-
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
Yeah but you were having that same problem when you drove a Lexus so not sure it is a Tesla or FSDS 12.3.4 thing.
Trip report on 12.3.3. And it's a bit of a doozy: From NJ to Dallas, TX; from there to Southern Indiana, and then back to NJ. Yes, a lot of highway driving. 2021 AWD LR MY with USS. It got the 2024.3.10/12.3.3 update about two days before we left.
First: A heck of a lot of highway driving.
Speed limit bug cured: Before 12.3.3, FSDS had trouble with those double speed limit signs. You know, the type that has, on the top, a 50 MPH limit and, below, a little placard that said TRUCKS and a different, lower speed limit. Some certain percentage of the time one would get the truck speed limit inserted into the car instead of the car limit. If one happened to glance at the display when passing such a sign, one would see two such signs, jittering around rapidly, and the car would, apparently, pick the last one passed. Or something. And pick the wrong one in some proportion. On this trip, whenever encountering one of those signs, there was only one speed limit sign on the display, not two, and it was (at least for me) the correct, "car" one. So, that appears to be one bug squashed.
Dry wipes. Sigh. The SO's car, which had been running EAP from last year before the 12.3.3 install, had been notably free of dry wipes. I mean, once in a great while, but not as often as my daily driver with $RANDOM FSD(S)'s on it. On the way out to Dallas, going through Rain in all its permutations, started getting dry wipes. I had guessed that the excessive numbers of bugs on the windshield (spring, right?) might have been responsible for that, and cleaned the windshield thoroughly. Followed by much fewer wipes. On the way back, it was Springy April Showers (and thunderbumpers, and deluges, and generalized sprinkles) all the way back, with occasional long stretches of dry weather, but clouds overhead. The dry wipes came back with a vengance.
Interestingly, if it was raining, the wiping software did a vaguely creditiable job, although at exactly the right level light level it might not start up on its own; below that level, it would start before the window got obscured; above that level, it would start with no problems and keep up.
But if the rain completely stopped, after a couple-ten miles of No Rain, it would start dry wiping. Definitely no insects at this point. So, eventually, would do the right scroll wheel and "Wipers off" and then, when the rain picked up again, "Wipers Auto".
I wonder.. When the sky was sunny, it wouldn't dry wipe. But with all the clouds up there, it would. Hmm. Somewhere at Tesla, there's somebody trying and trying to get this algorithm to work right, and that somebody has ulcers. I suspect that if there's a Sun in the sky, that's a factor in trying to turn off the wipers.
All I can think of at the moment. In case any of you wondered: South of Dallas the weather report was Clouds during the eclipse. There were highish clouds all morning, but they burned off around noon, apparently a common event in those parts. This left rather large fair-weather cumulus clouds with 90% coverage of the sky, with a very rare bit of blue sky. There were supposed to be Major Thunderstorms starting around 5 p.m., not a good sign. Unstable atmosphere, the weatherpeople kept on repeating.
Then, the eclipse started around 12:30. Almost immediately, the clouds started shrinking. By 12:45 it was more like 50/50 coverage and shrinking and, with a pair of eclipse glasses, one could make out the sun with a cut-out with no problems. And then.. starting from the South, all those blame clouds up and disappeared. By 1 p.m., they were almost gone; by 1:15 pm, they were gone. Look to the East; look to the West: There were clouds out there, far, far away. But none along the path of the eclipse. It got noticeably cooler.
Clear skies and a total eclipse. Notable, as per usual, the solar corona streaming away from that eyeball in the sky. The bottom right quadrant had a red prominance over the edge of the moon: Not an artifact, that sucker was red.
And the sky stayed clear of clouds for the rest of the day. There was a thunderstorm around 2 a.m., but that was it.
Guess all those weather mavens need to add "eclipse shadow" to their computer models.
First: A heck of a lot of highway driving.
- No ping-pongs. The car stayed in lane and didn't wander back and forth. Interestingly, before we left, on the advice of some $RANDOM internet page, the car was taken into a local AAA facility and its alignment checked. Turned out that the toe on two wheels, front and rear, was incorrect. The SO and I (this is her daily driver) don't remember hitting any potholes or curbs over the past few years and the car doesn't have any rim rash. So, as delivered? A pothole that we forgot about? Who knows. But no wandering on the road.
