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.
Part of the new testers group here and wanted to share my initial experience.

Want to start by saying that it is incredible how far tesla has come with FSD. Given that this is the worst it will ever be, there is a lot to look forward to.
With that being said - as far as my experience goes, this software is not ready for wide release yet and I would be very nervous if Tesla starts adding 1000/day like Elon mentioned. I might catch some heat for this but I cant say with confidence that they should add any more people even with v10.3.

I understand experiences vary regionally but on my not too complex drives, it has been a little scary.
My observations on the last few drives:

- Making a right turn onto a main road from my neighborhood: No cars coming, the car turned too fast over turning into the left lane, then trying to correct itself back into the right lane pretty aggressively and getting inches away from the curb on the right. I disengaged.

- Turning into a one way "Do not enter" McDonald's exit: The entrance was a little further up. I think this is okay for now as some maps don't really give you the exact entrance to the destination.

- Making a right turn at a red light: The car stopped a little before the line, and then sped its way onto the road. I could hardly see the oncoming traffic, so I cant imagine the car saw it. There was a car coming, I pressed the accelerator and disengaged. If I wasn't very alert, I can see this easily leading to a collision.

- Making an unprotected left into a shopping center: The car had to cross a 4 lane road, and there are always plenty of gaps with no oncoming cars. The car was too slow and then started slowly crossing when a far pretty far started getting super close. I helped with the accelerator. It was a Model Y so I hope they would've understood...

- Making a right turn into a side street: The car starts driving towards the left side of the lane like it wants to give itself more space? Almost feels like the car is about to cross over to the left lane while trying to make a right onto the street. I disengaged here as I got scared. Happened a few times.

- Some other things: Plenty of steering wheel twitching, often towards the opposite direction of the intended turn. Much slower initiating a turn but then taking it pretty fast. hesitating and slow execution when crossing over multiple lanes.

There have been instances where it has done things really well, some turns, stopping & going at lights, changing lanes and stopping behind cars has improved significantly and feels much more natural. Overall, it's very exciting to see how far Tesla has come and plenty to look forward to. Seeing where its at today, I don't doubt one bit that Tesla will achieve its goals with FDS and eventually RoboTaxis. But I am nervous of any further expansion in the coming weeks
 
I'm actually referring to severe slowdowns (like going from 35 to 20, for no apparent reason).
Yes, I was referring to that above as well (in the first paragraph). Those types of events were evident in the early FSD beta videos, but it’s often not commented on. I often override with accelerator, especially on my unprotected left roller coaster ride (reminds me of one of those old-school rides which jerk around corners on the track), but have not yet had a disengagement due to doing that.

The pulsing is something entirely different, and so far I have only noticed it once, on an slight uphill grade at 45mph (the same section of road pictured in my 49.9% unsafe following example I posted). I noticed 3 pulses in that instance.
 
Last edited:
I had a great, but not a lerfect. FSD drive yesterday on city streets and Freeway.
Well, nobody's lerfect. ;)
because it knew that within 0.1 miles of the right turn it needed then turn left onto another road.
See quoted post below and response. :)
I had a what the hell is it doing now session with FSD Beta. There were three lanes in my direction going 45 mph. I had been in the right lane for a while, and a right turn was coming up in 0.3 miles. FSD decided to move over one lane to the left. I thought maybe it was going around the car ahead of me and would merge back in, but it didn't try to overtake the other car. It just hung out in that lane. Meanwhile, that right turn is rapidly approaching, so I used the turn signal to force it back into the right lane. Again, FSD moves over to the left lane. This time it is kind of overtaking the car ahead, but barely. I'm about 0.1 miles from the right turn now, and there is only one obvious hole in the right lane traffic ahead of me, so I forced fsd with turn signal again into the right lane. Right then the right turn only sign appeared for that lane. Fsd was able to get left again in time and make the turn it had always wanted to.
This part of FSD Beta is one you gotta watch out for. Sometimes, it feels like it is actually planning its lane usage according to upcoming turns. Other times, it's completely out to lunch. I've had it switch out of the right lane to the middle lane when a right turn is less than .1 miles away, with no break in right lane traffic available to get us back in the lane in time for the turn. It does this very often. Today I'm going to put the lane change for slower traffic setting down to minimum and see if it helps.

I've been reporting these with the camera icon, but I'm wondering if the people/computers that examine the footage are actually able to put into context why the event was reported with the camera icon in the first place. In fact, there are a ton of instances where I've hit the camera icon and thought, "good luck putting that into context."
how well do we expect computer AI to read social cues like this one with the dog walker signaling “go ahead first” by just stop walking (i.e. - not waving at me to go first)?
Great example of something that we do all the time as human drivers without ever even thinking about it.

