FSD v12.x (end to end AI)

[AlanSubie4Life]

I can't speak to 12.4, but I know with 12.3.6 it seems pretty hit or miss on seeing animals. I've had it react to a dove that was sitting near the side of the road, but also completely ignoring a grown deer walking across, even though it was in the visualization (as a dog.). Not to worry, I braked before it got too close to the deer because it showed no signs of slowing down at all.

Yes that is what I would expect.

 

[swedge]

I can't speak to 12.4, but I know with 12.3.6 it seems pretty hit or miss (pardon the expression) on seeing animals. I've had it react to a dove that was sitting near the side of the road, but also completely ignoring a grown deer walking across, even though it was in the visualization (as a dog.). Not to worry, I braked before it got too close to the deer because it showed no signs of slowing down at all.

After we hit a deer, disabling the car, I learned that deer strikes generate a significant fraction of body shop work. In reviewing the dash cam footage, the deer was visible well before it lept into the road which was when I first saw it and reacted. So the car could have reacted sooner than I did, but it car did not respond at all. If AutoPilot or forward collision systems could better spot deer and take effective action, it would save a lot of grief and expense. Maybe, eventually, help lower insurance rates for Tesla cars. Also, deer tend to cross at the same places, so mapping these places when they are seen by the cars would also help.

 

[gforce2002]

After we hit a deer, disabling the car, I learned that deer strikes generate a significant fraction of body shop work. In reviewing the dash cam footage, the deer was visible well before it lept into the road which was when I first saw it and reacted. So the car could have reacted sooner than I did, but it car did not respond at all. If AutoPilot or forward collision systems could better spot deer and take effective action, it would save a lot of grief and expense. Maybe, eventually, help lower insurance rates for Tesla cars. Also, deer tend to cross at the same places, so mapping these places when they are seen by the cars would also help.

Absolutely. Deer are everywhere around here in this neighbourhood (I often find them sleeping on my lawn in the morning), so I get lots of examples of the car seeing/not seeing them. It's at best 50/50 even for them to be visualized. Given the prevalence of deer strikes, or even worse moose, surely Tesla could implement their detection better.

 

[Todd Burch]

12.3.6 has been doing very well with speed humps.

It's interesting. Multiple times I've been on a small suburban neighborhood road that has relatively hard-to-perceive speed bumps and it slows down perfectly for them every time.

Today on 12.3.6 I was in a shopping center parking lot where they have those UBER obvious large speed bumps with the yellow paint across the blacktop. Huge amount of contrast. The car didn't slow down at all for them. (Was going about 10 mph but 2mph is an appropriate speed for these bad boys).

I wonder if it was because it was in a parking lot and there haven't been any training sets with parking lot speed bumps? Shrug. Bit of a head scratcher.

 

[gforce2002]

It's interesting. Multiple times I've been on a small suburban neighborhood road that has relatively hard-to-perceive speed bumps and it slows down perfectly for them every time.

Today on 12.3.6 I was in a shopping center parking lot where they have those UBER obvious large speed bumps with the yellow paint across the blacktop. Huge amount of contrast. The car didn't slow down at all for them. (Was going about 10 mph but 2mph is an appropriate speed for these bad boys).

I wonder if it was because it was in a parking lot and there haven't been any training sets with parking lot speed bumps? Shrug. Bit of a head scratcher.

Yeah, there doesn't seem to be any rhyme nor reason. I've had it slow down perfectly for one and then the next one, the same size and painted identically only about 50 metres away, it completely misses.

 

[enemji]

but also completely ignoring a grown deer walking across,

It probably calculated the deer to cross the road before it got there

 

[enemji]

Today on 12.3.6 I was in a shopping center parking lot where they have those UBER obvious large speed bumps with the yellow paint across the blacktop. Huge amount of contrast. The car didn't slow down at all for them. (Was going about 10 mph but 2mph is an appropriate speed for these bad boys).

Yup....same with me too. It seems to think those are flat pedestrian crossing or something like that. Need training data for that too.

 

[enemji]

No there IS a minivan.

View attachment 1055022

Well at least it validates one of the points I had made earlier - these 1.2mp cameras (or the HW4 5mp cameras) are picking up way more than what we can see in bad weather or low light situations. The question has always been the decision making, not the data.

 

[MP3Mike]

In reviewing the dash cam footage, the deer was visible well before it lept into the road which was when I first saw it and reacted. So the car could have reacted sooner than I did, but it car did not respond at all. If AutoPilot or forward collision systems could better spot deer and take effective action, it would save a lot of grief and expense.