- Not sure why, but the car in "Average" FSD mode, but on the highway where supposedly the driving stack is straight out of the EAP that we've had on this car forever, likes to set the initial speed to 10% high. Religiously. Speed limit 65? It sets to 72. Speed limit 70? Set to 77. Speed limit 75 (hello, Texas/Arkansas!), set to 82. That last one gave a little trouble, since, as others have noted, if one is doing TACC/LK at 85, the car gets mad and disables both until one stops and parks somewhere. That happened twice, once with the SO, once with me. In my case it was to speed up to pass two semis; the one in front was going slow and the one behind the slow one wanted the left lane, so, in a fit of Being Nice, sped up so the rear semi wouldn't have to brake hard. Memo: Turn off Autopilot with the shift lever before doing that. The SO's experience was similar. Luckily, in both cases, it was 30 miles or so until the next stop, so no biggie. Except, I guess, that TACC/LK does a superior job keeping the car centered in the presence of cross winds and the like, so it feels less safe without the automatic stuff turned on.
- As delivered, the FSDS on highways really likes the left lane on two forward lanes.
- It very definitely will shift from that left lane to the right lane when someone is coming up from behind, without prompting.
- If one is in the right lane the car, from time to time and, as far as i can tell, spontaneously, will shift to the left lane, with the message, "Shift to avoid merging traffic." That's not a bad idea in principle.. but when it does it five or seven miles before the next interchange, it raises eyeballs.
- If one is in the right lane (typically, by hitting the turn signal and moving there at some point) and one comes up on an intersection, one can often actually see cars on the on-ramp and with good old wet-ware, figure that the left lane would be a better place to be. The car doesn't typically see those on-ramping cars until one is right on top of them, at which point braking, semi-confusion, and all that kick in. So maybe the Tesla guys tried to avoid all that sturm and drang by just hanging out on the left in the first place. Which makes sense, I guess, except it kind of violates the concept of those signs that say, "Left lane for passing only" that one sees periodically.
- Generally, as one is chugging along and there's a vehicle up in front, the car is Very Good about getting into the left lane in plenty of time to pass said vehicle. If there's nobody behind one after doing the pass, then it stays where it is; if there is somebody coming up, then it will shift to the right. I did a lot of manual shifting back to the right lane with the turn signal lever, being that kind of guy.
- Speaking on that last: Say that there's some car up ahead on the right that's going a bit slower than one, and there's somebody behind one that's coming up at a somewhat faster rate of speed. FSDS will definitely shift to the right, slow down a bit so the faster car can get past (and any other cars in train with that car), then shift back to the left and pass the bit slower car. I've seen plenty of humans that aren't that considerate.
- Only real interventions were for the obvious stuff, and only happened a couple of times on the +2000 mile trip. An obvious case: three forward lanes, semis galore in the right lane, car going a mile or two above the speed limit (normal) in the center lane, and an idiot in the left lane, stuck on the left rear corner of that center lane car, resulting in backed-up traffic all over. Sigh. Move into the center lane, move up, intervene, move up far enough so that the idiot is behind, then speed up, move left, and pass at speed. Then get back into the center lane. In the rear view mirror, see idiot wake up, move up, get over, and a plethora of cars chug on down the left lane. Sheesh. How the heck is FSDS going to handle idiots?
Speed limit bug cured: Before 12.3.3, FSDS had trouble with those double speed limit signs. You know, the type that has, on the top, a 50 MPH limit and, below, a little placard that said TRUCKS and a different, lower speed limit. Some certain percentage of the time one would get the truck speed limit inserted into the car instead of the car limit. If one happened to glance at the display when passing such a sign, one would see two such signs, jittering around rapidly, and the car would, apparently, pick the last one passed. Or something. And pick the wrong one in some proportion. On this trip, whenever encountering one of those signs, there was only one speed limit sign on the display, not two, and it was (at least for me) the correct, "car" one. So, that appears to be one bug squashed.
Dry wipes. Sigh. The SO's car, which had been running EAP from last year before the 12.3.3 install, had been notably free of dry wipes. I mean, once in a great while, but not as often as my daily driver with $RANDOM FSD(S)'s on it. On the way out to Dallas, going through Rain in all its permutations, started getting dry wipes. I had guessed that the excessive numbers of bugs on the windshield (spring, right?) might have been responsible for that, and cleaned the windshield thoroughly. Followed by much fewer wipes. On the way back, it was Springy April Showers (and thunderbumpers, and deluges, and generalized sprinkles) all the way back, with occasional long stretches of dry weather, but clouds overhead. The dry wipes came back with a vengance.
Interestingly, if it was raining, the wiping software did a vaguely creditiable job, although at exactly the right level light level it might not start up on its own; below that level, it would start before the window got obscured; above that level, it would start with no problems and keep up.
But if the rain completely stopped, after a couple-ten miles of No Rain, it would start dry wiping. Definitely no insects at this point. So, eventually, would do the right scroll wheel and "Wipers off" and then, when the rain picked up again, "Wipers Auto".