On a similar note, I can't tell you how many times I've avoided being in an accident just by looking at the driver in the potential accident car; just being able to see what the other driver is looking at in certain situations can be the deciding factor on whether or not a collision happens.
 
I know this is a stupid question I’m probably wasting my time but has anyone with a 2021 refreshed LR or plaid gotten the beta yet? It’s been about 20 pages since I last checked but I’m still driving my Plaid like a 90-year-old with a very slim chance that I’ll get it
 
This is kind of the holy grail of FSD. The second one is tougher then the first.
- How to predict what others are going to do
- How to predict how their actions will change based on your action

I thought this was actually addressed pretty well at AI day (at least for other vehicles, pedestrians will be another matter).


He states in scenarios where FSD needs to guess where another driver wants to go, it runs the planning algorithm from their point-of-view. This might not work very well right now because the planning algorithm tends to make movement decisions unlike a human, but as the planning algorithm gets better, so too will the vehicle's ability to intuit another driver's intentions.
 
A note of caution for FSD Beta testers is that neural networks can behave quite differently even for the same intersection that it has done fine multiple times in the past (even on the same version), so you'll need to pay extra attention especially if the visualized path prediction seems to be jumping around.

For example, I have an odd angled left turn where one time FSD Beta 10.2 tried to drive into oncoming traffic lanes. It didn't require a disengagement other times because sometimes there was oncoming traffic already waiting giving the neural network a hint that it should find a different path, and other times there was a lead vehicle acting as a guide of what path to take.
 
Is there a way to get it to only creep forward without making a turn?? There is a big one I go on every time I leave the house and I am trying to figure out how it thinks it has enough visibility- maybe it isn’t calculating that cars are coming on left from behind a hill 10mph over speed limit? Not sure how to attach the video I think it is too high resolution.
 
  • Like
Reactions: ahoen117
I thought this was actually addressed pretty well at AI day (at least for other vehicles, pedestrians will be another matter).


He states in scenarios where FSD needs to guess where another driver wants to go, it runs the planning algorithm from their point-of-view. This might not work very well right now because the planning algorithm tends to make movement decisions unlike a human, but as the planning algorithm gets better, so too will the vehicle's ability to intuit another driver's intentions.
Yes - that first part (figuring out what others might do) - is something everyone tries to do. The second part i.e. others behavior conditional on some action on your part is very difficult and few have started doing that.

BTW, this was as stated by George Holtz (comma.io).

ps :

You are talking about the static world. In the AI interview with George Hotz, he explains this well.
- Static world
- Dynamic world, where lots of actors are moving and you have to predict their behavior
- your action is going to change the way the various actors are going to change their behavior and you have to predict the changes depending on your choices

pps : You could potentially go further - how your behavior would change their behavior which in turn could change your behavior etc. Its like in Chess - how far ahead you think - how many future moves do you simulate. AI chess players have shown they can do this much better than humans - thinking far ahead and coming up with the best moves.
 
Last edited:
A note of caution for FSD Beta testers is that neural networks can behave quite differently even for the same intersection that it has done fine multiple times in the past (even on the same version), so you'll need to pay extra attention especially if the visualized path prediction seems to be jumping around.
IMO, that's worth yelling out loud.

We've all gotten used to the public release, and this ain't the same. Sure, the public release will occasionally not demonstrate consistent behavior, but for the most part, it does; I know exactly where it's going to automatically click off, take a curve too wide, etc. Quite predictable.

But with the Beta, you're very likely to get a different result VERY often.

The time you think, "Oh, I know it's gonna do xxx here," you'll find yourself in a ditch before you can finish the thought. ;)

Not to sound like a nervous nelly, etc., but you just gotta stay on top of the AP (Beta or not) ALL THE TIME. Drive this thing like its primary goal is to kill you when you least expect it. 'Cause that's exactly what it'll do.
 
Last edited:
Is there a way to get it to only creep forward without making a turn?? There is a big one I go on every time I leave the house and I am trying to figure out how it thinks it has enough visibility- maybe it isn’t calculating that cars are coming on left from behind a hill 10mph over speed limit? Not sure how to attach the video I think it is too high resolution.
No - we can't control its creep etc.

You can upload the video to youtube and put the link here.
 
  • Like
Reactions: jebinc
Then it also does this weird thing in my area. My neighborhood has these corners where instead of the street just doing a 90 degree turn, there is a big cul-de-sac area for the turn. The fsd treats it as a stop sign which it should not. Again, nobody was behind me so I let it go, and I'd say it stops there about 80% of the time. The other 20% it does it correctly.

Same, also Austin.
 
