FSD v12.x (end to end AI)

[KArnold]

It seemed it moved forward back 1 foot but I was not sure.

Does your car or phone tracker show any backwards details?

 

[rlsd]

Does your car or phone tracker show any backwards details?

I forgot to save dashcam. Stupid me.

 

[primedive]

Not sure if these have been mentioned before. I have been using FSD v 12.3.3 for a while and have about 500 miles on it. I agree with all positive remarks on how much it has improved. Few areas where I am consistently seeing issues and will look out for future improvements :

1) Too late to signal ahead of turns, lane changes.

2) Too many unnecessary lane changes, even with the Minimal Lane Change setting taken for the trip. Sometimes I just want to stay in the lane moving at speed limit and flow with the traffic. FSD just wants to make lane changes unrelated to route following etc. When it signals I signal the other way and often the car has started moving in the direction FSD wanted to change to and then it swerves back to my lane which creates passenger discomfort as well as confusion for the cars behind.

3) Yellow caution speed limit signs that are placed ahead of tight turns/curves are ignored. This results in the car speeding through tight turns often causing risk of going out of lane. Either map data or cautionary speed limit signs need to be used in future to slow the car to speeds that are suitable for the curves ahead.

4) Multiple lane left or right turns at traffic lights - Rule is that each vehicle should keep to its lane using the turn radius markings on the road. If I'm on the left lane FSD initiates the turn and cuts over to the right side ending the turn in the right lane. This happens almost 100% when the right lane is free and it happened at least once with another vehicle on the right lane resulting in my car cutting in front of the other guy. Often it does not signal either and just creeps to the adjacent lane.

5) Creeps too close to the car ahead on freeway sometimes. EAP/Autopilot has the 1 to 7 setting on how close we like to follow car ahead. FSD settings are on most conservative/Chill side and it gets closer than what EAP would do on proximity setting = 7.

Overall I am very pleased with the progress and now use FSD approximately 95% of the time. I take over in known challenge situations and hope that these will get less and less over time.

I’ve experienced all of these as well, so it seems to not be uncommon. Now I just disengage the moment the car starts misbehaving in an inappropriate way. It’s better for safety with the other traffic and that way we can also send a voice report back to TSLA which will hopefully improve the system in the next release (or have them do a nav patch).

 

[primedive]

Maybe Tesla is using data where the car is aggressive in making unprotected turns and so V12 is learning that behavior. Here's the thing though: that aggressive behavior may be appropriate in certain scenarios but not in others.

Because of the lack of communication from the system on what it wants to do, I always keep my foot over the brake anytime there’s people walking near the car or I’m waiting on an unprotected turn. When it commits, it sometimes accelerates like we’re escaping out of prison, when we’re just turning onto a small road with one lane per direction. I would proceed more carefully and make sure nobody is coming around the parked cars. I also always put the thing in chill but it insists on racing. It doesn’t always do this so somehow the neural net just gets confused on what type of road we’re on.

 

[primedive]

Maybe my description of encroaching was too generous, but these are cases of 12.3.3 switching lanes in the intersection if I did not disengage when it starts to encroach. Here's a couple examples of approaching the intersection that I made sure was safe to allow it to continue straight instead of the correct behavior of shifting left:

View attachment 1037158

Even the lead vehicles weren't enough for end-to-end to realize it should stay left. Would you agree that this 12.x behavior is wrong and worthy of calling it a regression if 10.2 fixed this behavior long ago? I suppose to be fair, humans do sometimes just cut across too, so is it just preference for smoother ride?

Shouldn’t be cutting but it’s also poor road design imo. I get shoved into the parking lane by bad drivers in places like that too. Maybe FSD was trained on those drivers?

 

[hobertg]

OK so in the past I have complained about navigation. Today something happened... I have a certain route I go frequently for an appointment and today it changed that route automatically...so STUPID me, I decide to go the normal route FSD takes instead of the NEW route since occasionally the routes are not optimal. 80% to the location the lights were out (4-way stop) so traffic was backed up and if I had went the NEW route I would have missed the congestion.

I didn't think anything of it when I began the trip but I should I have looked closer to see why it chose a new route - Maybe it could alert you if you change the route based on its awareness of congestion for that route?

 

[joelkfla]

I had an incident when leaving my driveway with the navigation to turn left. Since there are a lot of cars parked and hard to see I’ve seen it alter its course and just turn right and go around another block instead. This is what a human driver would do.