And then complaints of phantom braking would increase as it slows down and the deer doesn't cross.

 

[AlanSubie4Life]

And then complaints of phantom braking would increase as it slows down and the deer doesn't cross.

The nature of the complaints are exaggerated here. The issue with slowing down is the way it is done. Just ease off if there is a chance of a deer incursion. No need to do a Slamarama. Behave like an alert human.

Anyway, there's no reason to think the visualization will or will not indicate something which results in a slowdown. The visualization can show objects which it slows for. The visualization can not show objects which it slows for. The vehicle will stop for animals. And the vehicle will not stop for animals. It is all pretty consistent.

It's reasonable to expect the vehicle will sometimes slow for animals, and they will sometimes show in the visualization. It's best to stay alert and avoid collisions with animals if necessary via disengagement. When alert you'll usually see the animals before FSD reacts. Obviously no guarantee.

 

[MP3Mike]

Anyway, there's no reason to think the visualization will or will not indicate something which results in a slowdown. The visualization can show objects which it slows for. The visualization can not show objects which it slows for. The vehicle will stop for animals. And the vehicle will not stop for animals. It is all pretty consistent.

If that is an example of being consistent, what would be an example of inconsistent?

 

[JulienW]

Any bets that a few TMCers will wake up tomorrow and see 12.4.1 waiting. I'm kinda thinking it may be closer to the weekend but hope I'm WRONG (not wrong as in longer though ).

 

[swedge]

Any bets that a few TMCers will wake up tomorrow and see 12.4.1 waiting. I'm kinda thinking it may be closer to the weekend but hope I'm WRONG (not wrong as in longer though ).

Typically they roughly double the number release to "pending" each day. That would mean 6 tomorrow. ;-)

In TeslaFi, I've not seen a release of only 3 cars before in the FSD updates. 3 is pretty timid... Anyway, past performance does not guarantee future performance.

Anticipation is making me late, is keeping me waiting...

 

[AlanSubie4Life]

If that is an example of being consistent, what would be an example of inconsistent?

It's hard to say! It's consistent that there is inconsistency, as would be expected. No surprises.

 

[kabin]

Any bets that a few TMCers will wake up tomorrow and see 12.4.1 waiting. I'm kinda thinking it may be closer to the weekend but hope I'm WRONG (not wrong as in longer though ).

I doubt it but I've been surprised before with these wide FSD releases with questionable improvements. If anything I'd guess a mind blowing 12.4.2 release in two weeks and this time it really should be called v13.

 

[LowlyOilBurner]

Absolutely. Deer are everywhere around here in this neighbourhood (I often find them sleeping on my lawn in the morning), so I get lots of examples of the car seeing/not seeing them. It's at best 50/50 even for them to be visualized. Given the prevalence of deer strikes, or even worse moose, surely Tesla could implement their detection better.

I feel it worked better (detecting and reacting to deer) on 11.4.9(?, whatever was around in fall of 2023).

 

[kabin]

Any bets Chuck spares Elon the agony of uploading a 12.4.1 unprotected left turn video?

 

[PR0FESSOR]

FSD V12.x.x has been so bad for me, that I had to add to the complaints on this thread. In the area where I live, v12.3.6 is:

Slamming on the brakes for railroad crossings
Turning too close to medians on left hand turns
Turning too close to the right hand side of the road on right turns
Not staying centered in lanes; sometimes riding the rumble strip on the right and sometimes riding the rumble strip on the left

V12.x.x has been much less trustworthy for me than any or all of the 11.x.x versions I used. I could trust 11.x.x to perform well on freeways and major highways, but now, with V12.3.6, I cannot even trust it there.

Ugh.

Joe

 

[rlsd]

Typically they roughly double the number release to "pending" each day. That would mean 6 tomorrow. ;-)

In TeslaFi, I've not seen a release of only 3 cars before in the FSD updates. 3 is pretty timid... Anyway, past performance does not guarantee future performance.

Anticipation is making me late, is keeping me waiting...

Not many people are on Teslafi. I got v12.2.1 when Teslafi did not have many cars with v12.2.1.

 

[Twiglett]

And then complaints of phantom braking would increase as it slows down and the deer doesn't cross.

I remember watching a deer run across the width Route 22 in the NJ - it made a spectacular jump at the end to clear the Jeep driving in front of us