Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

FSD Beta 10.69

This site may earn commission on affiliate links.
Just a note of reality/depression for all us 10.12.2ers hoping for 10.69.1.1. Looks like 10.69.1.1 may have petered out and we must face the reality of "...probably end of the week..." and however long that is before it starts the slow several weeks rollout process.

Screen Shot 2022-09-06 at 12.48.02 PM.png
 
The irony. Just drove a 125 mile trip on 10.69.1.1 and has “poor weather detected” and wasn’t even able to see how it would work 🤦🏻‍♂️🤣
FSD beta used to enable in heavier rain (see Chuck's old videos), but now that FSD beta is under more scrutiny, Tesla has decided to be more conservative.

100% so. I had a 125 mile trip each way on Labor Day. Wasn't bad weather, mostly drizzle, and had "full self-driving unavailable" for almost the entire trip.
 
  • Funny
Reactions: Ramphex
The irony. Just drove a 125 mile trip on 10.69.1.1 and has “poor weather detected” and wasn’t even able to see how it would work 🤦🏻‍♂️🤣

100% so. I had a 125 mile trip each way on Labor Day. Wasn't bad weather, mostly drizzle, and had "full self-driving unavailable" for almost the entire trip.
Did you have AP available even though EAP/Beta was unavailable? I’ve been in torrential rain storms where AP kept working long after Beta became unavailable.
 
Which is interesting, as I've never had any trouble with that (in Seattle!). I wonder if those who have trouble have a M3P with the rear spoiler, which might change the aerodynamics just enough?
I also have trouble keeping the rear camera clean during rain (and snow) on both cars (S & X), especially the legacy MX.
 
Last edited:
  • Like
Reactions: Ramphex
Yes. It is responsive when scrolled up to a higher speed limit but barely responsive when scrolled down for lower speed. It’s response is “broken” but apparently it’s not been a priority to fix it.
I wouldn't be surprised if this has evolved by design as a way of smoothing unintended braking events aka phantom braking and other situations, unless they can tie stronger deceleration specifically to scroll wheel usage

But in general I'd imagine that ramping up braking aggressiveness would ramp up complaints from unintended braking and who knows what else, there's probably a fine balance with those G forces
 
I landed home from Florida this morning. Had to drive a whole 0.3 miles to pickup lunch, however I have lots of experience with FSD and this left hand turn into my community which has a divided entrance. It has always been jerky making the left and some releases have tried to drive straight into the middle where the sign and tree are.

I have to say 10.69's first attempt is the best I've had yet on FSD beta, by far the smoothest and most confident.

 
Last edited:
Also possible they are being even more conservative with this early release. They have said they are being extra cautious.

Another drive to work, another failed AULT for me. This time it once again proceeded too far (though not sticking out quite as far as before).

Furthermore it also blocked the view of someone turning right next to me. Common practice is to pull up just far enough, so the person turning right can pull up a little further, to get visibility around me to their left, while not obscuring visibility to my right (until they go of course). It was also slow to get into position which exacerbated this (if it had been faster it would never have been an issue because the person turning right would not have arrived).

Anyway it was a total cluster and I just overrode it and went. On video of course. Reported!

So far 1 of 5 on my simple unprotected left. The one success had zero traffic.

Did about 5-6 reports today on a 16-mile drive. Lots of herky jerky on well-marked parkways, little hesitations and unnecessary braking (the reasons for it were usually clear, but the slowing was clearly not needed). Reported several instances of lack of assertiveness where it just does not get up to speed (for example entering freeway on ramp). Also got some good video of a hideous stop at a traffic light, with pulsing regen and braking. It started off smoothly then just fell apart. Plenty of time to use regen only and stop smoothly and gracefully.

10.69.1.1 also suffers from the same long-standing AP issue where it doesn’t quickly and smoothly get to the set speed, nor does it close gaps that appear in traffic.

I do think that overall 10.69.1.1 is much improved at steering. Had a couple very nice turns; so good that it mimics my steering and I don’t end up disengaging as I steer. It just goes around corners and I don’t have to take over! Will keep evaluating; still a small sample size, and I definitely expect inconsistency. (The issue with my ULT, aside from the terrible creeping, is that it doesn’t seem to know where to go, so the steering falls apart, which is a different issue.)
 
Last edited:
10.69.1.1 also suffers from the same long-standing AP issue where it doesn’t quickly and smoothly get to the set speed, nor does it close gaps that appear in traffic.
Except when under 10 mph. Then it's accelerate-brake-accelerate-brake to keep right on the bumper of the person in front of you in bumper-to-bumper slow traffic. No subtlety whatsoever. I feel that the people behind me are saying "what's wrong with this idiot" as they watch my head snap back and forth.
 
While this doesn't mean that Andrej helped with the occupancy network, I think the occupancy network's foundations were conceived for FSD V11, but it took them a lot longer to get it all polished enough for deployment.
Yes, there was plenty of early investigation of these last year even going back to Karpathy talking about 3D reconstruction and self-supervised depth in 2019 Autonomy Day, but it seems like practically, they did not really pursue it until Plenoxels, which was published December 2021. There are many ideas of what neural networks can do, but many of those potential approaches are limited by the ability to efficiently train the network. So with Plenoxels, Tesla now had a way to quickly get "ground truth" for a lot more of the data that was being sent back.

Starting with FSD Beta 10.6 (December 2021), there has been release note entries about "general(ized) static object network," and there's been ongoing improvements including adding control, so the conversion to a video occupancy network instead of images from single time steps was quite significant. Unclear if the 10.69 goal to "solve" unprotected left turns required this significant change, e.g., replacing visibility network, or if it just happened to be ready in solving various other safety disengagements.

Karpathy did talk about improving robustness to temporary occlusion at 2021 AI Day, but at least from 10.69 visualizations, that's still not implemented with "classical" cubioid moving object detection still has disappearing vehicles. I wonder if it's because that "legacy"(?) approach will be superseded by the occupancy network predicting volumetric flow, i.e., moving objects, as presented by Elluswamy at 2022 CVPR.
 
Except when under 10 mph. Then it's accelerate-brake-accelerate-brake to keep right on the bumper of the person in front of you in bumper-to-bumper slow traffic. No subtlety whatsoever. I feel that the people behind me are saying "what's wrong with this idiot" as they watch my head snap back and forth.
Correction, to keep about 20 feet back from car in front of you.
Why does it appear to be smoothly stopping a normal distance from the car in front of you and then suddenly decide to brake hard to stop way behind them? Drives me nuts and I'm sure the people behind me don't appreciate the sudden unnecessary increase in braking force.
 
Yes, there was plenty of early investigation of these last year even going back to Karpathy talking about 3D reconstruction and self-supervised depth in 2019 Autonomy Day, but it seems like practically, they did not really pursue it until Plenoxels, which was published December 2021. There are many ideas of what neural networks can do, but many of those potential approaches are limited by the ability to efficiently train the network. So with Plenoxels, Tesla now had a way to quickly get "ground truth" for a lot more of the data that was being sent back.

Another possible explanation was that Tesla was limited by their training cluster for consuming vast amounts of video, so they weren't able to fully realize the occupancy network until their cluster was powerful enough. The Nvidia A100 GPU was released in 06/2020 (80gb variation in 11/2020).

Just a bunch of speculation, but I'm sure the AI team was constantly looking for ways to solve general objects in "4D".
 
Something that I haven't seen anyone else mention is how long it takes to slow to the posted speed after a speed limit change. I often drive on 2-lane highways that pass through small towns. The speed limit can change from 55 to 45 to 35 and even 25 in short order. FSD beta (10.12.2 still for me) takes a very long time to slow down. Sometimes I can be through the town completely before it reaches the lower speed limit. That is a very good way to get a ticket!

As a result, I typically have to disengage every time I enter one of these little towns. Once I slow the car down, then I can reengage.
Have you tried taping the speed stalk down multiple times in advance. I have been noticing that the car hits the brakes or regen and slows the car down alot faster. Im is the same boat ith a small town and Barney Fife is right there waitinng. I got pulled over once, he didnt see a notible decrease in speed but didnt site me.