Self driving cars are still decades away

Here’s the relevant sentence from the Illinois statute.

Except when directed to proceed by a police officer or traffic-control signal, every driver of a vehicle approaching a stop intersection indicated by a stop sign shall stop at a clearly marked stop line, but if none, before entering the crosswalk on the near side of the intersection, or if none, then at the point nearest the intersecting roadway where the driver has a view of approaching traffic on the intersecting roadway before entering the intersection.

Texas’ law is similar. If there’s a stop line, you’re supposed to stop there, or before entering the crosswalk.

Okay, thanks. Ignorance fought.

(Hoped-for loophole closed. :angry:)

This is true. But I’ve also been nearly hit when running because of this behavior, and I’ve even bounced off a couple cars who rolled into crosswalks before I could stop. Too many cars look for pedestrians in the crosswalk, and don’t think about ones who might be approaching the crosswalk at anything more than a slow walking pace.

Allowing a slow moving non-stop through the crosswalk to a point where you can see down the cross street would work if drivers could be relied on to actually look for all kinds of pedestrians. But drivers clearly cannot. Properly designed self-driving can do a much better job of this. But I don’t like the example it would set for all the human drivers. I’d prefer to start off with self-driving strictly following the rules, then loosen those up when self-driving vehicles becomes more ubiquitous.

V12.4 is being rolled out to a lot more users now. All of them on the 2023 branch of the regular software. Newer cars on the 2024 branch (like me) are in limbo.

FWIW, FSD 12 seems to be pretty good at this. It’ll wait for pedestrians that are somewhat far away but moving at a good clip. On the other hand, it’ll go for it if pedestrians are just stopped at the corner. Seems to have some kind of predictive model based on movement.

Just got the update to FSD 12.3. Of course, it happened after my 217 mile round trip drive into the mountains today.

I used 12.3 for a drive through the neighborhood to the store. I did notice the weird speed thing. If the set point was 26, it might get to 26 quickly, or it might get to 18, and then stay there. If I used the accelerator to get to the correct speed, then it would hold the set speed.

Unfortunately I didn’t get to test it with the one thing 11 really had trouble with. At the only traffic light on the trip there may be a flashing yellow left turn arrow, 11 did not handle well. This time the light was red, so I got a green arrow, which 11 always handled fine. 12 did it fine, too.

Nice! Looking forward to hearing more reports from your neck of the woods. The speed control is a definite weak point, though improved from 12.2.1 to 12.3.

Do you get the “phantom braking” issue with FSD? We drive a lot on 2-lane highways (not divided) and I’ve now had the third car freakout. One has happened when a semi approached and the shadow was fully in our lane and the other two have been when we approached a curve to the right and an approaching semi appeared to be in our lane.

How well does “self driving” work in those countries like mine, where we drive on the left side of the road?

I imagine the UK & Australia being the biggest markets, but there are others.

Just mirror-image the camera feeds; it’ll be fine :slight_smile: .

It’s probably going to be a while before it reaches left-hand drive countries, and I suspect it’ll be a while even before it reaches right-hand drive areas of Europe and elsewhere, simply because the signage, etc. is so different. But the good news is that if the new end-to-end training strategy works out, it should “just” be a matter of collecting training data from the areas in question and training with that.

Data collection and “cleaning” is probably the single hardest problem in AI right now, or at least the single most important one (it’s why GPT-4 is ahead of its competitors). So the problem shouldn’t be underestimated. But Tesla at least has a large fleet of vehicles and the data collection lessons they learn in North America will apply everywhere. So the next few geographical regions should be easier than the first.

It’s possible they could train a “universal” model–after all, human drivers can usually transition with some effort. But they do have memory and processor power constraints, and a universal model could be less efficient than a geographically-limited one. Maybe they’ll be able to train a universal base model and then fine-tune based on local rules and laws.

This is logical…different signs, different laws, different sides of the road, etc will take time to learn because it requires different logical solutions.

Now…what about the countries where logic just doesn’t apply? :slight_smile:

Will full self drive ever be available in places where chaos is the only rule ? Bankok, for example.

First real drive on FSD 12, and it was fine, but really pretty much just as fine as 11. Some things are better on 12, but it didn’t fix all of the problems, and it introduced new ones.

The big thing it fixed which I noticed is that it slows down for the places there are storm drains across the road (speed dips instead of bumps). Problem was, it slowed to 10mph (the right speed), about 100 feet from the dips, then accelerated to get closer to the dips, then slowed to 10 again, and finally accelerated once past them. Just one trial, so I don’t know if the slowing was unrelated phantom braking, or just confusion on the dips.

Lane changes on the freeway seem better, but the decisions still seem poor. I was about to pass a semi when 12 decided to get over to let someone by, but instead of passing the semi and then moving over, it slowed down to let the semi get ahead, and then move in behind it.

The reported set speed issue is a problem. Every time it crosses a speed limit boundary, it might not accelerate all of the way to the right speed. This even happens when the speed limit is the same. Speed limit is 30, car is going 30, drive past a sign that says “30”, and then car slows to 27.

Lane choice is still a problem, but at different intersections. There were certain intersections where 11 would usually try to go into the left or right turn lane, instead of the straight lane. 12 managed that intersection correctly, but then inappropriately went into the right turn lane at an intersection 11 never had a problem with.

Inappropriate freeway exiting is also still a problem. This morning it tried to take a regular exit and a bus-only exit even though the route says to continue on the freeway. 11 had this problem, too, which makes me think maybe it is a line tracking or labeling issue, but the lines shown on the screen appear correct, so I’m not sure what exactly is happening.

Finally, if you’re not bored yet, it behaved a bit odd in my parking lot at work. I have my waypoint set as a location in the parking lot, because using the street address puts me in the wrong place. 12 properly drove to the waypoint, then drove down the row of cars, hesitated a bit, and then pulled in taking two parking spaces. I’m not sure if it was trying to park, just got confused, or what happened.

Just a quick note, as far as I know freeway driving is still using the FSD 11 stack. There are probably some minor differences there, but the end-to-end training only applies to surface streets.

That’s good to know, because it certainly wasn’t any better. I wonder where the cutover happens between the two, because I had the 12 speed issue in the merge lane. The car showed the speed limit went up to 65, but the car stopped accelerating around 45.

Maybe that’s in the training data, though. Many people around here don’t know what a merge lane is for.

Not sure how the cutoff is handled. Could be just a speed thing. I haven’t done any testing beyond about 45 mph. And bugs in the transition wouldn’t surprise me either, like if it could get stuck in FSD 12 mode and be limited to 45 while on the on-ramp, and not properly switch to the high speed highway mode, or something.

Holy shit, I’ve never thought about this: can you get a car to drop you off and then go park itself? Then summon it when you’re ready to be picked up? That would seem to be a magnificent feature, not having to traverse huge parking lots on foot.

Yanno, I’d been looking for a cite for the claim with no success… but it turns out that it’s right in the release notes:
Imgur

So yeah, surface streets only. Though it still doesn’t say how the cutoff happens.

Earlier today, I was making a left turn that I could only make when the light just turned red and before the cross traffic started to move. How do self-driving cars handle such turns?

I’ve wondered about that case too. Haven’t seen a real-world example yet. But in principle, if it’s something drivers do, and they have examples, it’s something that FSD can learn.