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

FSD Beta Videos (and questions for FSD Beta drivers)

This site may earn commission on affiliate links.
I'm sorry but while Kim is takeover happy. AIDriver is not an objective beta tester at all. He excuses alot of mistakes FSD Beta makes.
For example, It looks like FSD saw a ped crossing late and was very late to slow down for it leading to it having to slam on the brakes late.

AIDriver called it a "fluke" because it happened once during his drive. I'm like wtf. There are NO FLUKES. Another point he's talking about the ability to reroute when you can't make your turn. Which is great.

But right after there were several moments where it could have easily changed lanes and had alot of space and it didn't. Finally AIDriver makes mention of it, but completely glosses it over.. Infact at an earlier point he glorifies it. At 11:50 he says AP cancelled the lane change because of the stopped cars. However this was a mistake because it needs to make a right turn literally ~250 feet ahead and a human driver would have easily made this turn. which AIDRIVER called "dang impressive".

And then tack on his nonsense about oil conspiracy against tesla, lidar, hd map is doomed, no one else can do this or his consistent stupid jab at Ford Cruise.
Makes his videos almost unwatchable.

Following anything Tesla related is often maddening because of the lack of objectivity.

Whether its a beta tester or a poster in here.

There seems to be only a few people who are truly objective who can point out good things while also pointing out bad things.
 

Brandon’s vid. He started out by saying he had been having trouble with the system in the morning, so he re-calibrated the camera which seemed to help a lot. The Tesla handled the small roundabouts he had pretty flawlessly. Some routing issues due to incorrect lane selections later and one instance where the car might have been trying to go against a red, but he notes that it might have stopped but he wasn’t risking it. That might be one issue with the car being willing to go against the Nav, it might be resulting in the confusing behavior that Brandon and the guy in North Carolina were seeing where the car is deciding to maneuver around instead of slavishly following the GPS.
I live in roundabout country, and FSD take those way to slow and stakato. But it went through them at least, so probably an improvement.
 
Doesn't matter. You can't assume FSD Beta will handle a situation correctly. And other videos are not the same circumstances so they can't predict how FSD Beta will handle Kim's situations. The job of a FSD beta tester is to be safe, especially on public roads with other road users. Not to mention that it is there car, so if they get into an accident, it's their insurance that has to pay or them. Tesla can review disengagements later in the simulation to see if the disengagement was warranted or not, and make changes to next version as needed. Ultimately, if a safety driver gets into an issue where they feel they need to disengagement, then the FSD has already failed. FSD needs to be able to handle those situations confidently enough that the FSD Beta does not feel the need to disengage in the first place.
Great, but the video was shown as FSD being a system which needs too many interventions. You cannot conclude that from just watching the Kim’s video. AIdriver intervenes much less. The video of AIdriver in San Fransisco therefore shows much better the capabilities and limitations of the current system.
 
Don’t think so. She just doesn’t let the car do it’s thing. She intervenes too quickly. The video from AI DRIVR in San Fransisco shows the system is much more capable than Kim thinks.
I would certainly say those were for most part needed disengagements. She is looking for a safe, passive and comfy ride. Seems FSD isn't able to deliver that yet. Don't trust the unpredictable.

Look also how she has to adjust speed down to feel safe. System doesn't choose correct speed.

I am predicting that very soon one of the more mind blown test drivers are gonna crash because they try too hard to prove how great FSD is. But they rather have patience.
 
I live in roundabout country, and FSD take those way to slow and stakato. But it went through them at least, so probably an improvement.
To be fair, I’m pretty sure the roundabouts Brandon has have stop signs along the route he’s traveling on. Also, his car had a tendency to basically ignore the concrete circle on previous versions and would fairly consistently try to drive over it even with it displayed on the screen, so this was a big improvement.

Dirty Tesla’s drive had a very nice roundabout where it just blew through the yield sign because there weren’t any cars in the roundabout, which is a solid improvement from previous versions where it would treat yield signs as stop signs. Still need to see how it would handle traffic in roundabouts on a yield sign.
 
