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.
@EVNow Not here. I successfully made it through a roundabout in Covington, WA several times (v10.2) - day and night. While the transition was anything but smooth, it didn't stop first, before entering. It just yielded.
Great to here. But like EVNow and others have posted FSD stops before entering the rotary then plays a game of slow creeping and stopping before proceeding. And then in my experience going around the rotary is an adventure sometimes requiring a disengagement and/or intervention.
 
  • Like
Reactions: impastu
It seems to handle this one just fine, and it isn't even mapped as it is new construction:

After seeing it again, I wonder whether it even recognized it as a round-about. I want to see that with vehicles around.

I won't be surprised if it just handled like it would handle twisty roads. Thats why no speed reduction or any kind of hesitation.

The ones I've near me are small neighborhood round-abouts that have steep curvature and need considerable speed reduction to navigate.
 
  • Like
Reactions: impastu
Great to here. But like EVNow and others have posted FSD stops before entering the rotary then plays a game of slow creeping and stopping before proceeding. And then in my experience going around the rotary is an adventure sometimes requiring a disengagement and/or intervention.
I feel bad for all of our members that have rotary's... I have 1 in 50 miles in any direction. Please let me know when Fsd is perfect so I can do the 40 mile round trip to test it. I will post the results at that time, which by my estimation will be in 2030.
 
  • Like
Reactions: aronth5 and impastu
Well, I decided to try out 10.3.1 for the first time, just now, and reaffirmed my plan to halt testing until at least v11, whenever that may be. I took a 10 mile round trip drive, on lightly traveled rural (and sometimes hilly) route. Conditions: Roads was damp, but it was not raining. Sun at times.

Issues encountered on this very short, and mundane, route include:

1. Crossed over double yellow centerline three times, once severely. The severe cross-over was on a hilly road, while on the rise. Thankfully, I disengaged quickly - as when I crested the hill, there was an oncoming school bus... F me!
2. Tons of phantom braking and wild yoke jerking (abnormal wear and tear continues)
3. Drove more on the extreme left side (not center or right) on private roads than 10.2; and 10.2 was bad.
4. Tried to run a T intersection (same as 10.2, so I was ready for it) - Private road to main public road.
5. "Chill" FSD mode seems to only set the "Chill" acceleration toggle. Nothing else felt "Chill" at all.

Sending email to Tesla.
 
Great to here. But like EVNow and others have posted FSD stops before entering the rotary then plays a game of slow creeping and stopping before proceeding. And then in my experience going around the rotary is an adventure sometimes requiring a disengagement and/or intervention.
Fellow Boston area FSD tester here (Somerville/Cambridge to be specific) and agree with your statement about rotaries. I've had it drive through some rotaries relatively well (a little jerky steering inputs but made it through without causing any drama) and I've had it fail while in the rotary or not enter the rotary at all.

The one rotary where it fails is the one in Powderhouse Circle (next to Tufts Univ) b/c with as pylons/white plastic poles lining the entire rotary to force ppl to use only one lane (and to make room for bicyclists). This rotary freaks out FSD big time haha. Here's a pic of the lanes that are enforced by the pylons throughout the rotary:

Screen Shot 2021-10-27 at 12.52.04 PM.png
 
Every roundabout I've seen the car stops at the entrance, takes time to figure out what to do and then slowly moves. The stops can be a few seconds long and when the round-about is empty and cars behind us - will feel forever.
Just one second is an eternity to a computer running at GHz clock rates (and Teslas have two of them). You'd think that after a few billion instructions the computers would have settled on a solution. So, what were they thinking?
 
What you are saying makes complete sense why it signals one way and then another and struggles.

As users of beta how do we help FSD to learn in roundabouts ?

I pretty much disengaged on all of them today
Actually, I'd almost swear that I was taught you should signal right when you enter a roundabout, left when you're not exiting, and right when you are exiting. This is pretty much impossible on tiny roundabouts, and I fully disagree with the suggestion to signal right upon entry, but if that is codified, it could be exactly what FSD is doing in spite of the fact that no other driver would.

Disengaging when you need to is probably all you can do to help training. Don't be afraid to use the accelerator to enter when there is no traffic to yield to anyway, though, that might help, too.
I’ve had a few of those too. Also, does anyone else notice the performance of FSD tends to really suck in the rain? It’s been pouring here all day, and a few drives it did some really dumb $hit.
My first two days of my first beta version (10.3/10.3.1) were pretty much all rain, and I was impressed how well it did. In fact, there seem to be more things it did worse with the sun out than there are things it did worse in the rain. TBH, I never really felt like it was trying to kill me until the sun came out. Strange how different one experience can be from the next.
 
Fellow Boston area FSD tester here (Somerville/Cambridge to be specific) and agree with your statement about rotaries. I've had it drive through some rotaries relatively well (a little jerky steering inputs but made it through without causing any drama) and I've had it fail while in the rotary or not enter the rotary at all.

The one rotary where it fails is the one in Powderhouse Circle (next to Tufts Univ) b/c with as pylons/white plastic poles lining the entire rotary to force ppl to use only one lane (and to make room for bicyclists). This rotary freaks out FSD big time haha. Here's a pic of the lanes that are enforced by the pylons throughout the rotary:

View attachment 726358

never been to that one. looks nasty :D

Ever been to the double rotary on Fresh Pond (near start of Rt 2)? Total zoo.

