FSD v12.x (end to end AI)

[JulienW]

Well, I'll accept they're at least working on the issues. If everything we've talked about that's dangerous is fixed as soon as May (I'll settle for June even), that would make me ecstatic.

Yeah, as I have conjectured before that I bet having to "rush" 12.3.x into 24.3.x in a hurry to meet Elons proclamation is slowing 12.4 a little at least. And still needs to tackle the 30% or so cars on 24.8.x that are "Elon" promised a free trial. If it is truly NNs and "ChatGTP" type learning we should see some big improvements and not many steeps back like the old "follow the rules" code. I bet we see 11.4 rolling by early to mid May with some solid advancements, especially in speed control and hopeful some edge case-ish (like school zones and RR tracks).

 

[sleepydoc]

1) Doesn't even attempt go the speed desired by the user.
2) Won't respond in a normal way to red traffic lights.
3) Frequently can't handle stop signs in a remotely normal way.
4) Cuts it super close to curbs on many turns.
5) Simply stops in lanes of traffic in the middle of a maneuver.

1. Clearly an issue that they need to fix, but I’ve noticed something interesting with 12.3.3 - it accelerates quite well up to about 5-7 MPH under the target speed then sits there totally content. If I don’t focus on the speed and there’s no one behind me it’s not at all unpleasant to let FSD do its thing. Not what I would do but not dangerous in any way and not illegal.


2. how is it supposed to respond? My car stops on red and goes on green.

3. Mine doesn’t stop ‘normally’ but that’s because it does a full, legal stop that no one does. After that it generally proceeds fine and does a much better job than v11 did. For unprotected stops where cross traffic doesn’t stop it is definitely slow to proceed.

4. I’ve never experienced this - all of my turns have been in a manner that gives me no concern that it would clip the curb. I haven’t noticed - is it a continuing problem on 12.3.3 or was it only the earlier v12 versions?

5. Likewise, I’ve never had it stop in the middle of a maneuver.

 

[Therealm80]

My mirror fold saved location is not working anymore. Every time I pull into my driveway they fold, but as I start to pull into my garage, they unfold at which I refold them and click save location and it still does the same thing every time any ideas?

 

[sleepydoc]

FSD doesn't have OCR the sign to read what it says. FSD just has to know what the image means and what the car should or shouldn't do. Tesla used AI so FSD learned what Stop Signs and Traffic Light's look like and learned that some traffic lights are vertical and others horizontal. So maybe FSD/AI can learn what other signs mean too. I just wouldn't assume the problem is solved by improving map data.

I would expect if FSD watched thousands of videos with these types of signs AI would learn what to do from what the human drivers did. Could be wrong but AI has already done some amazing things for FSD.

View attachment 1035464

Here’s the problem - implicitly teaching a negative (I.e. don’t do something when you see this sign) is a lot harder than teaching a positive.

 

[JulienW]

Here’s the problem - teaching a negative (I.e. don’t do something when you see this sign) is a lot harder than teaching a positive.

Another problem is sign location and styles. Stop signs and Speed Limit must meet standards. However "don't turn left" can be in many different styles and in any location.

 

[ArchHamster]

My mirror fold saved location is not working anymore. Every time I pull into my driveway they fold, but as I start to pull into my garage, they unfold at which I refold them and click save location and it still does the same thing every time any ideas?

Had this happen after an update forever ago. I just cleared the mirror fold location and then set it again in my garage. Has been solid since. I think there is some uncommon bug which clears or alters driver save data afterban update. Every couple of updates I have some user setting revert. Usually it's benign however. Like changing the location of the blinker cameras.

 

[sleepydoc]

It took 6 risks in a 1 mile drive that I may not have been able to respond to fast enough to avoid a crash (I could not always see that it was clear before the car aggressively crossed a lane with ROW). Luckily, no one was coming in most of the situations, and the one with a car coming I was able to intervene and avoid a crash.

It changed lanes with no signal, took a UPL into my street with no signal, signaled left at a T intersection, then abruptly turned right. Any one of these are disqualifying.

I have no way to square any of this behavior with anything close to a safe product.

I must be holding it wrong.

Must be.

To expand on my previous comment - I drove to the airport and picked up my daughter last week. 0 interventions beyond tapping the accelerator to get the speed past the last 5 MPH. I took over around the loading zone for obvious reasons. After picking her up I set the GPS and it drove the entire way home, again with 0 interventions. This was not an Omar-esque ‘let’s see if it can make it no matter how badly or dangerously it drives,’ it was a completely natural drive and my daughter didn’t notice that I wasn’t driving until I pointed it out to her.

 

[lzolman]

sleepy, that's the holy grail I think we're all pining for, to have it "just work"...

 

[sleepydoc]

I've wondered this myself (specifically, if manual driving is more beneficial vs. FSD driving w/disengagement data.) I kinda wish Tesla would offer some sort of "training" mode where we could submit manual drives for training. Allowing us to toggle this on/off per drive would let us adjust our driving behavior so that we could attempt to drive "perfectly" for that trip when we're in the mood to do so.

Of course, "perfect" is subjective as can be seen in these threads, but it might still be helpful...maybe? None of us really knows what on earth Tesla does with our data, or what they're looking for at any given time...

It would make sense for them to collect disengagement data then collect real driver data at the disengagement locations to further train the system.

 

[Hiline]

The biggest complaint I have about FSD v12 so far is the stopping distance. My car sometimes stays an entire car length or more behind the next car at an intersection or half a car length behind the line at a stop sign. My second biggest complaint is the hesitation in making a requested lane change. The third complaint is the inability to move over when an emergency vehicle is approaching.

 

[sleepydoc]

Another point to consider for those asking “how in the world could they let this version out for wide release???”

