FSD v12.x (end to end AI)

[AlanSubie4Life]

That’s what humans do half the time - light turns yellow, they start to brake then decide they can make the light and proceed.

Incorrect when it is not the controlling light.

 

[sleepydoc]

Elon said it was a lie.

Probably by one of those bots he promised to get off of twitter

 

[primedive]

Recent drives -- car does better in the snow than I expected, though I had to disengage to get it away from the snow banks a few times.

Lately, I've had it disengage a lot because I refuse to let it drive over potholes or sinkholes, and it continually has poor lane selection. There's also some sort of annoying behavior where for up to 10 seconds after a disengagement, the car will spazz out about crossing any lines in the road or getting close to the edge of the roadway, even though I might still be a good 3-4 feet away from it.

While waiting for a left turn, the car suddenly jerked the wheel 90 degrees to the left, hard enough to take me by surprise and I wouldn't have been able to stop it without having a death grip on the wheel. Kind of scary, and would have been exceptionally dangerous had we been moving quickly.

FSD is a cool tech demo / party trick, but for me personally, I find it more stressful than manual driving with the poor lane selection, speed behavior and refusal to move around potholes / sinkholes most of the time. I also don't find the highway FSD to be better than auto steer; if anything, it's worse. It exhibits poor lane selection and exit/entrance handling, so I'd rather just deal with that myself. Auto-steer and TACC just need to work more like other cars and: 1) Not arbitrarily change the speed on you. Leave that crap for FSD or navigate on auto-pilot, or at the most, slowing down for curves. There should be a way to not have it do that, 2) Allow steering the car without disengaging auto-steer (so we can move around potholes etc... or adjust the position), 3) Keep it engaged through manual lane changes.

When FSD is good enough to be a robotaxi, like Waymo, I'll be ready for it. Until then, not really interested in being the guinea pig.

 

[sleepydoc]

Incorrect when it is not the controlling light.

Thought you were talking about a stoplight turning yellow but people do that with the other lights, too. It turns yellow any the first instinctive response is that all the lights are turning yellow.

 

[AlanSubie4Life]

I think the term ‘NHTSA stop’ comes from Chucks’s videos. I was under the impression that it was NHTSA that forced Tesla to do away with the human-like rolling stop. Is this not true?

NHTSA requires car to stop at the stop. Not 5 feet to 10 feet before.

At sufficient distance from the intersection, the stop doesn’t count. 5 feet is too far. 10 feet (it happens!) certainly is. It is then illegal to proceed without stopping at the line. The line is placed so other traffic and road users can SEE you. This is very important.

So actually it is breaking the law.

With current behavior in busy areas (say La Jolla Cove area) it confuses people when it stops that far back because they don’t think you have made your stop yet. But it thinks so and will proceed (boldly in some cases!), leading to obvious problems.

This is one of the things that is broken with stopping. I think they can fix it.

I would like it to not tap the brakes, then proceed, when the left-hand-turn light goes yellow and my light stays green

Thought you were talking about a stoplight turning yellow but people do that with the other lights, too. It turns yellow any the first instinctive response is that all the lights are turning yellow.

Has not been my experience at all on massive thoroughfares. People know how it works. Would make no sense for lights to work that way most of the time.

 

[sleepydoc]

Why don’t these limitations (bugs) apply to simpler turns? Or are you suggesting only right turns, if left is unprotected?

Uhh.. because they’re simpler?

 

[AlanSubie4Life]

Uhh.. because they’re simpler?

What I was asking is why he thinks the car will not do the same (stall in traffic lanes) on simpler turns.

I for one know that if I had never experienced a stall on a simple turn, I would assume with very very high certainty that the car could stall on such turns. It just makes sense to me.

And then it follows to ask whether we should just do right turns at all such turns to avoid that happening.


It’s interesting when the car stalls. Everyone should experience it.

The bet was for FSD (beta) but this was FSD (supervised). I think the safety drivers were testing FSD (unsupervised) (beta) which probably has a couple nines of performance by now. FSD (supervised) is a driver assist so it's going to require driver interaction. Maybe they're trying out ways to make an EU legal version? Chuck was merely confirming actions that FSD (supervised) was already taking.

This seems fair; I did say this was like taking candy from a baby, so I am not too broken up about you reneging on the deal. Beer makes me fat anyway.

