FSD v12.x (end to end AI)

[David Wang]

I've been pretty bullish on fsd v12 as about 90% of my drives have been disengagement free in my little suburb city in the SF Bay Area.

However, yesterday I took a trip to San Francisco itself and I was really looking forward to what FSD could do and I had a bunch of disengagements that it was pretty disappointing. I don't know how Whole Mars Catalog can post all these disengagement free drives in San Francisco. I'm guessing he only posts a small subset of his drives and any drives that have disengagements, he doesn't post them to youtube. Given all the disengagement free drives that whole mars catalog posts, I was looking forward to a good fsd performance in SF, but alas that wasn't the case for me.

I'm currently on v12.3.3. v12.3.4 hasn't been pushed to my car yet.

 

[Blue_Model_S]

Would anyone else like to see an improvement when choosing a destination for FSD to have an option to automatically zoom into the destination allowing you to place the pin where you want the drive to end/park? Sometimes I do this but often forget which then becomes a hassle to try and change when driving. Assumes of course Tesla does a better job of ending at the Pin or close to it like my driveway.

It will do that if you touch the navigation icon (with distance information) on the right of the destination address.

 

[aronth5]

It will do that if you touch the navigation icon (with distance information) on the right of the destination address.

I've been doing that all along but it's not all that helpful unless I'm doing it wrong. First, I have to remember to do that. Second, I don't want to have to touch the display to begin with to initiate the zoom. I want navigation to auto zoom to the destination so the only thing I have to do is move the pin. Make it an optional setting since not everyone will want to do this. Third, that still requires me to zoom into the destination using the finger pinch. Thats a pain. KISS principle. For example if I'm gong to Home Depot I know which entrance I'm going into so I want FSD to auto zoom to the parking lot so I can just move the pin.

Robotaxi will have to do this for a new mobile app so the requester can pinpoint where to be picked up and where to drop them off.

 

[lzolman]

Would anyone else like to see an improvement when choosing a destination for FSD to have an option to automatically zoom into the destination allowing you to place the pin where you want the drive to end/park? Sometimes I do this but often forget which then becomes a hassle to try and change when driving. Assumes of course Tesla does a better job of ending at the Pin or close to it like my driveway.

Never in 6 years of FSD has the idea of placing a pin at the destination ever even occurred to me. But now that you mention it, if it were that easy to select a particular target spot I might start using it.
Arriving home, the nav is just hopelessly lost (tries to take me into a nearby developemt) and I always just disengage. The map doesn't even show my 800 ft. driveway and I have no idea how to change that, so have just considered myself SOL.

 

[enemji]

Its a computer, it just feels like it should respond instantly, and delete any delay from training that's only due to human reaction times.

It would be possible only if all the vehicles were talking to each other, and not anticipating

 

[tmoz]

@JulienW I spent a bit more time with the speed limit blindspot weirdness and found something "interesting".
FSD is using the type of road to figure out which stack to use and is using V12 stack on 65-70 two lane roads, it only goes to V11 when its divided.
At least in this local Texas area, I can see it happen on roads that flip from two lane to divided and back. Its really rather annoying.
Can't test that until this weekend, but at least that will be with FSD12.3.4 unless they push another version

I'm 99% sure that's what it did to me on roads like Tx Hwy 287.

 

[Nakk]

Haha, I'm 65 and I still haven't set foot in a "Gentelmen's Club" (or even a strip club).

Seriously? Where the heck do you get breakfast????

 

[OxBrew]

So what's an accident that can be avoided by that eternally vigilant computer with 50 ms reaction times, but not 500 ms reaction times? (Assuming that FSD is that delayed)

One of those last minute lane changes where the car in front of you reveals a stopped line of cars, you have may e 2 seconds to avoid it? Burning one second just because humans do is a really bad strategy, if you ask me. Maybe you want FSD to just wait a second to react for the hell of it, be my guest. I'm not signing up for that.

Its typically a full second to react, not a half second, for a human.

 

[rlsd]

Haha, I'm 65 and I still haven't set foot in a "Gentelmen's Club" (or even a strip club).

You are not a member of the OG?

 

[SilverSp33d3r]

I have been experiencing some strange behavior with 2024.3.10 in parking lots, not committing to turns then after it’s too late it attempts to make a turn.

 

[Dewg]

Its typically a full second to react, not a half second, for a human.

I've always thought human reaction time is about 250ms

 

[aataskin]

I've been pretty bullish on fsd v12 as about 90% of my drives have been disengagement free in my little suburb city in the SF Bay Area.

However, yesterday I took a trip to San Francisco itself and I was really looking forward to what FSD could do and I had a bunch of disengagements that it was pretty disappointing. I don't know how Whole Mars Catalog can post all these disengagement free drives in San Francisco. I'm guessing he only posts a small subset of his drives and any drives that have disengagements, he doesn't post them to youtube. Given all the disengagement free drives that whole mars catalog posts, I was looking forward to a good fsd performance in SF, but alas that wasn't the case for me.

I'm currently on v12.3.3. v12.3.4 hasn't been pushed to my car yet.

It should all be fixed in the next four months. Or there will be no robotaxis in SF, only in little suburb towns like yours.

