Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register
This site may earn commission on affiliate links.
Pothole avoidance
On my list is something similar.
- when going over speed bumps (in parking lots), drive such that one wheel is on the speed bump and the other is not, when possible. I wonder whether they will ever train this.

Ofcourse the main features I’d look for to see if V12 progress is not limited like 11 and 10
- School zone driving
- School bus recognition and appropriate driving
- usable summon & banish in parking lots
 
  • Like
Reactions: edseloh and GeoX750
On my list is something similar.
- when going over speed bumps (in parking lots), drive such that one wheel is on the speed bump and the other is not, when possible. I wonder whether they will ever train this.

Ofcourse the main features I’d look for to see if V12 progress is not limited like 11 and 10
- School zone driving
- School bus recognition and appropriate driving
- usable summon & banish in parking lots
I’m looking for progress on stopping.
 
I’m looking for progress on stopping.
You mean where it stops … yes, ease of bug fix is an important thing to figure out. Unlike V11, all this would have to be done through selection of training data ?

BTW, if it’s really all NN, why does it not stop close to the line as most people do ? Or is this because people who come to a full stop, usually stop way behind the line and that’s what Tesla has chosen as the training data because of NHTSA ?
 
  • Like
  • Informative
Reactions: MP3Mike and GWord
BTW, if it’s really all NN, why does it not stop close to the line as most people do ? Or is this because people who come to a full stop, usually stop way behind the line and that’s what Tesla has chosen as the training data because of NHTSA ?

I have to take annual, professional driver training as part of my profession. We are specifically taught to stop before the white line becomes occluded by the front of the vehicle. It’s been beaten into my head this way for the past 20+ years of courses. There’s probably some truth to your hypothesis, assuming it’s mostly only professionally trained drivers coming to a full stop, and that’s the training data they have.
 
You mean where it stops … yes, ease of bug fix is an important thing to figure out. Unlike V11, all this would have to be done through selection of training data ?

BTW, if it’s really all NN, why does it not stop close to the line as most people do ? Or is this because people who come to a full stop, usually stop way behind the line and that’s what Tesla has chosen as the training data because of NHTSA ?
My hypothesis is that it’s trained with simulated data.
They stop where the front camera can see the line because otherwise the car “forgets” that it’s there and somehow that causes problems.
 
I have to take annual, professional driver training as part of my profession. We are specifically taught to stop before the white line becomes occluded by the front of the vehicle. It’s been beaten into my head this way for the past 20+ years of courses. There’s probably some truth to your hypothesis, assuming it’s mostly only professionally trained drivers coming to a full stop, and that’s the training data they have.
Doesn’t that cause issues for signals controlled by cameras …? Hopefully all of them are setup to monitor wider area, but I have seen had issues with cameras if you are not close to the line.
 
I have to take annual, professional driver training as part of my profession. We are specifically taught to stop before the white line becomes occluded by the front of the vehicle. It’s been beaten into my head this way for the past 20+ years of courses. There’s probably some truth to your hypothesis, assuming it’s mostly only professionally trained drivers coming to a full stop, and that’s the training data they have.
This is kind of an arbitrary rule because the result depends on driver position and the car dimensions. It’s not a great rule. This leads to being 8 feet from the line or so in Model 3 (depending on driver height, seat position, etc.). This is too far.

Stopping too early leads to confusion at stop-sign-controlled intersections. It also leads to several obvious significant negative impacts on safety at stop signs and traffic lights as discussed earlier. Having the ability to see the intersection is super important.

Stopping so that you can see the tires of the vehicle in front is a similar rule which has some arbitrariness to it.

In the end should probably leave a couple feet to the line from the front of the vehicle. Rather than following a rule leading to sub-optimal results , just do that.

And much more room to the vehicle in front, so that rule is probably a better one for that application. It’s not as important to pull to the bumper in that case, and is obviously a problem for several other reasons.
 
Last edited:
  • Funny
Reactions: Dewg
First commute for me with 12.3.6 was mostly pretty good.
On a four lane road with double yellows and cats eyes for a divider, unsurprisingly it decided the left of our two lanes was the best, but then was so far left that the car was bouncing the left tire on the cats eyes. There was no oncoming traffic so I left it there to see if it would improve its position - no it didn't and it was getting with the constant noise so I cancelled.
Still hard acceleration to 46mph from rest then a granny like imperceptible increase up to 55mph.
On the highway, several times it signaled to switch lanes but chose to do it when there were already cars in the way.
The rubber banding behavior in traffic has definitely got worse. It will get within one car length of the lead car then allow it to get several hundred yards away before speeding up to one or two car lengths, rinse and repeat. That was old, old AP behavior.
So definitely drives more human like, just an a-hat human driver :D
 
My hypothesis is that it’s trained with simulated data.
They stop where the front camera can see the line because otherwise the car “forgets” that it’s there and somehow that causes problems.
A lot of the time it stops closer to the right place.

