to the existing test group
- do a very limited single-stack public release, to just the OG / very tenured beta testers (I'd love a peek at how it's going!)
- speed limit single-stack to 60 - 65 mph and let production AP handle driving faster than that (up to the current 85 mph limit)

Any wild speculation based on no data?

I have a divided highway near me that, for some reason, FSD stays engaged on. It's a windy-ish road with typical speeds around 60mph. Mostly two lanes. FSD Beta does quite well, actually. My main issues have been mild slowdowns on occluded curves (pretty understandable); nothing that would really explain the multi-month delays we've seen.

EVNow · Dec 27, 2022

gsmith123 said:
Anyone want to speculate on what's going on with V11 and what sort of issues they're running in to?

NOA has a disengagement rate of 1 in 100 miles now. V11 needs to be around that for them to release (i.e. on freeways).

They are probably tracking other metrics too.

BTW, just to clarify .... there are certain bugs that could give a "no go" to a release, but also quality metrics that need to be met for a "go" decision.

jebinc · Dec 27, 2022

EVNow said:
NOA has a disengagement rate of 1 in 100 miles now. V11 needs to be around that for them to release (i.e. on freeways).

They are probably tracking other metrics too.

BTW, just to clarify .... there are certain bugs that could give a "no go" to a release, but also quality metrics that need to be met for a "go" decision.

I suspect the wait will be much longer than days, unfortunately. And, when it does release, it will initially be much worse than current NOA/AP.

EVNow · Dec 27, 2022

jebinc said:
I suspect the wait will be much longer than days, unfortunately. And, when it does release, it will initially be much worse than current NOA/AP.

The way I look at it .... a LOT of the features of V11 have already been introduced in FSD Beta. Only the "single stack" remains .... while marginally useful, its not something I personally care much about at all.

- 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

jebinc · Dec 27, 2022

EVNow said:
The way I look at it .... a LOT of the features of V11 have already been introduced in FSD Beta. Only the "single stack" remains .... while marginally useful, its not something I personally care much about at all.

- 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

Understood, but my forecast remains.

AlanSubie4Life · Dec 27, 2022

Yeah not sure why there is all the anticipation for single stack. It just is not going to be that good at first, and prospects for way better (long term) are also pretty slim (AP is not that bad overall so significant improvements can happen but are hard to achieve).

I just want them to work on how to stop the car. I disengage FSD on virtually every downhill stop from above 40mph so that I can avoid unnecessary brake use and a pulsing stop. It’s pretty awful, has been for a while (it has never been good), and you can see it in all the videos of stops on downhills or flat from 40-50mph or above.

Maybe V11 will fix it but I don’t see why it would (except maybe they can figure out how to properly handle high speed control loops, which seems to be a major problem).

AlanSubie4Life · Dec 27, 2022

JulienW said:
Here you go and a couple of bonus including trying to run a red light.

Sadly I did not quite get the camera rolling in time (I like to have it rolling, so I can capture a repeat failure, but failed to have it turned on today, sadly), but tried to run a red on 2022.44.25.5 today. Same situation as before (prior video clip from older version linked below).

It requires the right combo of green and red lights, and no lead vehicles, and then sometimes it’ll just try to go through the red. I think this is probably somehow linked to the weird “two greens in a row slow behavior” that Chuck has tested before (the car exhibits this behavior on these double greens sometimes as well). Spans multiple releases over many months so it is a deep-rooted problem of some sort.

This is just the second time I have had it fail here - but I think mostly because the conditions required for failure rarely occur - I think it has only stopped properly on a couple occasions in the scenario of interest. In any case running a red light is bad!

Google Maps

Find local businesses, view maps and get driving directions in Google Maps.

maps.app.goo.gl

FSD Beta 10.69.2.3 Attempts to Run Red Light

Presumably confused by multiple lights?You can see my foot move to the brake pedal and the car continues to apply power (I was not pressing the accelerator d...

youtu.be

gsmith123 · Dec 27, 2022

Yeah not sure why there is all the anticipation for single stack. It just is not going to be that good at first, and prospects for way better (long term) are also pretty slim (AP is not that bad overall so significant improvements can happen but are hard to achieve).