 
Well, I decided to try out 10.3.1 for the first time, just now, and reaffirmed my plan to halt testing until at least v11, whenever that may be. I took a 10 mile round trip drive, on lightly traveled rural (and sometimes hilly) route. Conditions: Roads was damp, but it was not raining. Sun at times.

Issues encountered on this very short, and mundane, route include:

1. Crossed over double yellow centerline three times, once severely. The severe cross-over was on a hilly road, while on the rise. Thankfully, I disengaged quickly - as when I crested the hill, there was an oncoming school bus... F me!
2. Tons of phantom braking and wild yoke jerking (abnormal wear and tear continues)
3. Drove more on the extreme left side (not center or right) on private roads than 10.2; and 10.2 was bad.
4. Tried to run a T intersection (same as 10.2, so I was ready for it) - Private road to main public road.
5. "Chill" FSD mode seems to only set the "Chill" acceleration toggle. Nothing else felt "Chill" at all.

Sending email to Tesla.
It might be interesting to meet up with your S and my Y and go back to back on the same route and see if both cars perform the same.
 
Sorry if this has already been mentioned somewhere in the thread. I've been trying to follow this, but it's been a little challenging to keep up with all the posts.

I'm wondering if anyone knows anything about how a particular path is chosen when you navigate to a destination, because I've had some inconsistencies and curiosities. There is one route I travel almost every morning and afternoon and I nearly always have to disengage at one point, because it wants to take me on a route that requires a U-turn and it NEVER does the U-turn properly. The path it wants to take isn't shorter or faster, but it's not longer either. To my surprise today, it chose the path I always manually drive and I was just wondering why it chose a different path today. Does it actually learn or was this just a fluke? This was probably the 5th time I've traveled this path on 10.3.1 but the first time it's chosen the "correct" route. There are some other areas where the route it's chosen is just bizarre, but not entirely wrong. These usually are related to shopping areas, though, and I guess it really doesn't do well with parking lots. It's just picked some random side streets and unnamed roads rather than taking a standard, direct path.

I really wish Tesla would allow you to drag the navigation route to "correct" things like this, sort of the way google maps does it.

One other thing I've been wondering about is speed bumps/humps, which seem to be completely ignored. There are a number of different types of speed bumps, speed humps, and speed tables around here and 25mph is way too fast for most of them. They're mostly clearly marked (usually with a sign as well), so it seems like this is something that could be accounted for. There's one area in particular where it's a narrow, unmarked road that curves right near a speed hump. Since the car wants to go in the middle and NOT slow down for the speed hump, if I'm behind another car, it pretty much starts passing the other car and I have to intervene.

Final note, the other day I was navigating home but it missed the turn onto my street and I looked down and noticed there was no destination set! So, I said "navigate home" again and I was just about to pass the next street after mine when the destination loaded up and the car suddenly took a FAST left turn onto that street. It didn't slow down for the turn AT ALL. So, I'd say be very careful with setting new destinations while en route. Maybe this was just a random one off, but I'm not going to assume that at this point.
 
  • Like
Reactions: impastu
There are a ton of roundabouts here, and I see both what @EVNow and @jebinc are reporting, depending on which one I'm going through.

Performance varies from completely unacceptable to surprisingly good.

What is really entertaining to watch is its usage of the blinker ("turn signal" out here) entering a roundabout. It signals back and forth like it just can't decide. What I thought was really odd, is how it always turns on the turn signal prior to entering a roundabout. I thought (but again, could be wrong) that proper blinker usage is that you don't turn on the turn signal prior to entering a roundabout, but you did use it as you were exiting a roundabout, if it is appropriate for the layout of the roundabout you're taking.

My Model 3 likes to use it non stop, from prior to entering the roundabout, during the roundabout, and getting ready to exit the roundabout, with no apparent logic involved as to which direction it's going to signal, often switching between right/left numerous times without the car ever even changing lanes.

There are a few roundabouts where it doesn't go nuts with the blinker, but it does use it before entering every single one of them.
 
Just one second is an eternity to a computer running at GHz clock rates (and Teslas have two of them). You'd think that after a few billion instructions the computers would have settled on a solution. So, what were they thinking?
I would suggest tot he community that part of the reason why FSD beta is so jerky and indecisive is due to this, data processing speed. Maybe by HW4, 5, or 6, this will be less prevalent?
 
  • Like
Reactions: Phlier and impastu
I would suggest tot he community that part of the reason why FSD beta is so jerky and indecisive is due to this, data processing speed. Maybe by HW4, 5, or 6, this will be less prevalent?
Although that's possible, then wouldn't the "jerky and indecisive" behavior exhibit some consistent pattern? It would either occur everywhere or it would happen more often under circumstances that required complex decision making.

My experience is that such behavior occurs quite randomly. I've seen my car traverse an unprotected left turn in heavy traffic and the flub an easy right turn. It indicates to me that the NN needs a lot more training and also that there are probably regional deficiencies in the quality of mapped data.
 
I would suggest tot he community that part of the reason why FSD beta is so jerky and indecisive is due to this, data processing speed. Maybe by HW4, 5, or 6, this will be less prevalent?
No - its not the computer or NN. Its the planning code.

Round-abouts (esp, single lane) are actually very easy. Easier than 4 way stops. You have to just look to your left and make sure there is noone coming. Its just a simple unprotected right turn.