 

[David Wang]

The problem w fsd v12 is that it still doesn't pass the wife test. It still makes too many basic mistakes.

My wife doesn't trust it, she won't enable it when she drives herself and she gets nervous when I'm driving the car w fsd enabled and she's a passenger.

When it passes the wife test then it will be ready.

She's been on rides where there have been 0 disengagements, but inevitably there's that one ride when I have to disengage due to a safety critical mistake and she doesn't trust it anymore. It's ok for fsd v12 to make mistakes, but it doesn't do it in a safe manner and that's a big issue.

She was coming around while we were driving in our suburb city and v12 was driving pretty good, but then we took a trip to SF and I had around 4 or 5 disengagements and that really set her trust back significantly.

That said I'm still going to subscribe when the free trial is over. I can see the potential.

 

[Yelobird]

The problem w fsd v12 is that it still doesn't pass the wife test. It still makes too many basic mistakes.

My wife doesn't trust it, she won't enable it when she drives herself and she gets nervous when I'm driving the car w fsd enabled and she's a passenger.

When it passes the wife test then it will be ready.

That said I'm still going to subscribe when the free trial is over.

Is it possible it’s a “wife” issue? I Kid!

 

[lzolman]

The problem w fsd v12 is that it still doesn't pass the wife test. It still makes too many basic mistakes.

My wife doesn't trust it, she won't enable it when she drives herself and she gets nervous when I'm driving the car w fsd enabled and she's a passenger.

When it passes the wife test then it will be ready.

That said I'm still going to subscribe when the free trial is over.

We must have the same wife!

 

[JulienW]

[rant]Had to drive across the city and back this morning and BY far the #1 for me is LANE SELECTION. It is constantly in the wrong lane and the last second looks for sympathy to get over. I HATE this I ended up making WRONG or missing turns/disengaging 6 or 7x because of this. When you are in rush hour bumper to bumper you can't toddle along until the last 500' and expect someone to let you in.

 

[OxBrew]

I've always thought human reaction time is about 250ms

It varies with age, 250ms is the fastest a young male can percieve the need to brake. Then add 500-750ms to move their foot to brake.

Overall, different situations use an average of 3/4 to 1.5 seconds from event to action for total reaction time. This assumes the driver is able to see the event.

Training data probably has a wide variety of reaction times, if it includes events where people were just not looking where they needed to be (checking the mirror?).

Point is, how can they program out any delay in reaction time? End to end would seem to not be able to be curated to eliminate it, and there are easily identifiable not-so-edge cases where it matters a lot. The promise or expectation (it's a robot! it should be instant!) is that an always aware computer should be able to out perform ALL humans in response to events.

 

[JB47394]

One of those last minute lane changes where the car in front of you reveals a stopped line of cars, you have may e 2 seconds to avoid it?

Okay, good example. That's certainly the sort of thing where I'd want Tesla to train the system beyond human abilities.

But keep in mind that this is all predicated on Mardak's assumptions. Tesla is supposed to be training the system with "good driver" video. If they do that, then the car should do well. If they augment it with simulations that take advantage of the car's vigilance, then it should do that much better. Better than you or me.

As an aside, this is another reason why I think it would be good to train from a "control representation" of the driving environment. I originally thought that it would be good to normalize the outputs from sensors on any vehicle (so you don't have to retrain the entire system for each combination of sensors), but I realize that it would also allow simulated driving environments to be worked in without compromise. Just make the control representation something that a simulation can produce with equal fidelity, and you can start training perfectly on all manner of dangerous scenarios. Like this one. Heck, it could start in on self-directed learning.

I've always thought human reaction time is about 250ms

JD Power claims an average reaction time of 750ms, where that's the time it takes the average driver to move their foot from the accelerator to the brake pedal. Obviously, steering changes can happen faster. I could believe 250ms. But imagine the scenario where that stopped car appears in front of you. You've got to brake and check your mirrors, then swerve. I think it's safe to say that the delay before swerving is the reaction time. FSD could react pretty much instantly, but if it's trained to behave like people, it'll be slower.

 

[rlsd]

[rant]Had to drive across the city and back this morning and BY far the #1 for me is LANE SELECTION. It is constantly in the wrong lane and the last second looks for sympathy to get over. I HATE this I ended up making WRONG or missing turns/disengaging 6 or 7x because of this. When you are in rush hour bumper to bumper you can't toddle along until the last 500' and expect someone to let you in.

FSD always wants to get to destination using shortest time approach. I think it should use the easiest path if Tesla want to do robotaxi.
I started reducing speed limit below the speed on the adjacent lanes to prevent FSD from jumping out of the lane since yesterday. I will continue to do this to see if it has any affect.

 

[OxBrew]

s an aside, this is another reason why I think it would be good to train from a "control representation"

That makes a lot of sense, if it can be done. Almost like labeling, "this batch of human reactions to situations all fit in this set of outputs" but with curation?

So basically funneling human training data through filters that better control the outputs? So the inputs are still valid, varied real situations, but the decisions and outputs are sanitized, in a way?

I have no idea what I'm talking about, or how any of that would work. Apologies for the most inexpert speculation, my least favorite kind of content here. I just hope someone who does really understand the weedy mechanisms of NN could explain what I'm talking about.