Part of it is just curiosity - it's the next shiny new thing!

But I think there are some notable improvements (or at least potential improvements), actually.

1. Lane merges and splits. AP doesn't seem to understand the lane topology well and will center itself in the "big" lane. FSD is already a lot better at this.
2. Lane changes. AP lane changes are less smooth than FSD and it's slow to initiate them.
3. FSD city streets to AP handoff. I get all sorts of weird behaviors right a the handoffs.
4. On ramp trajectory and merge strategy. The car currently stays as far right as it can in merges and only merges in when it hits the end of the ramp. People smoothly merge in and bias towards getting into their target lane quickly.
6. More aggressive acceleration. AP is very very slow to speed up.
7. Kind of related to 1, but a better ability to plan and negotiate interactions with other vehicles. AP doesn't have much sophistication for planning it's lane change and merge interactions with other vehicles to smoothly yield or assert correctly to other vehicles.

Those are the big ones that I think people will notice. Number 1 in particular has been complained about for years.

There's a few other minor ones I'm interested in too:

- For me on curvy roads both FSD and AP drive very uncomfortably to the outside of curves even if there is "danger" there (i.e. a neighbor vehicle in that direction).
- Snow conditions. This is admittedly a long ways from being very trustworthy, but FSD beta will attempt almost any weather conditions while AP will not.

I fully expect regressions for a while, but disagree that "prospects for way better (long term) are also pretty slim". I see significant room for improvements but the trick is introducing more sophisticated highway AP without introducing phantom braking.

AlanSubie4Life · Dec 27, 2022

gsmith123 said:
FSD is already a lot better at this.

No it is not. On all the test cases I have tried it just stays centered. If a bike lane marking disappears it edges to the right, for example (see video). You can also look at how it behaves currently on unmarked residential and commercial streets - randomly centering to the right on occasion (no data from very latest release).

gsmith123 said:
AP is very very slow to speed up.

So is FSD!

gsmith123 said:
but disagree that "prospects for way better (long term) are also pretty slim". I see significant room for improvements but the trick is introducing more sophisticated highway AP without introducing phantom braking.

I mean, that’s one of the issues. A much more complicated perception and control system is going to have a lot of failures. Just look at FSD on surface streets! We would not want that reproduced on the highway, that’s for sure.

Of course I agree that there are shortcomings of AP - definitely they are. But generally it does well in simple situations and the concern is having FSD be able to duplicate that. Certainly the potential for it being better is there! But realizing that is a long road (as evidenced by the years of development so far).

Do I think that eventually it is possible to build a system better and more capable than AP? Yes.

gsmith123 · Dec 27, 2022

No it is not. On all the test cases I have tried it just stays centered.

Respectfully, my experience is that it is better in these situations. I know it depends so much on where you are, how the road markings are setup, etc so I'm certainly not doubting your experience, but this is an area I have high confidence that FSD is better than AP for my driving.

I did notice your video was from 10.69.2.4 and if I recall correctly the 10.69.3.x releases had improvements to to lane topology functionality so this might be something to retest.

You can also look at how it behaves currently on unmarked residential and commercial streets

To be clear, this wasn't the situation I'm talking about. It's still not good at unmarked roads. No disagreement there. I'm talking specifically about well marked transitions from 2 > 1 lane or vice versa.

I do broadly agree with your last point that FSD is more complicated and likely to have more failures, at least in the short term.

gsmith123 · Dec 27, 2022

For instance, splits like this are notably better on FSD:

AlanSubie4Life · Dec 27, 2022

gsmith123 said:
Respectfully, my experience is that it is better in these situations.

