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.
Couple more good things I noticed - not sure if anyone reported this already but if FSD becomes unavailable due to rain, the lane keep function that remains can go on green without a lead car. Speed bump visualizations are back too. I've wondered if some of the random slowdowns on residential streets were related to false speed bump detections, now it'll be easier to see if that's the case.

Overall not impressed though. I replicated the bad cone detections I saw on the freeway the other night - I was passing some pylons that are part of the street; FSD panicked and made a sudden attempt to merge left to get away from "cones". Also stopped a couple car lengths short at an unprotected left; got stuck there even after the light turned green. Might've been because it saw some pedestrians on the corner on my side of the intersection. They didn't move even after they had the right of way for a few seconds, and neither did the beta. I disengaged, pulled out, waited for the cars going straight, made the turn, and saw them run to the median after the light turned yellow. Maybe random variation, but I've taken that turn multiple times with previous versions without issue.
 
  • Like
Reactions: impastu
turns the left signal on and start changing lanes in the middle of the intersection
View attachment 748577
Did the lanes change recently? The image you show has a straight arrow and a straight+right arrow for the two westbound 14th lanes crossing Peachtree whereas this November 2021 Street View seems to show two straight arrows?

street view 14th.jpg


I wonder if the neural network some reason believed the right lane was a right-turn only lane given the angled and shifted intersection, so it last second decided it needed to be in the left lane? I believe FSD Beta starting with 10.x started predicting lane turn connectivity to override map data for immediate intersection behavior (whereas map data can still be used for longer term "follow route" planning). I wouldn't be surprised if the neural network hasn't had much training on this "Nevada-shaped" intersection.
 
The engineering reality doesn’t match with the levels. Levels have become obsolete.

All the ADAS systems from lowly lane departure warning in our Toyota van to Tesla FSD beta are all L2 ? But the keep lane TACC of MB below 37 mph in 60 mph roads in L3 !!
The levels are definitely all about ownership of the DDT aka liability. If there were no levels, 100% of this tech would be considered just advanced driver assist requiring constant supervision until one day that final incremental improvement creates fully autonomous vehicles. When that could happen is anyone's guess, but I think full autonomy in the US is many years away and full autonomy around the globe is probably much further. And I think Tesla, or at least Elon, is still way too optimistic about when it'll happen.

But the Levels and particularly 3-4 provide the opportunity to register higher degrees of autonomy through ownership of the DDT within defined ODDs and narrower scopes of function before we reach full autonomy. I think this is a smart approach in a lot of ways if companies can create systems that are pretty broadly capable but that can do some specific things with extremely low error rates, low enough to accept the risk of taking responsibility for the driving task in any capacity and to add value to people by meaningfully reducing the driving burden. And improving safety by making up for our shortcomings where possible.

Working to accept liability for the driving task in narrower scopes likely doesn't make a lot of sense if you think full autonomy is right around the corner, and there's no shortage of very smart people who have spent much of the last decade telling everyone that it's right around the corner. When it doesn't materialize year after year, they just say to look again because surely it'll be there this time.


Interesting that Hotz also sells a device he says will give all the functionality of Autopilot for $1k and that he'll be 1-2 years behind Tesla whenever they reach full autonomy. The other car manufacturers and companies like Waymo etc are playing a much longer game because they clearly don't think full autonomy is happening any time soon and that there will be many, many small steps along the way.
 
Last edited by a moderator:
Just completed my first 10.8 drive. Overall very solid - but I was focused on one specific area that is unfortunately not improved in 10.8.

I'm curious for other people's experience with this. There are two three way intersections in my neighborhood (each with only a single stop sign). When my car approaches these stop signs to make an unprotected left turn, the behavior is pretty awful:

  1. The car stops waaaay before the intersection. I understand this is because there is no line at the intersection, so it has to guess where to stop. But it's not even close.
  2. It then begins a very slow creep towards the intersection. Meanwhile, I can clearly see that there is no cross traffic approaching, and the car's b pillar cams have an unobstructed view.
  3. More hesitation, more steering wheel jiggle, more creeping, often into the intersection. Plenty of visibility, and clear to me that no cross traffic is approaching.
  4. Finally when it's completely in the middle of the intersection, it makes up its mind and turns left.
This behavior is not consistent for all intersections. I've had better luck at other similar intersections, although it is still much more hesitant than a human driver.

Anyone else have a similar experience? It's actually one of the major remaining issues that makes FSD Beta unusable for my commute/drive.

EDIT - here's one of the intersections. Sorry for the terrible scribbling - the intersection is circled, and the arrow shows the direction I'm approaching to the stop sign.