Does that all the time, since my a***ole neighbor has been parking his giant pickup truck with the trailer right up to the right edge of my driveway. Navigation shows a right turn out of my driveway, but Tesla always takes a left. It's roughly the same distance out either way, so it's no big deal to me.

 

[AlanSubie4Life]

Pro tip:

Problem: As we know for most traffic lights, it charges the red light with stopped traffic, then does nutso stuff. No easing off.

Solution: I have found that if you wait until the car first starts to slow down (way before it engages friction brakes), even if it seems pretty late, then immediately disengage with the stalk, and let full regen take over, no pedal input, most of the time you will have a very pleasing stop indeed. It doesn’t use friction brakes and you can usually come to a stop with no pedal input. This is for flat surfaces only, and of course does not apply for very late stops, just the typical last-second nonsense where it doesn’t ease off for red lights.

It’s really great; I feel like I have unlocked an FSD v12 cheat code. It’ll unlock your insurance scoring too (not counting Tesla insurance which is masked presumably).

There will likely always be a few scenarios FSD can't gracefully solve. Unfortunately when they hit a brick wall Tesla resorts to gaming the design by adding a crutch and/or over weighting the data. Hopefully we don't start seeing weird behaviors like 10.69's omnipotent turn-right to turn-left.

If you ignore the terrible creep performance, and pausing in lanes of traffic, left turns are pretty good!

In the end, after all the hoopla, v12 is good but not that great. It is the incremental improvement we knew we would likely get, not the quantum leap we hoped for and had been promised. I really tire of disengaging for the same corners where it cuts it close every time, taking exactly the same line, and sitting wistfully as it charges towards the latest red light.

The only thing having me withhold judgement on it being a dead end is the next iteration. If 12.4 shows leaps and bounds of progress and solves the most pressing issues after the first round of training maybe there will still be a glimmer of hope for a quality L2 assist. It’s not far off, really.

We'll see whether Tesla correctly prioritizes and addresses the issues - and whether the car is capable of it. We only have one or two more iterations before HW3 is retired - and then that is the end of the road for my car. Once ASS and Banish are delivered there is no reason for them to spend any more time on it - they have booked the revenue. So barely any improvement left for HW3 - it is feature complete and nearly frozen. Praying they can figure out how to make it stop.

Today it changed lanes because someone was turning right into neighborhood and slowing down. That is progress! It’s these simple things that are key - not following the right route - can disengage and take over for those, since those things are easier to anticipate.

 

[swedge]

re. RSD rollouts:

It appears that 2024.3.x (12.3.x) has stopped rolling out with around 6,000 installed plus pending on TeslaFi, while there are still around 2,000 still on 2023.44.x (11.4.9). (Only a few 2023.44.30+ on V12.x are still left)

Meanwhile, 2024.8.x versions have installed plus pending around 7,000 on TeslaFi. But there is no longer an FSD version shown for these, though someone reported an 11.4.9 FSD was available on at least one one these versions. Are all of the 2024.8's also FSD 11? Does this include fully paid, plus subscribers plus free trial folks?

Does this mean that the offer of free FSD included V11? At first it looked like the free FSD offer was tied to V12's apparent improvement, but maybe not???

And has V12.3 been stopped short, perhaps pending a new version which will finally be offered to those 2023.44.x folks

Then there is the question of who is in the tester group? It appeared that this group was held back from 2024 until 2024.3.x, and is still held back from 2024.8.X.

It looks like the fleet is becoming fragmented into a few groups:

• testers on 2024.3 (12.3), and maybe some more testers left behind on 2023,
• non-testers with 2024.3 (12.3)
• non-testers with 2024.8.x (11.4.9)

I am a tester on 2024.3.10 (FSD 12.3.3.3) and V12 feels more comfortable and generally better than V11, so it seems odd to give the V11 in the free trial. But then it also seems odd to start the free trial while FSD still could use some improvements like curb clipping, stopping short, slow decisions, etc.

Maybe 12.4 will come soon in a 2024.8 or higher package, and resolve the fragmentation and solidify the performance. One can hope, I suppose. But with 2024.8.9 rolling out quickly already without FSD 12, who knows.