From what I’ve seen in the comments, the variability in experiences between users is much greater with this version than it has been with previous versions. Even if we leave @AlanSubie4Life out of it I don’t think we can chalk it up to personal preferences. I don’t know if it’s related to location, hardware configuration or something else but it’s definitely real.

Whatever the reason, if the testers (and Elon) have my experience then it would make perfect sense to do a wide release. If they have others’ experience it wouldn’t.

 

[Electroman]

The way to use FSD Supervised now is to treat it like a drive function. Take all the disengagements just as another function. So, put in your Navigation route. At the turn that deviate from the route you want, just disengage, drive manually through the turn, reengage after the turn. The system would then reroute, again disengage on the next turn you want to make, do the turn manually. Repeat if the re- route is not what you want.

This way you are using FSD Supervised on portions which you feel comfortable with.

This is what I do even with Autopilot. I drive where I am not comfortable FSD driving. And I let it drive in other places. 80-20 rule. And I am quite happy with that.

 

[phxazcraig]

Does anybody think the quality of FSD varies by location? There are a ton of Teslas in Austin, for example, so a bunch of the training data invariably came from here. California in general would be a similar story. Does it work better in Tesla heavy areas than less Tesla heavy ones? I’ve been shocked at how well it works. I wonder if someone in west Texas would have the same experience.

Definitely. An example I've seen is that my car will neither get into nor out of HOV lanes around Phoenix - most of the time. But on trips to California, HOV lane transition works fine. The rules are different state-by-state, but it seems crossing wide lane lines are allowed in AZ but not CA. The car (11.4.9 FSD on highway) still doesn't think it can get across that line. (I manually enter the HOV lane.) Nav will start blinking a mile ahead of an exit when I'm in the HOV lane, but the car never moves out of the lane. Eventually it drives by all the exits until the HOV lane itself ends, miles down the road.

 

[zazim]

V12 is better by almost every measure. But unfortunately, it still cannot handle construction with shifting lane markers. The GW Parkway in DC is down to a single lane each way for the next 18 months. At times FSD will go at 45 mph and then slow down to 5-10 mph fairly abruptly when the barriers surrounding the single lane shift and then speed back up again when its preferred barriers come back. Better than v11 but only marginally so.

Lane drift in unmarked intersections and hurtling over speed humps is no longer an issue.

 

[JB47394]

Here’s the problem - implicitly teaching a negative (I.e. don’t do something when you see this sign) is a lot harder than teaching a positive.

It seems to handle No Right On Red okay. This is only one example, of course.

 

[phxazcraig]

Small win for FSD 12.3.3 here over all previous versions. It's better with gated community entrances and exits. I live with a gate consisting of In and an Out black wrought iron gates separated by a 4-foot-wide brick wall. Previous 12.3 or 12.3.2.1 versions would center a bit in the road and point at the brick wall while the gates opened. Then it would go through, slowly. With 12.3.3 it tends to point right at the gate in the correct lane, wait until it opens ALL the way, then slowly go.

It's still too slow though, as sometimes when I exit, by the time the car starts moving through the gate, it starts to close on me and I tap the accelerator.

 

[JulienW]

It seems to handle No Right On Red okay. This is only one example, of course.

It doesn't see No Right on Red signs. It is Map data that determines this. Most No Right on Red have been updated for a year or so for me in ATL.

 

[Hiline]

Small win for FSD 12.3.3 here over all previous versions. It's better with gated community entrances and exits. I live with a gate consisting of In and an Out black wrought iron gates separated by a 4-foot-wide brick wall. Previous 12.3 or 12.3.2.1 versions would center a bit in the road and point at the brick wall while the gates opened. Then it would go through, slowly. With 12.3.3 it tends to point right at the gate in the correct lane, wait until it opens ALL the way, then slowly go.

It's still too slow though, as sometimes when I exit, by the time the car starts moving through the gate, it starts to close on me and I tap the accelerator.

Piggybacking on this, my car drove itself out of the gated parking lot (once the gate opened) onto the street today. I didn't even know FSD could ever do this. Maybe one day it'll be able to drive from a parking spot to a closed gate, wait for the gate to open, then drive away.

 

[KLam]

This is what I do even with Autopilot. I drive where I am not comfortable FSD driving. And I let it drive in other places. 80-20 rule. And I am quite happy with that.

Another point I want to add is that I wish Tesla implement FSD Supervised more towards ADAS centric.

I was watching Chuck Cooks video this morning on protected left turn testing. I question the attention the driver need to monitor the car and surroundings during the turn. That is no way to make the user more "relaxed". For me personally, I would just drive to the stop line, one look at the left traffic, if it's very busy, I just relax and do my thing and come back to check the conditions after I think the traffic had passed. But not with FSD, I need to monitor conditions all the time, worry about FSD not making the right move, etc. That's not relaxed driving at all! Not ADAS at all. A true ADAS in this situation would tell me to relax and FSD is not going to do anything until it signals to me that its actively thinking of going ahead again, and I need to monitor its execution.

 

[355rockit]

Definitely. An example I've seen is that my car will neither get into nor out of HOV lanes around Phoenix - most of the time. But on trips to California, HOV lane transition works fine. The rules are different state-by-state, but it seems crossing wide lane lines are allowed in AZ but not CA. The car (11.4.9 FSD on highway) still doesn't think it can get across that line. (I manually enter the HOV lane.) Nav will start blinking a mile ahead of an exit when I'm in the HOV lane, but the car never moves out of the lane. Eventually it drives by all the exits until the HOV lane itself ends, miles down the road.

Is there a way to prevent the car from going into the HOV lane? On FSD, my Model 3 quickly got into the fast lane and then abruptly tried to change to carpool lane. I don't have HOV stickers and I was the only one in the vehicle. I had to disengage quickly and get back into fast lane.