I think the give-and-take among drivers is a very difficult part of parking lot "proper" behavior for FSD to attempt to do properly. We look at each other and behave differently based on many, many factors. Even here in the same community different parking lots seem to play by different coordination rules.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD v12.x (end to end AI)
- Thread starter Buckminster
- Start date
-
- Tags
- #fixphantombraking autopilot chill is not chill elon is an ass fix fsd extreme late deceleration fsd does not commit to go fsd does not maintain proper speed fsd hard and late braking fsd jerky driving fsd late deceleration fsd stops too early too quickly fsd stops too far behind line fsd tailgating in city streets
sleepydoc
Well-Known Member
True. There's a difference between disengagements early in the release vs. later. Early disengagement are more likely to be FSD failures while later disengagements are more likely to be anticipatory. Like you, I generally know where FSD will fail and usually don't bother letting it try. If there's not much traffic I'll still let it try once in a while, just for fun, though. I don't know if this matters, but I generally don't bother sending a report for proactive failures. Maybe I should just say 'known failure point?'
Like I've said before, though - it should be possible for Tesla to identify problem/failure points based on disengagements and then obtain data from cars not on FSD to form a video training library.
It occurred to me that parking lot behavior was not specifically trained. It emerged from many videos where the driver parked. But since it wasn’t deliberate there might not have been enough examples to cover all situations. Now that we’ve had a parking lot incident they have to train in all parking lot variations too ensure it won’t happen again. This could be a feature they planned later on down the road but now they are forced to do it immediately. That’s why it takes more time to fix.
In an end-to-end system, there is no object perception network. The system literally takes pixels in and spits out control values. All the intermediate internal states are emergent from the training data. Getting an explanation of what the network is thinking would require neurons and training devoted to getting that explanation out. Interestingly, getting that explanation would alter the network.
The objects that we see in the visualization are from a separate network taken from V11. That system relied on heuristics designed by engineers. The engineers required recognizable objects so that they could express their heuristics in familiar terms, which is why an object perception network was created. That network's output provided a nice visualization, which I've always considered to be primarily a debugging tool which turned into a nice marketing gimmick. That also distracts drivers.
Last edited:
sleepydoc
Well-Known Member
ok - so if that's the case, the graphic display in V12 is really not representative of what the system is doing (except on the highway.) That would also mean the planned path display is not accurate, either, except it seems to be based on some of the mistake videos posted. How do you reconcile those?
I would expect the planned path to be part of the output the neural net creates. In v12 ive noticed different behavior of the path compared to V11.
Daniel in SD
(supervised)
I think the planned path display would be easy to train the end to end NN to output. This is because you can generate an error function of the difference between the path plan output and the control output.
The sub-revisions in the hundreds are almost always custom factory firmware and won't get deployed to existing vehicles.
As others have said, the path is coming from V12, but the rest of it is V11. They may have retained the V11 visualization specifically to provide context for the V12 path noodle so they'd know at least that much about what it was planning to do.
If you watch AI DRIVR videos, he's pretty clear in how the visualization that we see is not related to V12's perception of the world.
Anyone else tired of the lower quality onslaught of v12 videos? So many have just one "iPhone" mounted to the sunroof and that is the only view you get. At lest you get to see their sun visors and floorboards. 
Add to that the pure crap resolution/bitrate of My Space (what the hell is up with having sooooo POOR quality?) is so bad you can't make heads or tails out of the screen.
O'well maybe AIDRIVR will bless us with his 3ed video this weekend since we don't have much else to look forward too.
Any bets on 12.3 dribbling by Monday (notice I didn't specify which week
)?
Add to that the pure crap resolution/bitrate of My Space (what the hell is up with having sooooo POOR quality?) is so bad you can't make heads or tails out of the screen.O'well maybe AIDRIVR will bless us with his 3ed video this weekend since we don't have much else to look forward too.
Any bets on 12.3 dribbling by Monday (notice I didn't specify which week
)?
powertoold
Active Member
The decision wobble is an emergent bad behavior that is indicative of an inconsistent decision within the curated training videos (some videos turn left, others turn right)
The only way to fix it is to re-do the entire video set in any situation where the ego approaches a thin curved island and can turn left or right
My speculation
The only way to fix it is to re-do the entire video set in any situation where the ego approaches a thin curved island and can turn left or right
My speculation
Double pull can not be selected if FSD beta is enables. It can only be activated if fsd is diabled - like the note on the first screen shot says. Here is a picture of FSD disabled where double pull can then be enabled.
Attachments
Maybe something going to employees this weekend? Or, maybe not. The rumor mill has been pretty quiet lately.Anyone else tired of the lower quality onslaught of v12 videos? So many have just one "iPhone" mounted to the sunroof and that is the only view you get. At lest you get to see their sun visors and floorboards.
O'well maybe AIDRIVR will bless us with his 3ed video this weekend since we don't have much else to look forward too.
Any bets on 12.3 dribbling by Monday (notice I didn't specify which week
AlanSubie4Life
Efficiency Obsessed Member
Well this has been a fun foray into the world of AI.
But unfortunately it is now over. You know when James “Doom” Douma starts talking about how FSD v12 is going to change everything, it’s all over, and a v13 (Beta?) is basically guaranteed. Reliable indicator! Then repeat the hype cycle.