 

[eli_]

V12 is mostly good but after several days of use my biggest pet peeve has become its indecision around turn lanes, especially dual turn lanes. Not every intersection causes this behavior but when it does it will oscillate back and forth between lanes sometimes 2-3-4 times, or just kind of straddle the line taking up both lanes before committing at the very end.

Also I'm convinced the FSD driving profiles have zero effect on the e2e stack, including the "minimal lane changes" option, and only affects the legacy highway driving.

 

[Todd Burch]

V12 is mostly good but after several days of use my biggest pet peeve has become its indecision around turn lanes, especially dual turn lanes. Not every intersection causes this behavior but when it does it will oscillate back and forth between lanes sometimes 2-3-4 times, or just kind of straddle the line taking up both lanes before committing at the very end.

Also I'm convinced the FSD driving profiles have zero effect on the e2e stack, including the "minimal lane changes" option, and only affects the legacy highway driving.

I agree with both of your observations, I see that too.

The former will get better with more training. I believe for the profiles to have an effect (I too can find no difference whereas I did see a difference on v11) they need some sort of C++ guardrail code on acceleration. But I also admit to not knowing how else they could accomplish this using e2e.

 

[Mardak]

V12 is mostly good but after several days of use my biggest pet peeve has become its indecision around turn lanes, especially dual turn lanes

Do people end up disengaging and leaving a voice drive-note? Hopefully Tesla has some generic data collection trigger to detect unnecessary indecision to reduce various forms of wobble. In some sense, either lane of dual turns could be equally good choices, so a prediction of say 45% inner lane vs 46% outer lane could easily flip the other way.

 

[Todd Burch]

Maybe my description of encroaching was too generous, but these are cases of 12.3.3 switching lanes in the intersection if I did not disengage when it starts to encroach. Here's a couple examples of approaching the intersection that I made sure was safe to allow it to continue straight instead of the correct behavior of shifting left:

View attachment 1037158

Even the lead vehicles weren't enough for end-to-end to realize it should stay left. Would you agree that this 12.x behavior is wrong and worthy of calling it a regression if 10.2 fixed this behavior long ago? I suppose to be fair, humans do sometimes just cut across too, so is it just preference for smoother ride?

Hard to say for sure. See if it follows the lane when there are other cars around.

If it goes from lane 1 to lane 2 only when there are no other cars around, I’d say it’s probably just doing it for comfort. If it does it with cars around that’s a problem and disengaging there would be best to let Tesla know there’s a problem.

Much of my driving life I believed changing lanes in an intersection was illegal. In many (most?) states it’s usually legal if there isn’t striping prohibiting it—though if there are cars around it can be unsafe or confusing to other drivers.

 

[JHCCAZ]

V12 is mostly good but after several days of use my biggest pet peeve has become its indecision around turn lanes, especially dual turn lanes. Not every intersection causes this behavior but when it does it will oscillate back and forth between lanes sometimes 2-3-4 times, or just kind of straddle the line taking up both lanes before committing at the very end.

I took a plane trip up to the Seattle area this past week to visit family. The only disengagement I had on the drive to the airport in Tucson was exactly due to the above, an indecisive wobble in selecting berween two right-turn lanes heading into one large intersection.

My son has an HW3 model 3 and HW4 model Y, and has the trial on both at the moment. We encountered the dual turn-lane indecisiveness a few times.

IMO his road and traffic situations are far more varied and challenging than what I typically encounter, but both cars seem to handle things mostly well, with no obvious difference between the older and newer hardware. Of course it was raining much of the time, but I can't say it caused any notable difficulty for FSD v12. As others have noted, Tesla needs to modify the poor weather "degraded" warning so that isn't just popping up constantly; that kind of thing simply encourages the user to stop looking at them, which is counterproductive.

BTW they took me to the Tesla showroom at the mall in Bellevue, where I got to see the Cybertruck for the first time!

Also I'm convinced the FSD driving profiles have zero effect on the e2e stack, including the "minimal lane changes" option, and only affects the legacy highway driving.

I think this is right; no obvious effect on city streets FSD (I also found this on v11in town). Scrolling the speed down a bit, using manual offset mode, is a more effective way to minimize lane-change attempts.