I think one contributing factor is that the cross street bends out of view, so I suppose it's possible a car could unexpectedly come through at an extremely high rate of speed. As you can see that cross street in a cul-de-sac and there is hardly any traffic, so I just know that I'm not going to see 50 mph cross traffic. But I suppose FSD doesn't take this into consideration?

View attachment 747863

I've been thinking more about this behavior. Specifically - it seems that FSD will hesitate for a minimum of 5-10 seconds before making an unprotected turn, no matter how good the visibility is or the complete absence of any cross traffic.

I wonder if it's one of two things:
1. There is actually that degree of latency in the system - i.e. it takes 5+ seconds to determine that there is absolutely no cross traffic. I'm not even talking about situations where there is some cross traffic, and FSD is trying to determine distance and velocity. I'm talking about plenty of visibility and NO traffic in sight.

2. There is some hard-coded minimum creep/pause for any unprotected turn, if for no other reason than to give the driver time to abort if necessary.

The first option seems highly unlikely - if it really takes the FSD stack 5+ seconds to determine the presence, distance and velocity of other vehicles, how does it perform any other driving tasks? Unless this latency is particularly bad for the B pillar cameras.
 
  • Like
Reactions: FSDtester#1
I've been thinking more about this behavior. Specifically - it seems that FSD will hesitate for a minimum of 5-10 seconds before making an unprotected turn, no matter how good the visibility is or the complete absence of any cross traffic.
The behavior you're referring to certainly isn't my experience. More than once I've had to disengage with a hard brake because FSD is making a quick decision with a car a second or two away. In my circumstance it's an unprotected left at T-intersections without a light or Stop sign. A light or Stop sign exhibits the behavior your're talking about. My area is full of these type of intersections.
 
The behavior you're referring to certainly isn't my experience. More than once I've had to disengage with a hard brake because FSD is making a quick decision with a car a second or two away. In my circumstance it's an unprotected left at T-intersections without a light or Stop sign. A light or Stop sign exhibits the behavior your're talking about. My area is full of these type of intersections.

Yes, you're correct - the two intersections in my neighborhood that I'm referring to are both governed by a stop sign (not a four way stop though). The question remains - with great visibility and no sign of any cross traffic, a human driver only hesitates at a stop sign for a second or two before making the turn (either left or right). I'm actually hoping it's a hard coded delay that they can back out when they feel more confident.
 
The behavior you're referring to certainly isn't my experience. More than once I've had to disengage with a hard brake because FSD is making a quick decision with a car a second or two away. In my circumstance it's an unprotected left at T-intersections without a light or Stop sign. A light or Stop sign exhibits the behavior your're talking about. My area is full of these type of intersections.
I had that issue this morning too. FSD arrived at a stop sign with no other cars on the road and it just waited and waited, then just as another car approached the junction FSD decided to go for it - except it wasn't a 4way stop. My hard braking intervention prevented FSD from getting us t-boned.
10.8 seems like some improved things, others not so much.
 
What happened - phantom braking ?

BTW, that is not really 10.8 related.
I disagree that it's not 10.8 related. I know it's not single stack, but I have zero doubt that the freeway AP software is different with Beta than it is with non-beta Teslas. We have a non-beta 3 and a beta Y, and I've been able to compare the two. Absolutely the freeway AP is different with beta. Our 3 AP is pretty much unchanged over the last four months. Our Y used to be much better than the 3 on the freeway, with the improvement coinciding with the upgrade to beta. The two APs are very different since beta, whereas they were identical before.

As far as what happened, first off both TACC and AP are now unusable on the freeway if it's very lightly raining, I mean less than a sprinkle. I get a message that max speed is limited due to restricted visibility and the car slows down to 35 or 40--and that's with no one in front of me. Before, AP would work quite well even in a downpour, now it's downright dangerous in light rain. I put on the turn signal to change lanes, and even though the lanes were correctly shown on the display the car started swerving like a drunken sailor. I honestly feared we'd spin out of control. I intervened, tried it again, still a drunken sailor. I now won't engage AP or TACC on the freeway if it's even sprinkling. Our 3 with the holiday update still drives fine on the freeway with AP. When it's dry and clear, beta seems to be smoother than it was, but it still wants to curb my wheels on right turns with right angled curbs.

The V11 UI is dangerous and will cause accidents and possibly fatalities. The font is now so small that I--and others I know--have to stare at many display elements rather than glance at them. That is dangerous. The defrosters, the wipers, the seat heaters, the trip odometers, and the tire pressures are now buried in menus. The notifications and other text based displays are now so tiny that significant time will be spent reading them rather than glancing at them. The blind spot camera is a good idea but a total fail in execution. It's too small, in the wrong spot, and unreadable at night in the rain with the blinker on. What's worse, is that Tesla got rid of the red warnings for cars in your blind spot when the turn signal is on. All of this means that your eyes will spend WAY too much time inside rather than out.