Are we going to take positions on FSD (Supervised) 12.4.x? I think we should look at all the testing done on a given main release (so 12.4, 12.4.x, etc.) since it seems like they reserve retraining for those main 12.x increments, and we cannot get to 10 tests on a given release. (There’s not compelling evidence of any 12.3 retraining within the 12.3.x series.) For that one unless they have hobbled it they should be able to incorporate the two-nines version. Also we’ll need more trials when we have that version.

I think it will fail to achieve one 9. (I also think I could finally be wrong, given the intense effort and the demonstrated incredible power of the neural net approach, which can solve any problem.)

 

[joelkfla]

Since yellow lights vary in timing, with no way to tell what that timing is, and the new NN are not coded but instead learn how to handle things by watching curated videos, I wonder if that's why we're seeing this behavior. Your yellows might have a 4 second timer, but the NNs were trained on 2 or 3 second timers, so it brakes out of caution?

For example, if your lights near you always have 4 second timers, and that's your muscle memory for years, then you visit another state where they have 2 second timers, I'd bet you run a few red lights before adapting.

In my case, it slammed on the brakes when I was probably less than a second from the stop line. It managed to stop beyond the stop line, but before the painted crosswalk. I'm glad that the driver behind me in the left turn lane was paying close attention.

 

[joelkfla]

There’s one place in my town at a turn that I would hit the curb I have to disengage at that 1 turn every time, all other curbs in town are fine

OK, bit if you're driving on an unfamiliar road, how would you know when you're at that 1 out 10 turns?

 

[AlanSubie4Life]

In my case, it slammed on the brakes when I was probably less than a second from the stop line. It managed to stop beyond the stop line, but before the painted crosswalk. I'm glad that the driver behind me in the left turn lane was paying close attention.

I’d recommend pushing through with the accelerator if it has made an obvious mistake and the slowdown will be severe (I’ve done this 3 or 4 times in the last two days). The car will gladly comply and allow you to maintain speed through the intersection. Definitely keep an eye on trailing traffic when approaching green lights, just as you would when driving manually. You need to be prepared to take the right action to avoid a collision; it is your responsibility as a driver to avoid having people behind you hit you.

 

[lzolman]

It's funny you mentioned this because that is EXACTLY what has happened the last few days.

2 things are happening:
I'm understanding the limitations of FSD and only using it in situations that I know it works well in.
And I'm getting used to it's quirks like sudden speed changes and hugging too close to the line.

I would be more impressed by it if I had the chance to experience the previous versions so I could be excited by it's progress, it's easy to be impressed if you watch something grow from nothing but if you stand back and look at it with fresh eyes it's not as impressive since you have nothing to compare it to.

I'm a Tesla fan boy, I have Tesla solar, batterywalls, two Teslas and $TSLA, I WANT FSD to work well.

It's funny how the most ecstatic reviews of V12 seem to be coming from folks like you who've never used FSD before. It boils down to what happens on each person's specific set of runs. If someone doesn't have a complete fail, the rest of the experience seems to be a lot better than it used to be when the driving logic was "robotic". Now that's more human-like (including our less-than-exemplary driving habits), absent failure it looks Real Good.

The problem is the failure rate. The march of nines needs a whole lot more of 'dem nines. I'd be impressed at this stage if they took ten random FSD newbies out for a demo, nine of them were happy with the experience.

We need that to be like 9999 out of 10000.

[EDIT: Or maybe only 9 out of 10 to be "happy", but 9999 out of 10000 (at least) to feel safe.]

 

[Mardak]

it wasn't clear what role the disengagement data would play, i.e. how these negative training examples could be used to teach what not to do

Similar to training 11.x perception that a vehicle is a bus and not a car, reinforcement learning for end-to-end can adjust control to favor a good example and avoid bad examples. For example, if 11.x had a prediction that a vehicle was 40% bus, 45% semi, 10% car, etc., the training data label could be 100% bus, 0% semi, etc. effectively shifting the neural network towards 40% bus -> 100% and 45% semi -> 0%, and once the bus % becomes the highest output probability (43% bus vs 42% semi, etc.), the visualization switches to displaying a bus.

Similarly, end-to-end has been trained on examples of what it can do such as going around parked vehicles or crossing double yellow lines, and these maneuvers might be fine most of the time, but disengagements are necessary when it incorrectly believes it's another situation that it should do this maneuver such as blind corner with oncoming traffic. Providing disengagement negative examples can then teach the network to reduce the probability of doing this overtaking action for this type of situation, and similar to how 11.x visualizations switch to the highest probability output, 12.x can switch to a different behavior it already knows about such as just staying in lane even without an explicit positive example of staying straight in this exact scenario.