I suspect one reason that it feels better is that the re-centering happens more slowly at FSD speeds so it is less disturbing. But who knows. Maybe it's better. Worth rechecking on this version for sure. Certainly V11 should be able to fix this issue but we'll see. The track record for fixing that sort of problem is not good.

gsmith123 · Dec 27, 2022

I suspect one reason that it feels better is that the re-centering happens more slowly at FSD speeds so it is less disturbing.

That part is better too. I'm saying both are better:

- fewer splits and merges have poor pathing where it centers in the big lane
- when does have undesirable centering (mostly on splits), it recovers more smoothly

AlanSubie4Life · Dec 27, 2022

gsmith123 said:
For instance, splits like this are notably better on FSD:

View attachment 889319

Forks are different than merges. The issues I've observed (video is above) are at merges (and it tends to be one of the things people complain about since they're a bit more common, occurring at every onramp). It does seem like forks should be better based on observations with FSD Beta so far. And they do occur on freeways of course, and AP sometimes handles them poorly.

AlanSubie4Life · Dec 27, 2022

gsmith123 said:
when does have undesirable centering (mostly on splits), it recovers more smoothly

On one of your test cases, if safe to do so, try this at 80mph. Just to see how it does. I'm curious.

Mardak · Dec 27, 2022

gsmith123 said:
Lane merges and splits. AP doesn't seem to understand the lane topology well and will center itself in the "big" lane

I just did a bunch of interstate driving for the holidays, and the more general problem of unclear lane line understanding was definitely the top / most frequent issue with current Navigate on Autopilot. Most of the time it's okay as there wasn't another vehicle to interact with, but the visualization shows the confusion especially with a left-side merge where the displayed lane rapidly switches between a really wide merged lane vs a trained / imagined narrow lane, which was probably more of a NoA "short-term hack" to reduce unnecessary centering when the dashed lines disappear at a merge.

Similarly, faded lines on curvy mountain interstates has the legacy stack searching for the outer lane line often resulting in rapid changes to the perceived lane edges in the visualization as well as jerky centering that I ended up preventing by gripping the steering wheel tightly and avoiding the unnecessary movement. Even more basic is the legacy stack has single-camera heuristics to determine the adjacent lane to switch to, and faded lines often results in lane changes being unavailable, but FSD Beta's unified 360º view probably would have done the lane change just fine.

Even when lane lines are clearly visible, the legacy stack very much wants to stay centered, and this resulted in a collision warning when the adjacent vehicle was slowly drifting out of its lane during a curve. I believe FSD Beta's stack would have more gracefully shifted over still staying within the lane to give space to the other vehicle and thus avoiding the beeping that startled all the other passengers.

Mardak · Dec 27, 2022

EVNow said:
a LOT of the features of V11 have already been introduced in FSD Beta

This versioning scheme matches up with a lot of other software where new features are added as part of a "minor version" perhaps behind a feature flag or some other condition while a "major version" bump like for FSD Beta 11 would be "breaking" / changing existing expected behaviors by default. A lot of hard work happens to add features that aren't actually "minor" but getting them to a quality level that's good enough to completely replace the legacy highway stack is definitely "major."

More broadly, FSD Beta 11's single stack represents progress of FSD Beta being good enough to match the quality expected of production-tested Navigate on Autopilot. If things work out, it should also result in more features and capabilities as the legacy stack was mostly in maintenance mode with Autopilot team focused on FSD Beta. Similarly, this is a natural step in moving all compatible vehicles off of the legacy stack even for those with basic Autopilot without FSD Capability purchased.

rpiolends · Dec 27, 2022

FSD for MCU1 cars?

FSD Beta v11.x

Well-Known Member

Well-Known Member

Member

Well-Known Member

Well-Known Member

Well-Known Member

Well-Known Member

Efficiency Obsessed Member

Efficiency Obsessed Member

Member

Efficiency Obsessed Member

Member

Member

Efficiency Obsessed Member

Member

Efficiency Obsessed Member

Efficiency Obsessed Member

Active Member

Active Member

Member

Similar threads