FSD v12.x (end to end AI)

[Peter N]

"Not following code, but making decisions."

You summarize the change well. My experience closely parallels yours.

 

[PianoAl]

Does not slow down when it is artificially boosted with accel? That's not good... So you can only use this when you are on a straight roads or at least you are at least 30 secs away from a curve?

Exactly.

 

[drtimhill]

Finally had time to do my standard test routes with 12.3, a few observations:

-- I wont repeat what everyone has observed, its FAR more human-like and smooth etc.
-- 12.3 seems more "relaxed" about lane centering. I found this interesting. It doesnt "wander" so much as its less "locked into the middle" the way 11.x was. The car tends to smoothly use the lane width e.g. when taking turns in the same way as a human (we're seeing training data I think).
-- Occasionally, this "relaxed" lane use got a bit too much .. with the car pretty close to the lane edges. However this didnt happen when car were in adjacent lanes.
-- Once, the car mistook a marked pedestrian lane close to an intersection with a turn lane, which caused an intervention.

 

[Dave EV]

Used 12.3 for a bit on my commute this morning with the new auto speed feature enabled and "normal" driving aggressiveness. Some behavior is the same, but some is new.

• Lane centering behavior is very worrying on my main route, which is a 2-lane 50-55 mph highway with erratic lane markings for turn lanes, etc. It often rode the bot dots on both sides of the lane for extended periods of time. Not the kind of behavior you like to see when the left side is opposing traffic and the right side is a bike lane that frequently has bikes in it. This alone is a huge drawback in my opinion.
• I did not observe any big changes in it being "human-like" or smooth. I still find that deceleration fails to anticipate slowdowns in traffic sufficiently and overuses friction brakes and that acceleration often falls into either category of accelerating too slowly or accelerating too quickly.
• The new auto speed setting tends towards driving too slow. I found myself nudging the accelerator quite frequently to either accelerate faster or close a gap in traffic. But then on another residential road, Tesla still has the speed limit wrong (40 mph in the car vs 25 mph posted in an area with lots of pedestrians / bikes on the road), so I had to disengage.

On the other hand, my wife reported that it gave her the best drive ever over 30 miles with just two minor dis-engagements with a mix of freeway and surface streets and did say it was human like, so not sure what's up with that. She does drive less aggressively than me, though, in general.

 

[JB47394]

What's the rule for turn signals? Seems quite individual, like many aspects of driving. Some want it early and some want it late. The only rules we have are traffic laws, which many states have as minimum 100 feet from the turn. This is going to be a goldilocks problem.

Well, I consider this stuff self-evident, but I don't work for the government. For example, traffic lights rules used a very simple formula until regulators started thinking about it all. With turn signals, the goal is to communicate with other drivers. If I've got a guy on my bumper and I'm going to brake soon for a turn, I'm going to make sure he knows that I'm turning. If there's nobody behind me, but somebody ahead that would be interested in my turn, then I'm going to signal them. Lastly, if nobody is around, I can use the 100 foot rule or whatever just because I may not see somebody.

This seems like a pretty good training opportunity such that the car ends up seemingly psychic, and folks start saying "It was just what I'd do!"

 

[AlanSubie4Life]

I did not observe any big changes in it being "human-like" or smooth. I

In order to experience the smoothness, you have to set aside and ignore all the jerky behavior!

I’ve been surprised at how smoothly it deals with certain scenarios. It’s definitely less robotic in some cases.

Overall your experience seems to match mine. I like the ability to drift in lanes, but it can be overdone, riding on the dots, or hitting curbs. It definitely drifts in response to other vehicles (sometimes!) and can create buffers this way (good), but it is not consistent. And it will still camp in blind spots. I feel like sometimes it tries to avoid that but not sure about that. Not enough time/samples.

 

[FSDtester#1]

Have to say 12.3 is pretty amazing solving nearly all my normal problem areas. I have had a couple of navigation issues where it didn’t get over to catch an exit but by and large this is a pretty amazing step forward.

I have a clip from the front dashcam that is sooooo good, I won't even give a big hint what fsd did. Reason? No sure if my IQ is high enough to upload it in the highest resolution.
If I can't figure it out, will give the details. Made the wife watch it.
She was impressed ...2 kids in the car when it happened, and we all scratched our heads at first going, "What is it doing"....I had some help from the kids to get her near the "Elon Bad Silent driving people killer".
Only hint, a bicyclist may have got seriously injured or became road kill if V12 wasn't on....
Ok, now to get out that damn USB drive, hidden behind God knows what, and find a working laptop.

 

[X Fan]

My 2-days in experience in NC’s Triangle area:
Local driving:
— as noted by others, vehicle always delivers a mph that is below speed limit/target unless I intervene and increase speed. That occurs in an open, straight road with no one in front of me.
— vehicle still gets confused about Cary. NC’s penchant for a mostly 2 lane road that vary between 2 and 4 lanes for brief moments. Vehicle always goes into the “extra” lane even though it will end shortly either as a turn lane right or just consolidate back to one (Cary forces developers to build the extra lane but if adjacent land isn’t built then you go back to one lane in your direction)
— vehicle does not slow down for speed bumps and requires intervention (often found on local streets here to slow down folks cutting through neighborhoods)
— this morning I was in a medical building parking lot looking to travel back to home. Vehicle ignored the immediate exit (6’ away) and proceeded to circle the entire parking lot and connected to an adjacent parking lot (rather than the closest exit again) to finally exit.
— Vehicle missed executing a u-turn and instead started attempting a three point turn on a typically busy divided 4 lane road last night (2 each way). I allowed the maneuver since it was later in the evening to see what it would do…(I would never allow it in normal hours). I intervened.
— it did handle two round about more successfully than typical
— turns are still very slow and much longer than humans driving

Highway-
Generally ok but still fails in a very normal highway exit to another highway. At i40/i540 exit , the vehicle first heads to the ”north” exit, then when I intervene it gets to the proper lane for the “south” exit. It then proceeds along the exit lane but doesn’t make the right turn to exit (wants to go straight again).
— rush hour traffic: still has tendency to leave the proper (far right) lane within 2 miles of your next exit veering into middle lane (when you know it will be very challenging to get back into the exit lane)
— program has a very bad habit of changing lanes into an area that is in an adjacent vehicle’s blind spot. I never do that recognizing the danger of making that type of lane change.