To be fair, I’m pretty sure the roundabouts Brandon has have stop signs along the route he’s traveling on. Dirty Tesla’s drive had a very nice roundabout where it just blew through the yield sign because there weren’t any cars in the roundabout, which is a solid improvement from previous versions where it would treat yield signs as stop signs. Still need to see how it would handle traffic in roundabouts on a yield sign.
It is not the stop for a stop sign I am reacting to, it is the stakato turning inside the roundabout.
Agree we need to see more of the ordinary ones, with and without traffic, with dense traffic, with 2 and 3 lanes, 3, 4 and 5 exits etc.
I even got AP2 to go through one back in 2018 without traffic!
 

tester in Milwaukee. Very weird, very wonky start.

5:25 - blows through a second stop sign after stopping appropriately at one a few meters before. I wonder if this is related to them maybe trying to fix a situation Frenchie had with the car stopping in the middle of the street because there were stop signs on both sides of the intersection?
 
  • Like
Reactions: DanCar
...tester in Milwaukee. ...
Just to note: This tester is Rob Maurer, who has one of the highest profiles of all Tesla-related YouTube creators with his Tesla Daily channel. Mostly finance / investment interpretations of daily Tesla company and product news, with a calm and balanced though generally Tesla-supportive take. Rob has also appeared on some national-TV business-news shows, providing knowledgeable balance to the usual zeroth-order, get-it-wrong Tesla headlines.

He was just recently added to the FSD beta-tester rolls and I think the v10 release may be his first opportunity to participate in early demo/review videos of FSD beta.

Why is this notable? Because his YT subscriber base is significantly wider than the FSD-interest and Tesla owner/enthusiast base of most other FSD testers online, so for better or worse it will expand the viewership of the FSD evolution. This is not to take anything away from the other hardworking public testers, many of whom have remarkable and informative FSD content. To Rob, I would suggest that his published FSD test videos could regularly fold in some newsy comments about what other testers are doing and discovering, which will expand the whole interest base of Tesla FSD development.
 
I don't think that is a major issue, but I can see other viewpoints. Left turns are tricky and should always be ready to take control. To me a major show stopper, is where it suddenly an inexplicably crosses the divider line heading straight, potentially causing a head on accident.
If anybody in the Bay Area with FSD Beta wants to try some roads that reproducibly do that on standard AutoPilot as a test case, try Redwood Estates, Old Santa Cruz Highway, Quito Rd., and County Road G8 near Uvas Reservoir (continuing onto Watsonville Rd.). Oh, and of course, CA-152. :D
 
  • Like
Reactions: Terminator857
She just doesn’t let the car do it’s thing. She intervenes too quickly.
At least she didn't need to intervene for the mid-intersection lane change issue that has been happening since at least January / Beta 8.1:
1631492765103.png


New Beta 10 behavior correctly stays in the right lane:
correct lanes.jpg


I believe this is with the new "Lanes" network to override the incorrect map data. If so, this is definitely worthy of the "Beta 10" version number bump as I would guess a lot of map data for the counts & types of lanes is incorrect/missing especially outside of the Bay Area, so this really does unlock a new level of capability in generalized planning.

I would also guess that this will be a significant source of regressions as this new Lanes neural network is taking on a lot of responsibility, but that should be "easily" fixed with more training data now that the network architecture can correctly predict it. Specifically from Karpathy's AI Day presentation, I would think this relied on the new Feature Queue that feeds into the Video Module to capture temporal and spatial information, e.g., driving past road markings and signage.
 

tester in Milwaukee. Very weird, very wonky start.

5:25 - blows through a second stop sign after stopping appropriately at one a few meters before. I wonder if this is related to them maybe trying to fix a situation Frenchie had with the car stopping in the middle of the street because there were stop signs on both sides of the intersection?
Good to see Rob added to the beta tester list. Probably the most listened to Tesla podcaster.
 
Anyone know/have a video of what "FSD" does when it encounters a yield sign, in the various cases of needing to yield vs not?
The YouTube channel "Dirty Tesla" has a very recent video (I believe it's his very first version 10 video) of a Yield sign at a roundabout entrance. The car... Ah, I'm not gonna spoil it. Go watch the vid. :)

Edit: OK, I'm not gonna make you go searching for the video... The Yield sign is in this one:



Oh, and btw... I'm not linking the time stamp for the Yield sign, because I think the entire video is worth watching. Especially noteworthy is how well the cameras do when the sun is at such an angle that the driver is darn near incapable of seeing out the windshield.
 