Another aspect of having more people use the latest 12.x is that each iteration of the neural networks will have new improvements and potentially regressions, so having disengagements more actively finds these cases where end-to-end did something wrong. There's a lot of "good enough" behaviors that can differ from how a particular human would have done some action but many drivers would also find them acceptable, so active usage can be more useful than general shadow mode triggers.

 

[Todd Burch]

Auto speed much better for me today, not sure why. Same road, same traffic, same conditions, same car settings…went 48 instead of 42 in a 45mph zone today. Much more reasonable.

Drove me from home to a park 20 minutes away, decent traffic, no interventions. Last lane change was within 0.2 mi of the turn so my rectum got sore at that point from all the puckering but other than that, all good.

 

[John Falconer]

Does anyone's FSD figure out the little green-red traffic light at a highway on-ramp that allows one car through at a time? Mine just blows right through it.

It works fine here in NorCal for me. Early 2020 MYLR.

 

[gottagofast]

OK, bit if you're driving on an unfamiliar road, how would you know when you're at that 1 out 10 turns?

The first time I could tell it was going to hit because of how sharp it turned, I disengaged 1 second before it hit

 

[tmoz]

Traveling SE on Texas 287 between Amarillo and Childress, the car doesn't read most speed limit signs (or they aren't programmed in), but the speed limit is 75 most of the time. The car behaves differently than when it was on I 40. I'm guessing it is treating TX 287 as not a highway. Many disengagements. It must not be using the freeway code, I think.

 

[kpanda17]

First, lots of drives
Love SFSD v12.3.3
Would like fixed:
Some hesitancy when approaching a light, grad slowing is too long
Jack rabbit off lights and stop signs unnecessarily fast
Random blinker on the highest, guess the old stack showing its warts
Still edge cases, that are challenges for human drivers

Excited that 12.4 and beyond will be amazing

 

[AlanSubie4Life]

Traveling SE on Texas 287, the car doesn't read most speed limit signs (or they aren't programmed in), but the speed limit is 75 most of the time. The car behaves differently than when it was on I 40. I'm guessing it is treating TX 287 as not a highway. Many disengagements. It must not be using the freeway code, I think.

Easy enough to tell. (auto/max & tentacle)

 

[PianoAl]

Auto speed much better for me today, not sure why. Same road, same traffic, same conditions, same car settings…went 48 instead of 42 in a 45mph zone today. Much more reasonable.

I notice this variability a lot and have been trying to figure out what's different. My best guess is a lead car.

Is it possible there is a lead car even if it's far away?

 

[SidetrackedSue]

Reading how 'super slow' FSDS can be convinced me to try it in my area.

So far, because of the day of week and time of day, I only managed a 1.5 mile drive. My neighbourhood has only two ways in/out and at this time of day one is completely blocked by traffic leaving the city and the other has cars cutting through our streets to avoid the blocked road (and make an illegal turn to rejoin the traffic leaving the city.)

Keeping in mind I'm supposed to judge if I feel safe with how it gets to the destination and not if it drives like me, it did very well. Speed was really slow but that's what is needed. That 3/4 of a mile has three different speed limits, 30 kph on the streets without a curb, 40 kph school zone, 50 kph on the 'minor collector' streets that have sidewalks and bus routes.

I wasn't thinking and one left turn with an immediate right got us delayed by the traffic of people cutting through our neighbourhood. The car did really well with that, two thumbs up from me given the steady stream of cars from the right (the cutters), the less steady stream of cars from the left (locals using the only way out of the neighbourhood at this time of day.) There was also an illegally parked car forcing FSDS to be over the lane marker at the stop sign and an overgrown hedge (that the city has ordered cut back so people can see to make the turn) complicating that off-set intersection.

Outbound was 50/40/30 kph speed limits. On the way back there was no 50 sign when we turned so the car was still going 30. Then the school zone 40 sign was missed because it is hidden in a tree (not FSDS's fault at all.) So I disengaged until after the school, and reengaged when the 50 sign was in view and the car did fine. I goosed the car through most of the stop signs (my little drive had 4 in each direction) unless there was a reason to stop and contemplate life, the universe, and everything while keeping NHTSA happy. I knew that would be uncomfortable, I have been practicing them on my own for the past 2 months and just can't sit there for what constitutes a legal stop. I did let the car stop, I just didn't allow the 'reflection time' unless there were cars in view in front of me.

On the other hand, in the future if it is those 'cutting through' cars behind me, I might just let FSDS do it thing. While fully stopped, I can reflect on the joy I feel when FSDS pisses them off.