This is occurring with my wife’s year old Y.

 

[Dewg]

It’s really not.

Just signal correctly.

Signal early, before slowing. It is to signal intent.

If it creates ambiguity, consider waiting if conditions and speeds allow (on a fast road you may want to signal early and position your car and maintain speed to avoid ambiguity while also providing buffer in case someone makes a mistake).

There is more to it than that of course.

But signaling 100 feet before the turn (what v11 seemed to do, if at all) is definitely not enough in many situations - that could easily be less than 2 seconds before you turn.

v12 is better than v11 with this. There may be situations where it is worse than v11, though (just have not seen).

Of course it is. v11 signalled to late, and people are saying V12 is signaling too early. So the correct signaling is somewhere in the middle, hense goldilocks. Not rocket science to understand the metaphor.

 

[TK211X]

I can't wait to wake up tomorrow and complain about not having 12.3.1 even though I don't even have 12.3.

 

[AlanSubie4Life]

Of course it is. v11 signalled to late, and people are saying V12 is signaling too early. So the correct signaling is somewhere in the middle, hense goldilocks. Not rocket science to understand the metaphor.

No. I think it is pretty well agreed that v11 did not signal properly.

There has been one clear complaint about v12’s pattern (???) of signaling too early, without detail. And a bunch of people saying it is much better.

Nobody is saying it is perfect. But it is better, and it is not a Goldilocks problem. (Assertiveness of acceleration potentially IS that sort of problem, for example.)

A complaint about v12 signaling early where ambiguity exists is NOT a complaint about signaling too early. It’s not the same thing. It’s not an implication that v12 should be signaling later. It’s saying FSD should be aware of the context, selectively delaying signaling depending on the situation.

 

[PianoAl]

I have a clip from the front dashcam that is sooooo good, I won't even give a big hint what fsd did.

It better include the street beaver.

 

[FSDtester#1]

It better include the street beaver.

Great idea! I will hit the record button while passing over it next time I leave the house .

 

[FSDtester#1]

No. I think it is pretty well agreed that v11 did not signal properly.

There has been one clear complaint about v12’s pattern (???) of signaling too early, without detail. And a bunch of people saying it is much better.

Nobody is saying it is perfect. But it is better, and it is not a Goldilocks problem. (Assertiveness of acceleration potentially IS that sort of problem, for example.)

A complaint about v12 signaling early where ambiguity exists is NOT a complaint about signaling too early. It’s not the same thing. It’s not an implication that v12 should be signaling later. It’s saying FSD should be aware of the context, selectively delaying signaling depending on the situation.

I would like to be Switzerland regarding this, however I think all of our driving styles are different. If we polled all of America about what they think the proper time to turn on the turn signals, I think answers would vary greatly.
Too subjective of a topic to make a bona-fide answer to.

 

[rlsd]

I wish Tesla can give us the option to record the touchscreen. Technically it's very easy to do. It's has nothing to do with FSD or NN or EE. But the downside is the YouTubers will be out of business, something Tesla may not want to do.

 

[AlanSubie4Life]

I would like to be Switzerland regarding this, however I think all of our driving styles are different. If we polled all of America about what they think the proper time to turn on the turn signals, I think answers would vary greatly.
Too subjective of a topic to make a bona-fide answer to.

I’m definitely not saying there is one right way. There’s a distribution of timings when humans are involved. V11 was way outside of that distribution a lot of the time (and sometimes failed to signal at all, by design apparently).

V12 is within the range of a typical distribution of timing on most signal events now. Special cases are likely poorly dealt with.

“Goldilocks” suggests it is too early now for most people, which I do not think is correct.

 

[sleepydoc]

Great idea! I will hit the record button while passing over it next time I leave the house .

We’re waiting on on a beaver clip!

 

[j0shm1lls]

We’re waiting on on a beaver clip!

I'm just waiting on the next wave...

 

[Kablooie]

Stop signs continue to be amazing for me so far. Solid A

There's a new stop sign near me and the V12.3 went right through it then stopped after it had passed. This probably indicates that it relies on mapped stop signs more than the video visuals.

 

[jabloomf1230]

An update on multilane traffic roundabouts. 12.3 does marvelously in even heavy traffic when exiting the roundabout's 1st exit (right turn equivalent) and 2nd exit (straight equivalent) because it stays in the outer lane. However, when exiting the 3rd exit (left turn equivalent) it becomes a blubbering mess. Before entering the roundabout, it first signals a left turn, then it swerves into the inner lane where it stays until it tries to exit. Then it puts on its right turn signal and cuts off all the traffic in the outer lanes. Admittedly, multilane roundabouts are a dumb design and further, most US drivers don't understand them. With regard to roundabouts, FSD seems to rely on Catch 22. If everyone else drives poorly in a roundabout, I'd be a fool to do otherwise.