One last thought about Beta FSD and yokers...

I've posted a bit about how I thought that the spinning yoke could be the reason (along with the Frankenstein hardware merge problem) that it hasn't been released for yoke cars.

A few people have countered that the disengagement torque is so minimal that you're not gonna get hurt.

Those cheap floor fans you can get at Walmart? If a guy were to hold his finger against a blade when he then turned the fan on would be able to completely prevent the fan from starting to rotate. There just ain't much torque. Even so, you still wouldn't want to stick your finger in that fan when it's spinning.

Same applies to the yoke. You have the entire drive column's worth of rotating momentum to stop, in addition to the torque being applied by FSD. And if you were to reach up and try to stop that when FSD is having a moment, it's going to hurt.

HOWEVER, a guy here on the forums (my apologies for not being able to give full credit but I don't remember his name), came up with a solution so completely obvious that I feel like a complete idiot for not thinking about it.

Tap the brake.

The only time that the wheel really gets spinning fast is when the car is either stationary or nearly so. Just tap the brake and disengage the spinning yoke of death. ;)

This would take some time for me to reprogram the grey matter, as I'm so used to grabbing the wheel to stop its shenanigans. But with some practice, I think it would eventually become second nature.

This being the case, I no longer am of the opinion that the spinning yoke problem is the reason they haven't released the beta for yokers.

Only Elon knows.

And he ain't talking. :(
 
Is there a way to get it to only creep forward without making a turn??

No - we can't control its creep etc.

I guess it depends on what you mean. I'm always pushing the accelerator to get it to move forward enough to actually be able to see with the cameras, or get to the expected position to make a turn (which is often past the stop line - the stop line is something we as humans typically ignore or use only as a general guide - stopping there briefly before pushing forward). It's far too slow to approach many intersections to get visibility (it doesn't seem to know when it can see - though sometimes it proceeds with aplomb nevertheless), so I invariably help it out when it's dawdling. However, it might make the turn or go for it unexpectedly - the accelerator input can push it over the "go" threshold, it seems (this is actually what I'm trying to induce on my unprotected left). Definitely want to practice some more to see how this meld of machine and human works, exactly.

If you mean, can you get it to move forward without turning the wheel at all? No, it does what it wants with the wheel as you push the accelerator. You might reduce the amount of scrubbing if you push through the point where it is vacillating though.
 
Last edited:
  • Like
Reactions: Phlier
One last thought about Beta FSD and yokers...

I've posted a bit about how I thought that the spinning yoke could be the reason (along with the Frankenstein hardware merge problem) that it hasn't been released for yoke cars.

A few people have countered that the disengagement torque is so minimal that you're not gonna get hurt.

Those cheap floor fans you can get at Walmart? If a guy were to hold his finger against a blade when he then turned the fan on would be able to completely prevent the fan from starting to rotate. There just ain't much torque. Even so, you still wouldn't want to stick your finger in that fan when it's spinning.

Same applies to the yoke. You have the entire drive column's worth of rotating momentum to stop, in addition to the torque being applied by FSD. And if you were to reach up and try to stop that when FSD is having a moment, it's going to hurt.

HOWEVER, a guy here on the forums (my apologies for not being able to give full credit but I don't remember his name), came up with a solution so completely obvious that I feel like a complete idiot for not thinking about it.

Tap the brake.

The only time that the wheel really gets spinning fast is when the car is either stationary or nearly so. Just tap the brake and disengage the spinning wheel of death. ;)

This would take some time for me to reprogram the grey matter, as I'm so used to grabbing the wheel to stop its shenanigans. But with some practice, I think it would eventually become second nature.

This being the case, I no longer am of the opinion that the spinning yoke problem is the reason they haven't released the beta for yokers.

Only Elon knows.

And he ain't talking. :(
It's not just about the torque though. With a round wheel, you can keep your hand in one spot and clamp down instantly when you need to disengage. yes you can tap the brake, but you would still need to reach for whichever random portion of the yoke you can grab in the event a yoke car does something incorrectly.

If you tap the brake but aren't already holding the wheel, that doesn't automatically mean you're okay.
 
  • Like
Reactions: Matias and Phlier
It's not just about the torque though. With a round wheel, you can keep your hand in one spot and clamp down instantly when you need to disengage. yes you can tap the brake, but you would still need to reach for whichever random portion of the yoke you can grab in the event a yoke car does something incorrectly.

If you tap the brake but aren't already holding the wheel, that doesn't automatically mean you're okay.
Agreed.

It definitely isn't a perfect solution, but it is at least enough of a solution that I don't think it's the reason they wouldn't release the beta to yokers.
 
  • Like
Reactions: AlanSubie4Life