FSD Beta 10.69

[powertoold]

Interesting breakdown thread of a slide from AI Day by Whole Mars Blog:

https://twitter.com/x/status/1576724117314641927

Basically, they're using crowd-sourced data from multiple trips through an intersection. But they're not using it to create an HD map that would become stale relatively quickly; they're using it to generate ground-truth for training data labeling.

This approach has me questioning:

If they simply overlay the prior trip data to the current trip, who is there to make sure that the current trip is only labeled with what the car can see? I know they have a human label evaluator, but do they even care that occluded lanes / geometries are being labeled in the current trip? Wouldn't labeling things that the car can't see cause problems? I'm sure they've considered this, but it wasn't explained during AI Day.

No one here knows this, I'm just spitballing.

 

[AlanSubie4Life]

don't think any driving behavior is determined directly by a neural network at the moment, is it?

But assuming perception quality is relatively equal across all circumstances, what would explain day-to-day variation in FSDb performance?

Seems like the planner doesn’t have to be NN-based to have quite poor behavior with small variations in the inputs. Seems pretty complicated to get right. They could also just be optimizing for the “wrong” things in the planner (for example they may be putting a high priority on not running into things).

Definitely could be variation in perception contributing though. Remember we have no way to know what the car actually sees. There is literally nothing provided that gives us any idea.

 

[DrChaos]

Example of FSD deciding it should avoid the crossing car by turning towards the telephone pole. One wonders what would have happened if I did not disgengage? As FSDb testers know the feeling in the car is much jerkier then it appears on video. There was plenty of time to slow down.

It's possible that the car crossing had little to do with it, and it was the expansion of the seemingly 'drivable area' to the right (grey pavement without lane lines) which erroneously encouraged the car to center itself in that direction to the right.

 

[ironwill96]

Most of my interventions over the weekend using FSDb were tied to the car just being weird and unpredictable for cars following close behind me.

Stop signs are a big one - FSD starts slowing at an appropriate distance from a sign but it does so with a LOT of brakes and regen, so you drop from day 50mph to 15mph 300 feet from the stop sign and then it creeps forward really slow to the stop sign ticking off everyone behind you. They really need to smooth out the deceleration curve in these scenarios.

The other problem is the late slam on brakes and throw on signal right before an aggressive right turn that slings all my passengers around and risks a collision from the drivers behind me. Ive sent snapshots of both of these.

 

[sleepydoc]

Nah, one is just a denier that doesn't like it when someone contradicts him with evidence. I get it, it's hard being wrong!

I'll have to make some direct measurements again tomorrow when I get off work comparing the deceleration of FSD with TACC and AP. I'm not sure that it will matter but I'll post them anyway.

Ok - I did a quick test on my way home from work. Slowing down from 55 to 50 on FSD took between 11 and 12 seconds while slowing from 55 to 50 on TACC took between 7 and 8. FSD was also going up a slight incline while TACC was on a flat section of road.

I’ll try to test more later. There was too much traffic this morning to get any more clean results. I also noticed that the last 1 MPH took significantly longer in both tests so the difference is actually a bit more marked if you only look at the 55 -51 mph times

 

[aronth5]

It's possible that the car crossing had little to do with it, and it was the expansion of the seemingly 'drivable area' to the right (grey pavement without lane lines) which erroneously encouraged the car to center itself in that direction to the right.

Possible but I take this road all the time and it's never happened before. It was an abrupt change in direction which is not portrayed well in the video. Certainly felt like it was reacting to the crossing car.

 

[bobsmy]

This is another reason why FSD needs speed of traffic awareness.

In some places I drive, the very casual and late-triggered slow downs is exactly what the rest of traffic is doing, so it’s a good thing there.

In the kinds of places you’re describing, yeah, totally the wrong thing to do and will result in getting a ticket.

A lot of my interventions would go away if I could configure FSD to go up to 10 mph over the limit if that’s what the rest of traffic is doing (eg I’ve reached the speed limit, but the cars in front are going faster than me leaving a widening gap and the cars behind me are leaving no gap).

