NHTSA requires the slowing and stopping before lights and signs
This requirement is
not related to the current complaints, at all! Not the least little bit related. It's not the same thing at all.
Put a different way: completely different thing. Everyone is on board with the car stopping at lights and signs as legally required.
I'm sure Alan would have an absolute fit with how soon it starts slowing down and how long it takes to get to the stop line and actually stop.)
Yes, it is very annoying. It's nearly completely solvable in typical cases by pressing the accelerator, fortunately, with no jerking at all.
Not a single hard braking event. (I don't think this has ever happened on this route before with FSD driving.) It's great.
Yeah, we'll see if that persists. As you know it can vary greatly from drive to drive based on very specific conditions (traffic, speeds, conditions, light timing, etc.). Not enough data for me with stopping yet with 12.3.4; just a short drive and I used my current strategy of
pressing the accelerator to prevent any hard braking, which works great in most cases. So I could have masked any improvements.
Look forward to letting it do its thing tomorrow and we'll see! Seems like something that could be tuned without neural net modifications. Super unclear how that would be done, of course.
The
left screenshot shows a dark black line at the end of the blue path indicating FSD was planning to stop probably because the light was red.
No. You grabbed screenshots a couple frames too early.
To be clear, my earlier references to "0.5s reaction time" were based on the regen bar in the yellow light video, not any visualizations. The regen bar is a fairly instantaneous measure of car behavior, and represents car reaction time (to me). That
yellow light video shows time 0.5 seconds. Since I was driving, I know that the car's intent was to stop at that yellow (since I stopped it from doing so), so it's a decent way to measure effective reaction time. However, I went back and looked at the planned path, and that seems even better since the regen bar has to blend in the reaction (which takes a few frames). So actual reaction time is closer to 0.35 seconds.
Effective reaction time still appears to be closer to 0.5s (onset of major braking).
Back to this case:
As you pointed out and I checked above in the yellow light case, blue path is a superior and accurate instantaneous representation of FSD intentions, but it can change at any moment. I went back and looked at it, and I was happy to see that the vehicle intent actually matched what I felt.
You can see in image 1 it extended the path
before the light turned, and in image 2 as the light is turning the planned path extends further intending to follow the lead vehicle (based on yellow light case we know
that was NOT in response to the light). It had no permanent intention of stopping at the red light; that is now 100% clear from video evidence.
This image shows the scene 0.5 seconds after the green, the fastest known reaction time I have seen demonstrated (as discussed it may actually be slightly faster, 0.35 seconds, but effective time is 0.5 seconds):
before it got there, it saw the green and maintained its acceleration without ever needing to decelerate for the stop.
As detailed above: Based on it being at 6mph, and more like 8mph at the earliest likely possible reaction moment (0.5 seconds after red), within just a couple feet of the line, as illustrated above, it never had any permanent intention to stop, even if the light had stayed red. It had done the commit and did not have plans to jam on the brakes if the light stayed red. It was going, d**m the torpedoes. I don't know why the path planner showed an initial plan to stop at the line - the acceleration did not match that intent, so it seems that it overrode its own planning.
. That is why you can't do it.