Last edited:
From the context of the tweet, the button was always just going to be a request for beta access rather than Tesla having to work through emails as they are doing now. I never got the impression that it would be a wide release, but rather limited to the around 10k they were aiming for the next group. Otherwise there is no point to the button, it'll just be a general release that you enable in the menu like all other features.
Depends on how far the firmware has diverged. It may or may not be feasible to keep both versions available simultaneously, in which case it would require a reinstall to switch modes.
 
Especially noteworthy is how well the cameras do when the sun is at such an angle that the driver is darn near incapable of seeing out the windshield.

I wonder if people who believe that Tesla will never achieve human level+ fsd with the current sensor suite realize that their concerns are being conquered by Tesla one by one.

I still many remember people who thought the cameras would get blinded by the sun, so Tesla would never achieve fsd.

Nowadays, some people are concerned about the b pillar camera being inadequate for unprotected lefts. Just watch.

On another note, we just saw in AI DRIVR’s video, V10 was able to do an unprotected right turn with:

Tons of pedestrians, bikers
A lane splitting motorcyclist within inches of the car
Cross traffic going up a hill
The v10 car turning into a narrow upwards-sloping road with backed up traffic
Etc. Etc.

Any single one of these circumstances would cause most fsd developers a lot of headache.
 
Last edited:
Dunning-Kruger Effect right? Let me guess you work some random job maybe real estate maybe some insurance guy who posts on the internet.

Are you ever not wrong man? :)

whatever. I’m a senior software engineer and project lead at one of the biggest automotive supplier.

That's adorable.

So you what write control code for fuel injectors for Bosch or something equally likely to not even be a job in 10 years? Would explain why you keep trying to FUD all over Tesla I suppose, keep your industry alive slightly longer...

What code, specifically, do you write for the self-driving elements of self-driving vehicles? How many neural networks have you designed or trained personally?

Is it none? Because it sounds like it's none.
 
Last edited:

Another from the apparent Tesla podcaster. Kind of a rough first left turn where it was kind of forcing itself in by creeping so far. A very impressive sequence where it screwed up by taking the wrong lane, fairly seamlessly turning right when it actually wanted to be in the left lane, and re-routing. It did pause after being forced to make the right turn, which is probably due to it waiting on the gps to give it another route. Goes back to what I said before: Tesla needs to have a few backup routes in the system ready to go that are recalculated after each maneuver so it has something to fall back on if it fails. It can’t just stop dead in the street waiting on a new route, and you also don’t want a situation where it keeps missing the turn on a re-route because it has passed the next turn by the time the gps updates.

Doesnt seem to handle flashing reds well yet. It was requiring driver confirmation to continue.

Huh, it pretty seamlessly changed two lanes over to get to a left turn lane immediately after a right turn. I’m not sure I’ve ever seen it do that before.

Wow. One of the better unprotected lefts I’ve seen. I was nervous as hell because it looked like the car was on the edge of deciding to try to go around the car in front of it, but it properly waited. And then it did one of the best lefts I’ve seen, anticipating a gap in traffic, slowly sliding forward as an oncoming car was passing, and then completing the turn after the car passed and before the next oncoming car could impede its progress. If the driver hadn’t specifically said he wasn’t doing anything, I would have thought he was goosing the accelerator.

And then another impressive maneuver pulling into a parking lot with people blocking the entrance to the lot. It waited well and then pulled in when it was clear.

Doesn‘t handle the narrow parking lot at all. It also doesn’t appear to see the barrels and chain blocking stuff off as it twice tries to accelerate through them. Tesla is going to need to expand the car’s recognition of obstacles like these barrels and Gali’s pillars in Seattle.

It does maneuver around a chain link fence blocking off part of the road, but the fence also had cones around it. I would have been curious to see if the car would have noticed the fence without the cones. It’s a toss up given examples of it missing other kinds of obstacles.

Forced disengagement on an unprotected left because a couple of pedestrians were confusing it. I was honestly confused myself as to whether or not they were going to cross at first since they looked constantly on the edge of doing so, and the pedestrians and also a an Intermittent stream of oncoming cars made that one tough. Driver eventually had to intervene and go with a Jimmy John’s delivery driver getting impatient behind him. I’m honestly not mad at the car taking so long here. Better safe and frustrating to people behind it than risking hitting a pedestrian in this phase with a driver who should be ready to take over. Obviously eventually this needs to be solved for anything above level 2, but for now good behavior, and far better than that first left.

‘Overall an impressive drive.
 
Last edited: