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

Wiki MASTER THREAD: Actual FSD Beta downloads and experiences

This site may earn commission on affiliate links.
While overall 10.11.2 has been a mild improvement I have several LARGE steps back that makes it impossible to not have a disengagement drive. Here is an example that started with 10.11.2 and is an easy right turn. today it even had a lead truck to follow. Every time it turns on right signal and then veers off to the left. This is a dangerous place to do that with 4 lanes in one direction. I have been honked at twice. I may try and go early in the morning and see if I can catch it with no traffic and let it play out.

I also noticed the beta starting to veer left more often before some right turns that don't really need it. It's like there is some threshold of visibility or how sharp a turn is where it starts doing that
 
I also noticed the beta starting to veer left more often before some right turns that don't really need it. It's like there is some threshold of visibility or how sharp a turn is where it starts doing that
Agree, happened to me twice today but each time there wasn't a car to my left. Both times I thought for a second it might not even make the right hand turn but it did, just took the scenic route.
 
I also noticed the beta starting to veer left more often before some right turns that don't really need it. It's like there is some threshold of visibility or how sharp a turn is where it starts doing that
Agree, happened to me twice today but each time there wasn't a car to my left. Both times I thought for a second it might not even make the right hand turn but it did, just took the scenic route.
Right.

This is the basic problem with cost based optimization. When we drive, we place a high "cost" for crossing the lane. Because we have to be sure no car is coming from behind etc. FSD always knows whether a car is coming from behind. Depending on the costs they have used for lateral acceleration etc., cost optimizer might think crossing the lane would actually be of lesser cost than that sharp turn.

I think in a lot of these case, the correct thing would be to reduce the speed at which FSD takes turns. If the speed is low, you don't need to cross any lines and still be able to take sharp turns with low lateral acceleration.
 
Right.

This is the basic problem with cost based optimization. When we drive, we place a high "cost" for crossing the lane. Because we have to be sure no car is coming from behind etc. FSD always knows whether a car is coming from behind. Depending on the costs they have used for lateral acceleration etc., cost optimizer might think crossing the lane would actually be of lesser cost than that sharp turn.

I think in a lot of these case, the correct thing would be to reduce the speed at which FSD takes turns. If the speed is low, you don't need to cross any lines and still be able to take sharp turns with low lateral acceleration.

Not sure turning speed (lateral g-forces) is the reason here. I've seen the car do this at very slow speeds. It's even come to a full stop with the car angled to the left, before making the right turn, even when no one is around. To me, this is the car preferring to see to the left with the front cameras.

Common problem around here is people who think I'm not planning to turn right because I left a ton of space to the right-side curb, and they try to squeeze by. They wouldn't try that if my car made the turn at a more normal speed (15mph).

And I know from my safety score driving days that I trigger the lateral g-force ding around 16-17mph on level ground for a sharp 90-degree right turn. I have not seen FSD make any sharp right turns at that kind of speed.
 
Not sure turning speed (lateral g-forces) is the reason here. I've seen the car do this at very slow speeds. It's even come to a full stop with the car angled to the left, before making the right turn, even when no one is around. To me, this is the car preferring to see to the left with the front cameras.
We're going to need the raw video to determine if the truck went over the curb. haha.
I think the root cause is that the perception is not all that accurate. That leads to a cost optimization problem where one of the costs is your rims.
Yes - definitely not curbing wheels is high the cost side. As it should be compared to crossing the line.

I'm not sure the cost optimizer is using a strategy of moving to the left to have a better look - I'd be surprised if it can pull something like this off.
 
We're going to need the raw video to determine if the truck went over the curb. haha.
I think the root cause is that the perception is not all that accurate. That leads to a cost optimization problem where one of the costs is your rims.
He is going over a manhole cover that is uneven. The curb is very low and rolls to street level. If you look close you can tell the tire is not up on the curb.

Also in 3ed pic here it is making a (different) turn and seemingly not caring too much about my wheels.:oops:

Screen Shot 2022-04-19 at 2.23.52 PM.png

Screen Shot 2022-04-20 at 12.11.33 PM.png


IMG_0756 (1).jpeg
 
Last edited:
  • Informative
Reactions: EVNow and impastu
He is going over a manhole cover that is uneven. The curb is very low and rolls to street level. If you look close you can tell the tire is not up on the curb.

Also in 3ed pic here it is making a (different) turn and seemingly not caring too much about my wheels.:oops:

That curvature of the right curb means, the car should be able to turn without issues.

Anway - on a slightly different note - yesterday I saw a UPS truck repeatedly cross the center line when driving straight ;)
 
Another run on my camping trip run. This time it really highlighted something I'd not seen in previous runs.
It has always had an insistence with using the left lane it if it there, especially when its the one lane into two lane transitions.
Previously AP would start for the left lane then lurch to the right in an uncomfortable moose test maneuver. FSD beta has been handling that quite well.
I should have been warned by some roads getting the constant attempts to move to the left lane to "continue on route", but the new trick on the odd times that I just give up and let it move into the left lane, it has now started to signal left and pull into left turn only lanes on divided highways, only to perform that annoying moose test to get back into the main lanes.
Yay - now I've got more roads where I turn off FSD beta :(
But at least I can change driver profiles while moving :D
 
Or it has to make a quick lane change. Going purely by DOT code, that is probably what is correct.
Doubtful since there is a wide line to the left of the right turn lane — I don’t think you’re supposed to cross those the “wrong” way.

Also, it still moves over and hugs the right side of the lane where the lane is double-wide for an unmarked right turn. So the car steers right then back to center after the intersection.

In another hugging case, rather than stay centered in the lane, it sideswipes overgrown vegetation.
 
Last edited:
So the UPS truck was autonomous and the center line was squiggly, then? It seems more likely to me that the UPS truck was being driven, and not in a straight trajectory...
No - just noting how human drivers drive.

There is actually low cost to crossing lines - when no traffic is present. Human drivers frequently cross lines.

ps : My wife's Toyota van has lane departure warning. She turned it off because it would beep so often ;)
 
No - just noting how human drivers drive.

There is actually low cost to crossing lines - when no traffic is present. Human drivers frequently cross lines.

ps : My wife's Toyota van has lane departure warning. She turned it off because it would beep so often ;)
I do agree with what you say here, however, a human is not driving a machine is and I don't have any information for Tesla or anyone who knows if this behavior (crossing the center line on bends in the road) will not occur with oncoming traffic. If Tesla would publish a formal explanation of how and when the car will behave we could all feel more confident in using the FSD in certain situations. My own perspective is that if I see the car consistently crossing the centerline on bends in the road and I am unable to ask the machine why it is doing this I make a assumption (to be safe) that it will do it always. If the machine (or its creator) tells me that it only will behave this way when there is no oncoming traffic I will accept that explanation and stop being concerned about the behavior.