When it stops where I can see the line clearly, the view of the cameras of the line must be nowhere near obscured.

It’s possible there is something like this involved but the inconsistency is so bad it might be something else. It also may be location dependent. Or related to the scene somehow.

Really hard to know. All over the map. Garbage.
 
It also leads to several obvious significant negative impacts on safety at stop signs and traffic lights as discussed earlier. Having the ability to see the intersection is super important.

Stopping so that you can see the tires of the vehicle in front is a similar rule which has some arbitrariness to it.

In the end should probably leave a couple feet to the line from the front of the vehicle. Rather than following a rule leading to sub-optimal results , just do that.

And much more room to the vehicle in front, so that rule is probably a better one for that application. It’s not as important to pull to the bumper in that case, and is obviously a problem for several other reasons.

I’m going to have to disagree unless you can produce quantitative evidence to back this up. If so, I stand corrected and I’m sure every professional driving system in North America would be all ears. My corporation does about 20M miles per year across 1500 regular drivers. I just did a quick query in our incident database and for the past 20 years of data there are no vehicle incident root causes that point to stopping too far back at a traffic control device.
 
I’m going to have to disagree unless you can produce quantitative evidence to back this up. If so, I stand corrected and I’m sure every professional driving system in North America would be all ears. My corporation does about 20M miles per year across 1500 regular drivers. I just did a quick query in our incident database and for the past 20 years of data there are no vehicle incident root causes that point to stopping too far back at a traffic control device.
You should DM Alan and let him know the name and contact info of your driving instructors so he can reach out to them and let them know they're wrong. 😁
 
You should DM Alan and let him know the name and contact info of your driving instructors so he can reach out to them and let them know they're wrong. 😁
There is definitely a lot of bad info out there. Not sure what the deal is with it.

Anyway the point is that for safety it is important to stop in a place that affords proper visibility of pedestrians and the intersection and does not confuse other drivers.

Just do that.

The law also requires that you stop at the line. If a motorcycle is stopped in front of you at the line you don’t get a free pass to just roll though the stop line! Then remove the motorcycle. There’s a reason for the law requiring this!!!
 
I’m going to have to disagree unless you can produce quantitative evidence to back this up. If so, I stand corrected and I’m sure every professional driving system in North America would be all ears. My corporation does about 20M miles per year across 1500 regular drivers. I just did a quick query in our incident database and for the past 20 years of data there are no vehicle incident root causes that point to stopping too far back at a traffic control device.
But I assume such drivers stop in the right place! So if anything that speaks to the importance of stopping in the right place.

Hopefully the dangers of stopping 8-10 feet behind the line (which happens with FSD on occasion for unknown reasons) are apparent to all.
 
There is definitely a lot of bad info out there. Not sure what the deal is with it.

My sense is that in this particular area it may be related to:

Dunning-Krueger effect

The Dunning-Kruger effect is a cognitive bias in which people wrongly overestimate their knowledge or ability in a specific area. This tends to occur because a lack of self-awareness prevents them from accurately assessing their own skills.
 
My sense is that in this particular area it may be related to:

Dunning-Krueger effect

The Dunning-Kruger effect is a cognitive bias in which people wrongly overestimate their knowledge or ability in a specific area. This tends to occur because a lack of self-awareness prevents them from accurately assessing their own skills.
Definitely!

It’s important to stop in the right place. On that we can agree.

Fortunately in most cases FSD stops such that the line is not visible (for my vehicle). But would be good to fix the errors, for sure.
 
Last edited:
  • Helpful
Reactions: GWord
My 12.4 wish list

don't use more than 2/3 throttle to accelerate to the speed limit on any road with a speed limit below 55 mph.
don't use more than 1/2 throttle to accelerate to the speed limit on any road with a speed limit below 45 mph.
don't use more than 1/3 throttle to accelerate to the speed limit on any road with a speed limit below 35 mph.
don't use the throttle to speed up when a line of cars ahead of you is stopped at a red light
don't keep the speed the same zooming by a long line of cars in one lane when approaching a red light only to hit the brakes hard 2 car lengths from the red light. Regen or at least glide at neutral approaching that red light.
don't use the throttle to speed up faster than a car ahead when a car in front of you just started to leave a red light
don't use the throttle to speed up then slow down then speed up. If you are switching back and forth between noticeable regen and noticeable forward throttle you are doing something wrong.