FSD v12.x (end to end AI)

[Mark II]

I think it is weird that some people have problems with FSD curbing wheels. I don't see it ever getting that close. Is it only an issue on some models? (I have a Model Y.) Is it some kind of weird road design that causes a problem?

I have not had curb rash (I have disengaged a few times, so it might have), however my tires have picked up stray screws & nails near the curb (at least that is my theory) FSD 12 causing flat tires (right rear)

I have a MSLR 2022, which it wider at the rear. One person in the above thread suggested they have not compensated for the wider rear vs M3 & MY with the new NN FSD.

 

[rlsd]

I went to change my username but what is this! Maybe by that time, we will finally have v12.4.9.3.6.

Oh first drive and park on v12.3.6 part2 as part of 2024.14.9. drives the same, just as slowly as before (no surprise), Autopark was just as slow as it has been, and it parked so close to my side, I couldn't open my door and had to manually repark. It always used to ignore where the adjacent cars were and just park between the lines and it still does the same thing.

View attachment 1056591

You need to recalibrate the cameras to get your username changed.

 

[JHCCAZ]

Are you that young and handsome?

Once upon a time

In a Ford Galaxie far, far away

 

[LSDTester#1]

I went to change my username but what is this! Maybe by that time, we will finally have v12.4.9.3.6.

Oh first drive and park on v12.3.6 part2 as part of 2024.14.9. drives the same, just as slowly as before (no surprise), Autopark was just as slow as it has been, and it parked so close to my side, I couldn't open my door and had to manually repark. It always used to ignore where the adjacent cars were and just park between the lines and it still does the same thing.

View attachment 1056591

To what?

 

[AlanSubie4Life]

One person in the above thread suggested they have not compensated for the wider rear vs M3 & MY with the new NN FSD.

Does not make a lot of sense since the car knows where it is relative to road objects and this would be important in general.

Also: trajectories are wrong on all vehicles. “Wrong” is my characterization. Tight trajectory is fine as long as it does not hit the curb.

Also: have examples of Model 3 (I think) curb hopping in parking lot causing thousands of dollars (if repair was desired; it appeared to scrape the pack) of damage.

 

[Teetoogrn]

I have not had curb rash (I have disengaged a few times, so it might have), however my tires have picked up stray screws & nails near the curb (at least that is my theory) FSD 12 causing flat tires (right rear)

I have a MSLR 2022, which it wider at the rear. One person in the above thread suggested they have not compensated for the wider rear vs M3 & MY with the new NN FSD.

I watch the camera view during right turns. It does get closer to the curbs than I would if driving myself, but it has never hit the curb. I think FSD stays more to the right of the lane when driving straight as well.

 

[rlsd]

I watch the camera view during right turns. It does get closer to the curbs than I would if driving myself, but it has never hit the curb. I think FSD stays more to the right of the lane when driving straight as well.

Yes. A lot of times it's too close to the double white lines to protect the bike lane. I think it moves to the left when bikers are present. Regardless, it should always be closer to the left lane marker though.

 

[Gwgan]

That’s not a strike.

Right, not a strike, my bad. I did not have to stop to reengage.

 

[AlanSubie4Life]

Right, not a strike, my bad. I did not have to stop to reengage.

Sounds like not much has changed. Except I did have to stop to re-engage in dense fog. I didn’t bother of course. It was just dead after a few red hands.

 

[stopcrazypp]

Does not make a lot of sense since the car knows where it is relative to road objects and this would be important in general.

I don't think that this how it works, at least if Tesla is still using the previous perception engine (or the NN is based on it)

Basically they had a camera rectification function that created a "virtual camera" that is positioned exactly like that of the NN training, and then the world model is built around that.

FSD v12.x (end to end AI)

It still works however, been busted a few times with the red warning ⚠️ 😳. Why does it have to yell at me every time I get in the backseat to get some shut eye? Try some of you blowup dolls. You have so to many to try so one may work. 🤣

teslamotorsclub.com

That's likely why it needs camera calibration (to gather the parameters necessary for that rectification function).

If that is the case, depending on how they model the vehicle(s) in the world model, it seems certainly possible it didn't account for differences like that.