I wonder.. When the sky was sunny, it wouldn't dry wipe. But with all the clouds up there, it would. Hmm. Somewhere at Tesla, there's somebody trying and trying to get this algorithm to work right, and that somebody has ulcers. I suspect that if there's a Sun in the sky, that's a factor in trying to turn off the wipers.
All I can think of at the moment. In case any of you wondered: South of Dallas the weather report was Clouds during the eclipse. There were highish clouds all morning, but they burned off around noon, apparently a common event in those parts. This left rather large fair-weather cumulus clouds with 90% coverage of the sky, with a very rare bit of blue sky. There were supposed to be Major Thunderstorms starting around 5 p.m., not a good sign. Unstable atmosphere, the weatherpeople kept on repeating.
Then, the eclipse started around 12:30. Almost immediately, the clouds started shrinking. By 12:45 it was more like 50/50 coverage and shrinking and, with a pair of eclipse glasses, one could make out the sun with a cut-out with no problems. And then.. starting from the South, all those blame clouds up and disappeared. By 1 p.m., they were almost gone; by 1:15 pm, they were gone. Look to the East; look to the West: There were clouds out there, far, far away. But none along the path of the eclipse. It got noticeably cooler.
Clear skies and a total eclipse. Notable, as per usual, the solar corona streaming away from that eyeball in the sky. The bottom right quadrant had a red prominance over the edge of the moon: Not an artifact, that sucker was red.
And the sky stayed clear of clouds for the rest of the day. There was a thunderstorm around 2 a.m., but that was it.
Guess all those weather mavens need to add "eclipse shadow" to their computer models.
Speaking of Strips Club(although I like to called them “Gentlemen’s Clubs”), Interesting that I have ZERO interventions, Zero Wheel nags, and ZERO issues with AutoMax driving like a 80yr old when the destination is the aforementioned.
I think I’ve discovered how to unlock Elon Mode!!!
FSDtester#1
Craves Electrolytes
Makes sense, when I go to see my parole officer, constant nags, interventions, disengagements... Ego doesn't want me going there...
j0shm1lls
Member
I checked now just for you... I'm still being new build blocked.
All those ADAS drivers at Chuck's unprotected left so far have been a nothing burger. Chuck's first attempt had multiple issues as well as his normal u-turn attempt thereafter. Also missed one speed bump in that narrow street neighborhood he drives through.
And the other unprotected left turn after the narrow neighborhood had FSD dramatically slowing in the roadway after entering at least 40mph traffic. Another speed control SNAFU with a max set speed of 5mph. Chuck says it's "dangerous" and wasn't there in v12.3.3.
In some ways v12.3.4 seems to be worse than v12.3.3. Back to the drawing boards.
And the other unprotected left turn after the narrow neighborhood had FSD dramatically slowing in the roadway after entering at least 40mph traffic. Another speed control SNAFU with a max set speed of 5mph. Chuck says it's "dangerous" and wasn't there in v12.3.3.
In some ways v12.3.4 seems to be worse than v12.3.3. Back to the drawing boards.
Last edited:
FSDtester#1
Craves Electrolytes
I'm sorry if this is a really dumb question. But you know to only check for updates when you're on a good Wi-Fi connection, right?
If anyone tells you it doesn't matter. They're wrong..
I am like Joe Isuzu, would I lie???
Tesla should teach FSD know how to go around the block and to do the easy way instead of the hard way.
AlanSubie4Life
Efficiency Obsessed Member
While Chuck may state otherwise in the videos, there's no evidence that any retraining of FSD for this release has been done.
Why would they?
We'll know when they retrain. The car will drive differently! Note that there have been no driving behavior changes for any 12.3.x release so far (only changes in speeds, etc., a few other behaviors adjustable outside of training, using the extensive C code). There are potentially hundreds of thousands of lines of code.
I'm looking forward to seeing the results of all those ADAS drivers at the turn. It'll be interesting to see if things improve once they release that end product.
Yeah this could just be a fluke or due to C code changes.
Using manual mode would have allowed scroll wheel adjustment rather than accelerator input (not that it matters at all).
This release was for additional cars and USB key (dash cam) fixes (probably the big fix - we'll see if it worked, going to be extremely interesting).
No need to go back to drawing boards. Let's just wait for a new release of FSDS. Then we can see - it'll be the first test of end-to-end.
So far, 12.3, 12.3.3 and likely 12.3.4 (haven't tried yet) seem to drive identically. There are just small changes in behaviors/assertiveness, etc. The fundamental driving behavior is the same though.