 

[arbitrage000]

re. RSD rollouts:

It appears that 2024.3.x (12.3.x) has stopped rolling out with around 6,000 installed plus pending on TeslaFi, while there are still around 2,000 still on 2023.44.x (11.4.9). (Only a few 2023.44.30+ on V12.x are still left)

Meanwhile, 2024.8.x versions have installed plus pending around 7,000 on TeslaFi. But there is no longer an FSD version shown for these, though someone reported an 11.4.9 FSD was available on at least one one these versions. Are all of the 2024.8's also FSD 11? Does this include fully paid, plus subscribers plus free trial folks?

Does this mean that the offer of free FSD included V11? At first it looked like the free FSD offer was tied to V12's apparent improvement, but maybe not???

And has V12.3 been stopped short, perhaps pending a new version which will finally be offered to those 2023.44.x folks

Then there is the question of who is in the tester group? It appeared that this group was held back from 2024 until 2024.3.x, and is still held back from 2024.8.X.

It looks like the fleet is becoming fragmented into a few groups:

• testers on 2024.3 (12.3), and maybe some more testers left behind on 2023,
• non-testers with 2024.3 (12.3)
• non-testers with 2024.8.x (11.4.9)

I am a tester on 2024.3.10 (FSD 12.3.3.3) and V12 feels more comfortable and generally better than V11, so it seems odd to give the V11 in the free trial. But then it also seems odd to start the free trial while FSD still could use some improvements like curb clipping, stopping short, slow decisions, etc.

Maybe 12.4 will come soon in a 2024.8 or higher package, and resolve the fragmentation and solidify the performance. One can hope, I suppose. But with 2024.8.9 rolling out quickly already without FSD 12, who knows.

I don't know of anyone who was offered the free trial when on 2024.8.X. I don't believe anyone with the trial is trialling V11.

 

[MP3Mike]

It looks like the fleet is becoming fragmented into a few groups:

Becoming fragmented? It has been fragmented for years, and I don't see that going away any time soon.

 

[kabin]

If you ignore the terrible creep performance, and pausing in lanes of traffic, left turns are pretty good!

In the end, after all the hoopla, v12 is good but not that great. It is the incremental improvement we knew we would likely get, not the quantum leap we hoped for and had been promised. I really tire of disengaging for the same corners where it cuts it close every time, taking exactly the same line, and sitting wistfully as it charges towards the latest red light.

The only thing having me withhold judgement on it being a dead end is the next iteration. If 12.4 shows leaps and bounds of progress and solves the most pressing issues after the first round of training maybe there will still be a glimmer of hope for a quality L2 assist. It’s not far off, really.

We'll see whether Tesla correctly prioritizes and addresses the issues - and whether the car is capable of it. We only have one or two more iterations before HW3 is retired - and then that is the end of the road for my car. Once ASS and Banish are delivered there is no reason for them to spend any more time on it - they have booked the revenue. So barely any improvement left for HW3 - it is feature complete and nearly frozen. Praying they can figure out how to make it stop.

Today it changed lanes because someone was turning right into neighborhood and slowing down. That is progress! It’s these simple things that are key - not following the right route - can disengage and take over for those, since those things are easier to anticipate.

Yep. It's easy to smoke out all v12 marketing BS. FSD regression wack-a-mo needs to be a thing of the past.

I fear HW3 will never solve the 1.5sec latency issue. It'll never have the chops to allow the driver to relax.

Regarding turns, I've started disengaging on right turns where ridiculous acceleration is most noticeable. Although a left turn, this guy's reaction to the turn sums it up for me.

 

[kabin]

If true, will self driving ever be cheap? There could be many more training computer updates needed in the future.

https://insideevs.com/news/715366/tesla-10-billion-self-driving/

 

[PianoAl]

charges the red light with stopped traffic, then does nutso stuff. No easing off.

Solution: I have found that if you wait until the car first starts to slow down (way before it engages friction brakes), even if it seems pretty late, then immediately disengage with the stalk,

I do the same, but disengage with a tap on the brake pedal.

 

[joelkfla]

Do people end up disengaging and leaving a voice drive-note? Hopefully Tesla has some generic data collection trigger to detect unnecessary indecision to reduce various forms of wobble. In some sense, either lane of dual turns could be equally good choices, so a prediction of say 45% inner lane vs 46% outer lane could easily flip the other way.

It needs to be based on what's planned after the turn. If there's a right turn within a quarter mile after the left turn, it should choose the rightmost left turn lane, even if there's a line of cars in it. If there are no turns closely following the left, then fine, choose either -- but MAKE UP YOUR MIND!