I am just looking forward to my mildly improved driver assist. No one is crazy enough to use v12 in parking lots, so who cares how it performs there, but maybe parking lots are a good training ground for the FSD team for figuring out how to avoid randomly wrong or indecisive behaviors. I’m not really looking forward to those on normal streets.
Also wonder if they had v12.5 planned for actual release to freeways? I guess they’ve reserved that enhancement for another hype cycle, where they can talk about how awesome and natural and incredibly safe it is on the freeways.
Seems to me we could get some idea of how they are training/limiting this NN, if we could get info from people with v12 about lane positioning, etc. As mentioned earlier, lane positioning should be all over the place (in a good way). E.g. Does it move for lane-splitting motorcycles in a timely manner (unlike v11, which moves for motorcycles)? If it’s not moving all over the place in lanes in response to slight shifts of surrounding traffic (people do this constantly and it is wonderful and good, but v11 oddly only does it for trucks), they’re putting some sort of guard rails on this thing, unfortunately.
I care much more about the basic behaviors than all the failures and amazing moves it makes. Who cares if it can get over four lanes quickly after a turn in no traffic? No one is going to let it do that anyway! It’s great that it does, no question about it. But less interesting. More interesting: Can it stop? Reliably, in a timely manner. Can it ease off? Can it go? I don’t know how exhaustively these features have been tested. Have seen some discussion, but people seem more fired up about all the “difficult” moves it makes successfully - which are less important to me. Is it actually comfortable? Would it pass a wife test, as long as you always disengaged when it could not handle the truly trickier situations? These are the things that matter. v11 fails all of them.
But unfortunately it is now over. You know when James “Doom” Douma starts talking about how FSD v12 is going to change everything, it’s all over, and a v13 (Beta?) is basically guaranteed. Reliable indicator! Then repeat the hype cycle.
I am just looking forward to my mildly improved driver assist. No one is crazy enough to use v12 in parking lots, so who cares how it performs there, but maybe parking lots are a good training ground for the FSD team for figuring out how to avoid randomly wrong or indecisive behaviors. I’m not really looking forward to those on normal streets.
Also wonder if they had v12.5 planned for actual release to freeways? I guess they’ve reserved that enhancement for another hype cycle, where they can talk about how awesome and natural and incredibly safe it is on the freeways.
Seems to me we could get some idea of how they are training/limiting this NN, if we could get info from people with v12 about lane positioning, etc. As mentioned earlier, lane positioning should be all over the place (in a good way). E.g. Does it move for lane-splitting motorcycles in a timely manner (unlike v11, which moves for motorcycles)? If it’s not moving all over the place in lanes in response to slight shifts of surrounding traffic (people do this constantly and it is wonderful and good, but v11 oddly only does it for trucks), they’re putting some sort of guard rails on this thing, unfortunately.
I care much more about the basic behaviors than all the failures and amazing moves it makes. Who cares if it can get over four lanes quickly after a turn in no traffic? No one is going to let it do that anyway! It’s great that it does, no question about it. But less interesting. More interesting: Can it stop? Reliably, in a timely manner. Can it ease off? Can it go? I don’t know how exhaustively these features have been tested. Have seen some discussion, but people seem more fired up about all the “difficult” moves it makes successfully - which are less important to me. Is it actually comfortable? Would it pass a wife test, as long as you always disengaged when it could not handle the truly trickier situations? These are the things that matter. v11 fails all of them.
Last edited:
I can say definitively that it does not pass the wife test yet. My poor wife has to suffer FSD everytime we go anywhere and invariably the car does at least one thing that makes her jump or yell. From my viewpoint as the driver they aren't much of a problem, I know I can grab back control anytime it goes too far, but these moves jitter up passenger nerves every time. They are much less frequent than V11 but even once makes passengers distrust the car.
powertoold
Active Member
Yes, but I feel a lot less safe without FSD. That's what I tell my wife, but she probably doesn't believe me
Like right now, my cameras are stuck at 95% FSD calibration, and I get a lot more anxiety and sensory overload driving without FSD.
AlanSubie4Life
Efficiency Obsessed Member
Sounds like a very permissive environment.
I am expressly forbidden from using v11. I have petitioned for use on freeways only, which has been granted on a probationary basis, but grudgingly, and use is not appreciated.
And I really can’t blame her.
She is dismayed at how horrible my driving has become with FSD in the loop, so she is pushing for complete banishment, to help restore allegedly impeccable & predictable driving behaviors.
In a way it is good, because it honestly is a lot more relaxing (and likely safer if I am not distracted somehow) to just drive myself, unless I am either on wide open freeway with no traffic, or stuck in a traffic jam. So it is good to be forced back to that flow state, with no FSD allowed.
I guess we will see what the first wide release version of v12 brings. It’s good to hear the bad moves may be less frequent - that gives me a better chance of being able to just disengage pre-emptively and pass the test, hybrid style.
I definitely have the opposite. Especially with sensory overload; it is too easy to get a bit out of sync with normal traffic checking and the “flow” of driving, and then I can find myself discombobulated and scrambling to catch up with the situation in the event of a required disengagement (which occur roughly every mile or so, of course). And anxiety for sure is higher in typical circumstances with FSD engaged. It’s simply a higher level one has to operate on to supervise properly.
Last edited:
Similar threads
- Article