That said, Tesla really shouldn't be leaving so little margin in the first place. This is one case where "lowest common denominator" may be justified (basically go wide enough that no variant will ever come close to curbing the wheels).

Also: trajectories are wrong on all vehicles. “Wrong” is my characterization. Tight trajectory is fine as long as it does not hit the curb.

Also: have examples of Model 3 (I think) curb hopping in parking lot causing thousands of dollars (if repair was desired; it appeared to scrape the pack) of damage.

 

[AlanSubie4Life]

If that is the case, depending on how they model the vehicle(s) in the world model, it seems certainly possible it didn't account for differences like that.

I’m not sure how construction of the world model is relevant here.

Why would they not account for this rather obvious difference?

I’m just saying it is important for them to know the size of the vehicle. Not sure how else they could make things work. The differences are not insignificant.

I’m also confused that you think the car does not know where it is relative to road objects. (What you quoted… ???)

Of course with Tesla any oversight is possible. Just seems unlikely.

Again, it’s not like we don’t have examples of Model 3/Y curbing ($13k+ in damage).

If you cut corners it is going to be more consequential in a car with a (slightly) longer wheelbase. Drive correctly and it’s not a problem.

That said, Tesla really shouldn't be leaving so little margin in the first place.

The only place they leave excess margin is stop lines, it seems.

 

[Dave EV]

That's likely why it needs camera calibration (to gather the parameters necessary for that rectification function).

I would think that ideally, the car would periodically be monitoring and making adjustments to the camera calibration. Having to reset the calibration seems like a bug.

 

[stopcrazypp]

I’m not sure how construction of the world model is relevant here.

It matters because if instead the NNs were based on relative position of the cameras to the real world and/or the views of the car body (fender cams can see the car body) then it's possible the different camera positioning can already naturally compensate for the differences with width and wheelbase of the vehicle, without Tesla needing to have explicitly separate car models as parameters.

Why would they not account for this rather obvious difference?

I’m just saying it is important for them to know the size of the vehicle. Not sure how else they could make things work. The differences are not insignificant.

Of course with Tesla any oversight is possible. Just seems unlikely.

It could simply be a "lowest common denominator" function that they didn't update for subsequent car models.

Again, it’s not like we don’t have examples of Model 3/Y curbing.

If you cut corners it is going to be more consequential in a car with a (slightly) longer wheelbase. Drive correctly and it’s not a problem.

Depends on the relative circumstances (front or rear, inside corner or outside/straight on) and prevalence as a proportion of the population. It would be helpful if people that end up curbing mention the model, and some details and then a tally can be done.

 

[AlanSubie4Life]

Depends on the relative circumstances (front or rear, inside corner or outside/straight on) and prevalence as a proportion of the population. I would be helpful if people that end up curbing mention the model, and some details and then a tally can be done.

I’m not sure what you’re arguing. Clearly FSD will hit curbs with any car model (is the argument that we have not seen a report from a Model Y?). Seems like a problem.

Is it possible it hits curbs more often with some car models? Sure. Not sure why that is relevant - it’s not like they “update the car model” and everything is going to be fixed.

That’s not the problem. I would be shocked if a Model Y had never hit a curb. It’s bigger than Model 3.

It could simply be a "lowest common denominator" function

Wouldn’t the lowest common denominator be the largest car? That seems like what would make sense. Anyway it does not really matter, since that is not the main issue.

What will fix this is driving differently. Sounds like they have it figured out since they don’t have interventions anymore according to Elon. So happy to hear it. Just a matter of time now. Elon

 

[stopcrazypp]

I’m not sure what you’re arguing. Clearly FSD will hit curbs with any car model (is the argument that we have not seen a report from a Model Y?). Seems like a problem.

I'm saying some curb hits may not be related to car size or trajectory estimation/prediction errors related to car size.

For example, outside or straight on curb hits are likely other errors (going too fast, misjudging the road path, etc).

Also the relative prevalence can point to car size/model being a possible factor (as opposed to playing no role). For example if the data shows a much larger percentage of Model Ss are doing curb hits, while a very small percentage of Model 3/Ys are doing so, then that might suggest car model playing a factor. If the percentages work out the same, then that suggests it's unlikely.

Saying just because there exists curb hits for Model 3 and thus model plays no factor doesn't make sense to me.

Is it possible it hits curbs more often with some car models? Sure. Not sure why that is relevant - it’s not like they “update the car model” and everything is going to be fixed.

Not all cases will be fixed, but it may fix many of them if that was an issue. That matters!

That’s not the problem. I would be shocked if a Model Y had never hit a curb. It’s bigger than Model 3.

Again as above, I don't think enough data has been presented to show if car model is a problem or not.

Wouldn’t the lowest common denominator be the largest car? That seems like what would make sense. Anyway it does not really matter, since that is not the main issue.

The largest car dimension is not a constant. It has changed over the years. An example of this problem of using a lowest common denominator (since I happened to encounter it) is Microsoft's 32GB limit for formating FAT32 partitions in their tools. It was good enough for decades, until drives got bigger, but they never bothered to update their formatting tool (third party tools did).

What will fix this is driving differently. Sounds like they have it figured out since they don’t have interventions anymore according to Elon. So happy to hear it. Just a matter of time now.

 

[AlanSubie4Life]

Also the relative prevalence can point to car size/model being a possible factor (as opposed to playing no role). For example if the data shows a much larger percentage of Model Ss are doing curb hits, while a very small percentage of Model 3/Ys are doing so, then that might suggest car model playing a factor. If the percentages work out the same, then that suggests it's unlikely.

Sure. But we’ll never get that data. What we know is:

1) Every vehicle model seems susceptible to curb strikes to some degree.
2) The trajectory FSD v12.3.x takes is notably different than 11.4.9, in a way that would be expected to lead to curb strikes. This is a known fact.

To me it seems like that points to a clear issue. And it’s not “need to adjust for car type.” That was all I was saying.

Saying just because there exists curb hits for Model 3 and thus model plays no factor doesn't make sense to me.

Yeah that doesn’t make sense to me either.

Not all cases will be fixed, but it may fix many of them if that was an issue. That matters!

Of course reduction in rate matters but that’s kind of a patch (if it even works which I doubt since they likely already try to account for this), not addressing root cause. As I have said several times they should take vehicle dimensions into account. It’s important for everything, not just curb strikes.

In the end they should drive correctly for the given vehicle. It won’t do to have Cybertruck rolling up on every curb it sees, lacerating pedestrians.

 

[SSDTester#1]

You need to recalibrate the cameras to get your username changed.

First, two wheel restart the forum though

 

[dachshund]

I have not had curb rash (I have disengaged a few times, so it might have), however my tires have picked up stray screws & nails near the curb (at least that is my theory) FSD 12 causing flat tires (right rear)

I have a MSLR 2022, which it wider at the rear. One person in the above thread suggested they have not compensated for the wider rear vs M3 & MY with the new NN FSD.

Mine veers to the outside of the curve, typically steering too far left. It’s surprisingly awful FSD is in my ModelY. I have to keep a hand on the wheel, and it gives up when I prevent it from hitting the curb. And then it asks me why I disengaged. I always think, kinda busy keeping you from crashing. Can’t talk now.

 

[bdwalters]

by now, my butt provides me better info

The regen indicator is kind of cool information, but I don’t see that is has much practical use.

 

[Ben W]

I have not had curb rash (I have disengaged a few times, so it might have), however my tires have picked up stray screws & nails near the curb (at least that is my theory) FSD 12 causing flat tires (right rear)

I have a MSLR 2022, which it wider at the rear. One person in the above thread suggested they have not compensated for the wider rear vs M3 & MY with the new NN FSD.

I got my first 12.x curb hit tonite. Older Model 3 on freshly updated 2024.14.9 + v12.3.6. The car was turning slowly right onto a narrow unlit residential street, and I could tell it was too close to the curb, but decided (unwisely) to let the car do its thing. Big bump against the curb, immediate Red Hands Of Death.

Given that this is the narrowest car Tesla makes, and that the hit was much more severe than just a bare graze, I'd call this a pretty significant fail. An argument for retraining the NN to stay farther away from curbs in general, but also an argument (in my opinion) for Lidar, which would have clearly detected and located the curb in dim light. Elon talked about FSD avoiding static objects, but you can't avoid what you can't see!