This will be an issue with all AI driving software as it is by nature designed to obey traffic laws while non computer assisted driving vehicles will continue to violate ( or stretch) traffic laws. I have experienced the reaction of angry drivers even when I am driving 8-10 miles over the posted limit but is not fast enough for them.

 

[bobsmy]

It's possible that the car crossing had little to do with it, and it was the expansion of the seemingly 'drivable area' to the right (grey pavement without lane lines) which erroneously encouraged the car to center itself in that direction to the right.

The issue of the car centering itself on the expansion of the seemingly drivable area has been around for over a year. It is one of my major complaints as if it occurs where the added pavement is for a turn lane it signals drivers possibly entering from the right and those behind you that you intend to turn right, then suddenly the car moves back to the left and continues straight. I am certainly not an AI expert, but it seems like a delay in lane centering would go a long ways to eliminating this.

 

[willow_hiller]

Maybe a first sign of 10.69.3 going out to employees as 2022.20.18. All the previous employee software updates first hit a Model Y SR:

https://twitter.com/x/status/1576728836653432834

EDIT: Others are claiming 10.69.2.3:

https://twitter.com/x/status/1576747610617966592

 

[Knightshade]

Much as it was hyped I'd kinda hope 69.3 is on a new base firmware anyway

 

[willow_hiller]

Maybe even customer installs, in a limited capacity, already. I don't think this guy is an employee, and as far as I can tell, his tweet is the source of this image:

https://twitter.com/x/status/1576736546157314048

 

[petersk]

Is anyone that has FSD beta incredibly disappointed with it? It gives you a violation and stops working with literally no warning. No sound nor visual cue. It seems like you have to exert way too much force on the steering wheels when prompted to, often taking it out of autopilot.

It spends way too much time thinking about turns, especially right on red.

Overall, I turned it off, because it’s actually worse than EAP.

 

[BBTX]

In fact, the last I heard, is that they have to use both processors to get FSD to run, so there is no longer any true “shadow mode”.

Yeah, but, the feeling I get from some of the AI presentation is that they are shipping isolated shadow NNs for testing. For instance, 10.69.2.2 could have 3 versions of some specific object classifier NN: the real one that's affecting car behavior, and two others that are in early testing and not affecting the car's decisions, and then probably looking for cases where the new code is regressive, or the flip side where the new network would likely have avoided an intervention. They could even be doing live A/B testing on us at this kind of level.

 

[flutas]

I agree with a lot of posts in here, this build seems a lot better than previous ones for me, but the same level of inconsistency, or close.

Still issues for them to work through though. Biggest ones for me revolve around dual turn lanes.

1. The car will always go to the left turn lane, even if a right turn is shortly after.
2. If I force the car into the right-most left turn lane, it tries to swing wide and always feels like it's going to clip the traffic going straight.

I was also stuck on 10.69.2 for a while (with broken voice controls). Going into service mode and triggering a reinstall prompted for 10.69.2.2 right after the reinstall was done.

 

[Jmolloy3]

Is anyone that has FSD beta incredibly disappointed with it? It gives you a violation and stops working with literally no warning. No sound nor visual cue. It seems like you have to exert way too much force on the steering wheels when prompted to, often taking it out of autopilot.

It spends way too much time thinking about turns, especially right on red.

Overall, I turned it off, because it’s actually worse than EAP.

Maybe moderately disappointed. I've always tempered my expectations, because I knew this would be immensely difficult for anyone, including Tesla. For me, it's annoying that 69.2 seems to be worse. The herky-jerky drive is is like an overcaffeinated Rain Man with a nervous tic on his right foot.

 

[MP3Mike]

I thought this too when I watched a video about it but........it is about the SAME as driving an unmarked road except MUCH safer. On unmarked roads you have VRU's on both sides roads, often occluded and must navigate. On a road with advisory bike lanes you aren't required to drive in the middle (stripped and NOT solid) and can move to the side but MUST yield to VRU's. The lines are just there as a visual cue that guides you away from the MUCH more dangers sides of the road which are highly likely to have VRUs. So while at first it looks like a stupid or counterintuitive idea, in actuality all unmarked roads would be much safer if they were painted this way. It is totally impractical to just widen all residual type roads to make them safe for VRUs and have 2 lanes of detected car travel.

Sure, but they are only doing this to three short disjointed sections of the road so in the matter of ~30 blocks you will go into and out of this format three times. And there are probably less than 5 bikes on this road per day. (There is a major road just a few blocks away that has continuous dedicated bike lanes.)

Maybe with these changes more bikers will move to this road, but I doubt it. (There aren't a lot of bikes that use the main road with dedicated lanes to start with.)

 

[EVNow]

I've not written much about 69.2.2 .... still collecting data.

Has a lot of "jittery" issues and lacks refinement. We can now compare to a new driver taking driving lessons. Needs the instructor to correct from time to time.

But I'm now doing many more zero disengagement drives than any other release. May be even more than all previous releases combined, even with the roundabout issues I have.

 

[Jersey_guy_202]

Is anyone that has FSD beta incredibly disappointed with it? It gives you a violation and stops working with literally no warning. No sound nor visual cue. It seems like you have to exert way too much force on the steering wheels when prompted to, often taking it out of autopilot.

It spends way too much time thinking about turns, especially right on red.

Overall, I turned it off, because it’s actually worse than EAP.

There's blue visual warning prior to disengagement, but I agree the time in between is too short. What I usually do now is just turning sound up and down, or slight force turning bottom right wheel with one finger (just slightly, not to the point of disengagement) and that usually does the trick.

If you feel turning is too slow, you can press on accelerator to speed through the turns and that will still keep FSD intact.

I had similar feeling as your the first few days, then I decided to trust the car more to let it do its thing. While FSD is not perfect, I got to understand the current limitation and now for the most parts I know where I need to be extra cautious or just take over, and can let it drive majority of time. I would say give it another chance.

 

[sleepydoc]

I agree with a lot of posts in here, this build seems a lot better than previous ones for me, but the same level of inconsistency, or close.

Still issues for them to work through though. Biggest ones for me revolve around dual turn lanes.

1. The car will always go to the left turn lane, even if a right turn is shortly after.
2. If I force the car into the right-most left turn lane, it tries to swing wide and always feels like it's going to clip the traffic going straight.

I was also stuck on 10.69.2 for a while (with broken voice controls). Going into service mode and triggering a reinstall prompted for 10.69.2.2 right after the reinstall was done.

I’ve actually had it pick the wider turn lane on its own a couple of times in 69.2.2 but generally my observation has been the same as yours. My assumption has been that it’s chooses the leftmost/tighter turn lane because easier to plot a course.

I've not written much about 69.2.2 .... still collecting data.

Has a lot of "jittery" issues and lacks refinement. We can now compare to a new driver taking driving lessons. Needs the instructor to correct from time to time.

But I'm now doing many more zero disengagement drives than any other release. May be even more than all previous releases combined, even with the roundabout issues I have.

With the exception of slowing too soon at stop signs, I’ve found 69.2.2 to be smoother overall, but I’ve definitely noticed a far higher rate of zero-disengagement drives.

The two major issues are still lane selection and unprotected turns. It will usually handle the turns but it still takes far too long to decide and commit. I often end up hitting the accelerator not because it’s doing something dangerous but out of consideration for other motorists.

 

[sleepydoc]

The issue of the car centering itself on the expansion of the seemingly drivable area has been around for over a year. It is one of my major complaints as if it occurs where the added pavement is for a turn lane it signals drivers possibly entering from the right and those behind you that you intend to turn right, then suddenly the car moves back to the left and continues straight. I am certainly not an AI expert, but it seems like a delay in lane centering would go a long ways to eliminating this.

I think the problem is FSD doesn’t look far enough ahead. When a human drives and the lane starts to widen you look 50 feet down the road and see that it’s actually becoming a turn lane so you pick the lane you want and head towards it. FSD seems to focus on the next 10-20 feet and then suddenly realizes there’s 2 lanes that it needs to decide between.