For those who have excellent intermediate vision you may not recognize just how dangerous this update is. But please remember that well over half of other drivers on the road do not have excellent intermediate vision. Your safety depends on those people's driving as well as your own. There is no reason--none--for the font to be that small. Bear in mind that I am a Tesla fan, and that I have many years of professional experience relevant to analyzing UIs. Overall, this update is dangerous enough to warrant an immediate recall. WTH happened to Tesla's commitment to safety?
 
Is your 3 with Radar ? If so, they will be different.
Yes, and that is one difference and probably why the even the TACC on the Y is now useless in the rain. But there is more than that. Immediately upon getting beta, I noticed that AP on the freeway did a much better job not swinging right into expanded right lanes from on ramps. The Y AP also does a much, much better job when several lanes merge into one. There is one freeway interchange here that the Y could never take before, and the 3 still can't. Once the Y got beta it could then successfully navigate the interchange. Prior to 10.8, The Y's NAO on the freeway was much better in many places than the 3. BTW, the 3 does not have the AP problems in the rain as far as lane keeping that the Y has now. That's not radar. Beta is not single stack, but beta equipped cars are definitely using different software on the freeway compared to non-beta cars. The difference is just too noticeable and consistent to be random.
 
Did the lanes change recently? The image you show has a straight arrow and a straight+right arrow for the two westbound 14th lanes crossing Peachtree whereas this November 2021 Street View seems to show two straight arrows?....
No, the lane assignments did not change and have been Left lane straight only (no left turn allowed) and Right lane straight AND right turn. Here is Apple Maps showing a car turning right and the next pic shows about the point when it turns on the left signal and stats moving over and then heads towards the island. Even if the Arrow was panted with a right tun added to the straight arrow this should not confuse Beta into thinking it is in the wrong lane.

Still it is crystal clear that is not a place to be changing lanes.

Screen Shot 2021-12-27 at 6.24.07 AM.png

Screen Shot 2021-12-27 at 6.25.20 AM.png
 
Yes, and that is one difference and probably why the even the TACC on the Y is now useless in the rain. But there is more than that. Immediately upon getting beta, I noticed that AP on the freeway did a much better job not swinging right into expanded right lanes from on ramps. The Y AP also does a much, much better job when several lanes merge into one. There is one freeway interchange here that the Y could never take before, and the 3 still can't. Once the Y got beta it could then successfully navigate the interchange. Prior to 10.8, The Y's NAO on the freeway was much better in many places than the 3. BTW, the 3 does not have the AP problems in the rain as far as lane keeping that the Y has now. That's not radar. Beta is not single stack, but beta equipped cars are definitely using different software on the freeway compared to non-beta cars. The difference is just too noticeable and consistent to be random.
You are still all the time comparing vision and radar AP. You need to compare VO AP of FSD beta and non-beta cars.

It’s quite likely VO AP is changing with every release. It’s probably changing for both FSD and non-FSD cars.
 
You are still all the time comparing vision and radar AP. You need to compare VO AP of FSD beta and non-beta cars.

It’s quite likely VO AP is changing with every release. It’s probably changing for both FSD and non-FSD cars.
I see what you're saying and you could be entirely correct; I have no way to test that. I can say with some certainty that you are correct in that VO AP is changing with every new beta release and is different than my radar equipped 3. When I first got beta, the change to AP was a very noticeable and appreciated improvement. VO AP and TACC with Beta 10.8 on the other hand is a huge downgrade in the rain, to the point that I don't feel it's safe to have on at all in the rain.
 
  • Informative
  • Like
Reactions: impastu and EVNow
Drove short distance today, overcast daylight; acceleration set to standard; FSD set to average. The drive was smoother that 10.5 the car seemed to make decisions faster. Still crossing the yellow center line on some right turns (not all). fewer phantom slowing's (possibly because of overcast or could be getting better). I need to check my drivers manual on this one thing, when in a designated right turn lane is it necessary to use a turn signal? The car did not use the turn signal and I think that is correct.
 
10.8 so far for me:
  • Less wheel jerking in various low speed situations.
  • Better at using bike lane for right turns to leave room for traffic going another direction next to my car (still needs more work).
  • New: Incorrectly tries to use center divider lane (crosses double yellow line) when approaching intersections for a left turn.
  • Slightly better turns through intersections.
  • Still ignoring No Turn On Red signs.
  • Phantom braking slightly improved.
  • Otherwise generally very similar to the prior version (10.5 for me).
 
Last edited: