FSD v12.x (end to end AI)

[Mardak]

V12 will begin a maneuver (for me) if the pedal is pushed, regardless of safety. I've tested it on a left turn to see if it would, then hit the brake. I wanted to see if it would go closer to the stop line, but the car just started going

I tested out various similar scenarios this morning with 12.3.6, and they were all safe while still also doing what I would expect most of the time. Your situation of needing to hit the brakes was because there was actual cross traffic to avoid or that it wasn't completing a full stop? If it was actually clear, I wouldn't be surprised if FSD treated it as a rolling stop.

• tested right turn following navigation pushing accelerator to go past the stop line for better visibility, and FSD correctly slowed down picking a more appropriate stop limit to see past street parking
• tested right turn ignoring navigation left turn pushing accelerator to suggest a right-on-red that FSD completed the maneuver once cross traffic cleared and we still had a red
• tested right turn ignoring navigation continuing straight pushing accelerator and same behavior as previous with FSD realizing I wanted a right-on-red instead of running red light by going straight
• there was one instance of an attempted right-on-red suggestion that I felt crept out too much forwards, but manually engaging the right turn signal made it immediately complete the turn, so it's not always consistent

So already 12.x can support collaborative acceleration where both the driver and FSD can decide to accelerate, and 12.x adjusts its steering and other behaviors to complete the desired maneuver (and people have been doing this to suggest to 12.x to maintain a faster speed). Collaborative steering currently isn't supported because it would be a disengagement requiring manual steering and acceleration.

Given the experience of 12.x working fine with collaborative acceleration, it would seem like driver steering to initiate lane change or other corrections like avoiding potholes or giving more space might work fine as collaborative steering with human initiation and FSD completion.

 

[OldManCan]

Yes there is ONLY ONE direction the car can go. So a signal to say you are going the ONLY direction you can go is redundant.

I agree with the logic here but feel its cheaper in the long run to just signal anyways.

 

[FSDtester#1]

Give us 12.4

 

[KArnold]

There was enough mode confusion for the NHTSA and Tesla to make a change.

Wish we could hold them accountable. Were there accidents due to this? Probably. But so dangerous we need the government to protect us from ourselves? Sounds like a stretch.

 

[KArnold]

But with this being end-to-end NN based, that means that it is likely that in the training videos human drivers only signaled when they thought it was necessary, not when they are supposed to.

We can do the math on this one.

Option 1: NHTSA wants both people and computers to obey all laws all of the time. They can't (yet) force people to comply but they can with computers. It complicates Tesla's FSD at a tactical level creating more work for them.

Option 2: Drive as humans drive. Why? Tesla's potential delay, if NHTSA forces, is not a good enough reason to do this - it's irrelevant. Neither is the fact that they cannot find enough training examples from the 12 people in CA who actually do this. So "why" from a NHTSA or even government perspective? Maybe it depends on the political winds de jour.

Which option is likely to win out?

 

[JulienW]

Give us 12.4

Relax gottogofast confirmed last night we would have it by this weekend so it is as good as a done deal. Feel better now?

 

[sleepydoc]

We can do the math on this one.

Option 1: NHTSA wants both people and computers to obey all laws all of the time. They can't (yet) force people to comply but they can with computers. It complicates Tesla's FSD at a tactical level creating more work for them.

Option 2: Drive as humans drive. Why? Tesla's potential delay, if NHTSA forces, is not a good enough reason to do this - it's irrelevant. Neither is the fact that they cannot find enough training examples from the 12 people in CA who actually do this. So "why" from a NHTSA or even government perspective? Maybe it depends on the political winds de jour.

Which option is likely to win out?

humans can't drive for crap so we want FSD to be more human-like?

 

[lzolman]

Yes there is ONLY ONE direction the car can go. So a signal to say you are going the ONLY direction you can go is redundant.

So no one behind you is ever distracted enough to not realize your lane is ending and you have to move over? In any case, I'm old now and signaling is the easiest way to do a head check...

 

[DrGriz]

But does the arrow show up in the visualization of the road for the correct lane (s)? That was my main question.

Again, whether the car can read and recognize the arrow is separate from whether it acts on it. It may weigh any map data more heavily, but that doesn't mean it can't read the turn arrow marking.

An example of this is the speed limit signs (it has been tested before that it can read fake speed limit signs that do not exist on a map), but whether it acts on that sign is a completely different manner.

Well, I will be honest and let you know that I was so involved in wrenching the car into the correct lane that I didn't notice, and I let my subscription lapse because I left town for 3 weeks and wasn't driving.

Nevertheless, if even sometimes, it registers the lanes, or road arrows, and doesn't respond, then it's unreliable and therefore dangerous (as well as useless). Which is why I haven't ponied up even $99 to extend it.

As for speed limit signs, I can say that even in AP (and FSD was the same) we pass many signs without it adjusting. These examples were true as of yesterday. On one of the main north-south highways the speed drops from 55 to 45 through one town, then goes back up to 65. The UI drops the limit to 45 appropriately and aligned with the sign, then, we pass the sign for 65 and the UI stays at 45 for another 1/8 mile and then changes to 65 (no signage at that spot).

In another section of road 20 miles away, the speed limit goes from 50 (south edge of town) to 65. The UI stays at 45 for the next 10 miles until it comes to the next 45 mph section. I have to keep my foot on the accelerator the entire time to go the speed limit, unless I use TACC. It passes several 65 mph signs. It used to do these sections properly.

So I have to conclude that it's following a corrupt map in all of these situations, and if it "sees" anything, it doesn't matter because it doesn't do anything with the information. Anymore. Perhaps there are subversions out there that still read the road. I don't know, but I am not the only person who has mentioned these sorts of things, including friends of mine who are in rural California.

It doesn't matter if it used to read and respond to signs. It wouldn't be the first time Tesla broke something with a version change.

 

[uscbucsfan]

I tested out various similar scenarios this morning with 12.3.6, and they were all safe while still also doing what I would expect most of the time. Your situation of needing to hit the brakes was because there was actual cross traffic to avoid or that it wasn't completing a full stop? If it was actually clear, I wouldn't be surprised if FSD treated it as a rolling stop.

• tested right turn following navigation pushing accelerator to go past the stop line for better visibility, and FSD correctly slowed down picking a more appropriate stop limit to see past street parking
• tested right turn ignoring navigation left turn pushing accelerator to suggest a right-on-red that FSD completed the maneuver once cross traffic cleared and we still had a red
• tested right turn ignoring navigation continuing straight pushing accelerator and same behavior as previous with FSD realizing I wanted a right-on-red instead of running red light by going straight
• there was one instance of an attempted right-on-red suggestion that I felt crept out too much forwards, but manually engaging the right turn signal made it immediately complete the turn, so it's not always consistent

So already 12.x can support collaborative acceleration where both the driver and FSD can decide to accelerate, and 12.x adjusts its steering and other behaviors to complete the desired maneuver (and people have been doing this to suggest to 12.x to maintain a faster speed). Collaborative steering currently isn't supported because it would be a disengagement requiring manual steering and acceleration.

Given the experience of 12.x working fine with collaborative acceleration, it would seem like driver steering to initiate lane change or other corrections like avoiding potholes or giving more space might work fine as collaborative steering with human initiation and FSD completion.

I tapped the pedal and the car began to turn into traffic until I stopped. My goal was to test to see if I could get closer to the stop line.

"Collaborative steering" is not something I ever see Tesla implementing. They are taking away user inputs, not adding, plus outside of the CT, the wheel is locked when driving to steer. I don't see any reason or path that they will allow users to turn and not disengage (even in 12.4/5 there's still wheel nag as well) to "make a correction".

 

[Dewg]

humans can't drive for crap so we want FSD to be more human-like?

It's definitely an odd human idiosyncrasy. When we are a passenger, in the back seat of an Uber for example, we don't stress as much about how the driver is driving. But when we are behind the wheel, even if the car is driving, we stress more and get more annoyed if the car doesn't do what we'd do.

I've enjoyed FSDS much more since I stopped caring about things, as long as the car is driving safely, I don't mind if it drives at, or just below the speed limit. I don't mind if it stops at a stop sign/light and waits a few seconds before going. And so on. I find I'm more relaxed now.

 

[gottagofast]

Give us 12.4

 

[spev18]

Give us 12.4

Think there are two possibilities:

1. Major bugs and/or issues with 12.4, in which case we all better turn on camp mode and chill because it's going to be a minute. Almost a full retrain sounds cool but a lot could've gone sideways.

2. Purely wishful thinking on my part BUT maybe they are syncing 12.4 up with the latest vehicle software and will surprise us with a 2024.15.x release. In that case we'll all lose our collective in happiness because then 2024.3.25 and 2024.14.x will both have a clear upgrade path for FSD 12.4 at the same time instead of only getting the latest vehicle or FSD version. Based on that Elon tweet I mentioned earlier, it's possible he meant that after the rollout of 12.4 it won't be an issue because the vehicle software and FSD sync will have already happened.

Smart money is on #1, but being optimistic for #2 is more fun.

 

[zoomer0056]

I agree with the logic here but feel its cheaper in the long run to just signal anyways.

A signal not only telegraphs intent, it also attracts attention and makes the car more noticeable. That lesson learned from many years of riding a motorcycle. It's a little thing, but games are sometimes won by just inches.

 

[dhanson865]

I got pushed 2024.14.8 / 12.3.6

Looks like everyone is getting it

Could it be 12.4 will go wider than 12.3.x did?

 

[FSDtester#1]

I got pushed 2024.14.8 / 12.3.6

Looks like everyone is getting it

Could it be 12.4 will go wider than 12.3.x did?

View attachment 1052003

That's a bad sign, unless you aren't an OG tester

 

[MP3Mike]

I got pushed 2024.14.8 / 12.3.6

Looks like everyone is getting it

Your FSD trial ended a week or two ago, right? So, you are probably getting rolled into the main deployment branch.

Not everyone is getting it. (Still ~30% of the TeslaFi fleet on 2024.3.25.)

Could it be 12.4 will go wider than 12.3.x did?

Highly unlikely. (At least not at first.)

 

[Dewg]

Think there are two possibilities:

1. Major bugs and/or issues with 12.4, in which case we all better turn on camp mode and chill because it's going to be a minute. Almost a full retrain sounds cool but a lot could've gone sideways.

2. Purely wishful thinking on my part BUT maybe they are syncing 12.4 up with the latest vehicle software and will surprise us with a 2024.15.x release. In that case we'll all lose our collective in happiness because then 2024.3.25 and 2024.14.x will both have a clear upgrade path for FSD 12.4 at the same time instead of only getting the latest vehicle or FSD version. Based on that Elon tweet I mentioned earlier, it's possible he meant that after the rollout of 12.4 it won't be an issue because the vehicle software and FSD sync will have already happened.

Smart money is on #1, but being optimistic for #2 is more fun.

If rumors are true, 12.4 removes steering wheel torque requirements in many instances, and it may be that they've found issues with the cabin camera vision tracking, where it's causing false positives on attention monitoring.

 

[David Wang]

Think there are two possibilities:

1. Major bugs and/or issues with 12.4, in which case we all better turn on camp mode and chill because it's going to be a minute. Almost a full retrain sounds cool but a lot could've gone sideways.

2. Purely wishful thinking on my part BUT maybe they are syncing 12.4 up with the latest vehicle software and will surprise us with a 2024.15.x release. In that case we'll all lose our collective in happiness because then 2024.3.25 and 2024.14.x will both have a clear upgrade path for FSD 12.4 at the same time instead of only getting the latest vehicle or FSD version. Based on that Elon tweet I mentioned earlier, it's possible he meant that after the rollout of 12.4 it won't be an issue because the vehicle software and FSD sync will have already happened.

Smart money is on #1, but being optimistic for #2 is more fun.

maybe elon should take that massive pay package and give some of those stock options to the people that actually do the work. Then they will have more incentive to get it out faster.

 

[lzolman]

It's definitely an odd human idiosyncrasy. When we are a passenger, in the back seat of an Uber for example, we don't stress as much about how the driver is driving. But when we are behind the wheel, even if the car is driving, we stress more and get more annoyed if the car doesn't do what we'd do.

I've enjoyed FSDS much more since I stopped caring about things, as long as the car is driving safely, I don't mind if it drives at, or just below the speed limit. I don't mind if it stops at a stop sign/light and waits a few seconds before going. And so on. I find I'm more relaxed now.

If I weren't in New England with traffic behind me, I wouldn't mind about those things so much either. But the behavior at stop signs around here is beyond embarrassing, and it gets worse now as the flora thickens up.