Yea looks like regen happens AFTER you disengage. Side note: Im surprised the over sensitive Youtube algorithm did flag your video for Copyright with the music playing.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD Beta 10.69
- Thread starter Buckminster
- Start date
AlanSubie4Life
Efficiency Obsessed Member
For now at least! Maybe if the clip is short enough they don't care. So far the AI hasn't found me out in the Copyright claims. Lots of background noise too.
Last edited:
EVNow
Well-Known Member
Had my longest zero disengagement drive yet. 25 miles with 15 on NOA.
Couple of interventions to (not confirm) lane changes on NOA - because the changes were dumb. Why does NOA want to change to right most lane when it's going to be exit only in 30 seconds - just so it's in the right most lane ?
Couple of interventions to (not confirm) lane changes on NOA - because the changes were dumb. Why does NOA want to change to right most lane when it's going to be exit only in 30 seconds - just so it's in the right most lane ?
AlanSubie4Life
Efficiency Obsessed Member
I guess the disappearance of the set speed indicator is a better indicator of the moment of disengagement than the steering indicator (takes time to flicker off)?
AlanSubie4Life
Efficiency Obsessed Member
Also does not visualize (though could have been out of range of the displayed visualization - I really have to use the enlarged view!), or apparently react to, a coyote.
No collision risk here so this was ok, though it could have gone either way.
No collision risk here so this was ok, though it could have gone either way.
Seems to drop from 45 to 30 so there was some braking.
AlanSubie4Life
Efficiency Obsessed Member
Yeah, that was me. Watch when the speed sign disappears at around 7.6 seconds. That is when I hit the brake, I think (2 seconds after I suspected things were going to go awry).
I mean, I assume FSD would have stopped, but it would have been unpleasant and I wasn't about to find that out. We'll never know. It might not have stopped, since maintaining speed would have meant the Nissan would have been less in my lane at the time I arrived at that vehicle's location. But it probably would have stopped.
In any case, there's clearly not a path for my vehicle at the moment of disengagement, and there was no reaction from FSD up to that point.
AlanSubie4Life
Efficiency Obsessed Member
As I added above, you can clearly see there is no path for my vehicle at the moment of disengagement. There was no reaction from FSD. So at a minimum, FSD is slower to react than a human (we all know that of course).
This is the moment before disengagement. There's nowhere for FSD to go. Note the strong application of power by FSD. Balls to the wall. (If I had been manually controlling the vehicle, there would definitely have been regen taking place at this point.)
Would it have slowed? Probably. I have no idea. It had not flagged the vehicle as an issue, so that doesn't seem great. I don't know the conditions under which it does that. Maybe it would have used the bike lane!
Last edited:
While I bet Beta wouldn't have actually hit the car it was at the point that the amount of jerk needed to avoid it was getting extremely high. Looking at it this is an easy and obvious avoidance maneuver for a human and Beta should have already been reacting to make it a more comfortable maneuver. Another 500ms and the amount of braking force and wheel turning beta would have to use would have gone up exponentially.
EDIT: Just to add if Beta is going to be better than all human drivers it should already be reacting BEFORE any of us humans have time to respond. So if we have time to take over BEFORE Beta reacts then it is a fail.
Last edited:
Ah! Got it. Thanks! I didn't see the disengagement. Roughly at 0:07.7. The UTube is too blurry to make out the blue steering wheel, but the max speed vanishes.
It looks like your silly car actually accelerated a bit, just before you disengaged.
Not that I would have waited to find out either, but I wonder what would have happened without intervention. Clearly there was still time to slow down when you disengaged, but a moment later it would have been a slam-on-the-brakes deal. (And a nasty safety score ding if Tesla measured FSD the way they measured us! )
EVNow
Well-Known Member
Difficult to make out in the video whether it would have slowed down or not.
Today on NOA - I could see the traffic slowing down in the front, but NOA wouldn't .... I thought I'd have to emergency brake .... but NOA slowed down well in time. On another day I'd have disengaged.
Basically FSD/NOA slows down later than we do.
Right .... with 1 disengagement in 10 miles, FSDb has to improve 100x to 1000x to equal humans.
ps : Ofcourse FSDb can still be better than humans *sometimes*. It slowed down for pedestrians once before I had noticed them .... (it was dark, and interestingly similar to the late slow down we all saw in one of the posted videos near a roundabout).
Last edited:
AlanSubie4Life
Efficiency Obsessed Member
I really did wait as long as was comfortable, and we’ll never know.
Yep. When we see that flip or come in line with human drivers, we can discuss whether we are getting close to human performance (and even then lots of hurdles of course). Currently not a discussion worth having.
EVNow
Well-Known Member
Yes - sometimes I feel the same way with roundabouts. I disengage a lot of times when FSDb doesn't seem to be slowing down with a vehicle coming from the left. I don't want to find out by waiting (what may be) for too long.
AlanSubie4Life
Efficiency Obsessed Member
I’d say they aren’t really regressions (except in documented cases of regressions of course).
....and every drive is a unique drive that can produce different results and impressions that may appear as a regression but it is just likely anecdotal.
EVNow
Well-Known Member
There is some talk of this on Twitter (DirtyTesla) as well.
I've felt it many times. It could be a number of things ...
- With every release, there is a little change in the way FSDb drives and it takes some time for us to get used to it
- Just random
- There is some local setting that is cached and used by the planner that gets reset with each release
Cool. My understanding also is that there is no “learning” in the vehicle - the Model is computed centrally - then runs on individual cars - there is no local AI.
EVNow
Well-Known Member
Yes, but doesn't preclude the planner (which mostly uses normal code, not CNN) from using some stored / cached information locally in the car.
Similar threads
- Article