Last edited:
Two obnoxious behaviours I hope they fix in 12.4:
1) When the "FSD degraded" message comes on, it has to come with an annoying "boop boop boop". Sometimes all it takes is for there to be a few clouds for it to say it's degraded. As soon as the message goes away, it comes back with another "boop boop boop". Sometimes it's beeping every 10 seconds. Farking annoying so I have to disengage.
2) I often disengage to go around potholes and such. Unfortunately, the car will start spazzing out about crossing a lane even though I have lane departure warnings turned off. This flag seems active for a minute+. After the first time it does this, it won't spazz out again, but it seems that some flag gets set with exiting FSD. I understand in case the driver fell asleep or something, but they have the cabin camera to check for that.
1) When the "FSD degraded" message comes on, it has to come with an annoying "boop boop boop". Sometimes all it takes is for there to be a few clouds for it to say it's degraded. As soon as the message goes away, it comes back with another "boop boop boop". Sometimes it's beeping every 10 seconds. Farking annoying so I have to disengage.
2) I often disengage to go around potholes and such. Unfortunately, the car will start spazzing out about crossing a lane even though I have lane departure warnings turned off. This flag seems active for a minute+. After the first time it does this, it won't spazz out again, but it seems that some flag gets set with exiting FSD. I understand in case the driver fell asleep or something, but they have the cabin camera to check for that.
Chuck said he noticed a couple of differences with the first unprotected left as well as the u-turn so those are driving behavior changes. He called it a slight regression from v1.2.3.3 and I would have to agree.
The ADAS drivers have been visiting Chuck's unprotected left for probably a year now. I think there were better performing v11.x versions for his unprotected left. I'm not convinced the C code bypass made a significant improvement. They still have much to prove with v12 and so far releases are back to the slow, slim to none to category of improvement.
AlanSubie4Life
Efficiency Obsessed Member
When I drive 12.3.4 I guess I'll have to see whether it seems different on my turn (which it does acceptably well except for the 10-20 second approach, and the occasional unpredictable stopping in traffic lanes (haven't seen in a week or so)). It's generally better than v11 was on my turn, and is notably different - it completely ignores the median filter lane, so now it drives the turn totally differently than I would - but is less jerky and robotic than v11.
I think it's very difficult to judge from very small sample sizes. I didn't see any obvious differences on that first unprotected left (really all I watched). It looked the same. Yeah it creeped out too far into traffic from the right, but that seems well within expected variation and Chuck has only done the turn 20-40 times so he may not have seen it before - it's just such a tiny sample (and many of his turn attempts so far have had basically zero traffic from the right).
I think doing the turn 200-300 times makes more sense - then you can probably assess differences on a new release.
I suppose it's possible they can attempt to retrain only small sections of behavior, but no idea how that works.
Even if possible, I doubt it is happening in 12.3.x.
Last edited:
You're just a perennial ray of sunshine. I'm interested in what they can do with performance and behavior just like everyone else here, but we've seen only one version of the software, and that's V12.3. When we've seen 12.4 and 12.5, then we may have enough information to make some basic observations about the pace of progress. But even then we won't know how much the development team is learning about this stuff. It's not like there's a book on how to implement an end to end vehicle autonomy system. Right now, they look to be immersed in procedural work, ensuring that all models have a working version, sending out the free trial, and so on.
According Tesla 12 percent of Tesla fleet on 2023.8.x Branch still doesn't have V12. They promise to release in the next couple days.
Last edited:
Think 12.3.3 and 12.3.4 both have the v11 highway stack
Cheburashka
Active Member
Wow v12 is awful for me here in the SF Bay area. v11 was much much better.
- kept driving over the ribbed part, over the lane markers, and over the botts dots on Hwy 17.
- blew right past the 17/85 interchange
- auto speed limit was set to 65 in the 50 zone where cops always camp on 17
- missed the exit from 85 to El Camino in Mountain View
- at intersections applies way too much throttle mid right turn for seemingly no reason
- was going through a roundabout and got the red both hands on wheel message out of nowhere
- got a Toollbooth detected message on 17 before a pedo overpass
I wouldn't pay even $49.99 for this.
- kept driving over the ribbed part, over the lane markers, and over the botts dots on Hwy 17.
- blew right past the 17/85 interchange
- auto speed limit was set to 65 in the 50 zone where cops always camp on 17
- missed the exit from 85 to El Camino in Mountain View
- at intersections applies way too much throttle mid right turn for seemingly no reason
- was going through a roundabout and got the red both hands on wheel message out of nowhere
- got a Toollbooth detected message on 17 before a pedo overpass
I wouldn't pay even $49.99 for this.
Similar threads
- Article
