This happens to me too. It happens when I engage the turn signal (half press of the stalk to get three blinks), but for some reason the car doesn't recognize that I engaged the turn signal. Then when I move across the lane marking the car freaks out because it doesn't realize that I want to switch lanes. I'm not sure why the turn signal engagement isn't recognized--it happens probably about 5% of the time if I had to guess. Maybe I'm just not applying enough force to it.
-
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 accidents
- Thread starter Terminator857
- Start date
Real world FSD is simply to difficult to achieve in the time frame that keeps getting put out and with the equipment presently onboard these cars. Have had FSD beta for a while now which has made improvements along the way but the basic real world messy issues continue to linger with no improvements at all. I really wish they do achieve it and hope they do it first but realistically we are at present rate 10 years away for Level 4 and Level 5 will not happen with current equipment IMHO.
Probably not great
I'm driving a rental 2021 Ford F-150 with basic lane keeping technology and very rarely use it. Even in that limited function, it does weird stuff often enough to not be worthwhile.
I tend to agree but I suspect they will do this in less than 10 years since they will update the current camera/sensor count and location sooner. Tesla could also try to get approval for controlled access highways only for some level of 4/5 as they work thru the city/streets approval which will be much more difficult to get approval for.
So clearly SJ needs to replace the green pylons with plastic replica kids!
Those guys have been driving that SJ route for ages pushing the car to see if it messes up on those edge cases. Guess they found its limits.
Barklikeadog
Active Member
Members of the jury, I'd like to present exhibits A and B....
We never know, because training data decides. The right angle, light, outfit, background, and suddenly poof! Invisible to FSD.
If only there was some device to identify "things" that should not be in the road...
Tam
Well-Known Member
I guess there's a special effect called "green screen" in cinematography that makes the green screen invisible and thus, in this case, invisible green pylons seem to be true to its name for FSD beta.
EVNow
Well-Known Member
This is "Level 3" by Honda. The car almost hits a truck changing its lane. The driver had to disconnect. Best thing is this is a demo by Honda !!!
Tam
Well-Known Member
After owning a Tesla for a couple of months and using auto-pilot / auto-steer pretty much wherever I can enable it, including some city streets, it's been my overall impression that the car is slow at reacting to its environment.
A very simple example is when someone cuts you off and the car does not do anything for a couple of seconds. Sometimes the car that cuts you off is long gone to the next lane and that's finally when the Tesla starts reacting by braking (there's no longer any obstruction in front of it whatsoever, obstacle is long gone).
This is a very repeatable behavior of the car.
I wonder where the delay is and how Tesla could sort this out purely by software (with the existing camera / computer hardware).
Is it a delay in the cameras sending the data to the FSD computer, or is it the FSD computer not processing and reacting to the data fast enough, or is it that the car signals commanded by the FDS computer take too long to get executed?
Where is the delay in reaction time? And how could they fix this?
I recall Elon recently discussing camera photon counting (instead of processing full camera frames) to shave a few milliseconds on the data capturing.
What are your thoughts on this?
For those with FDS BETA (which can be disabled on demand), which stack is faster at reacting to the environment, BETA or non-BETA auto-steer? Is there a perceivable difference?
A very simple example is when someone cuts you off and the car does not do anything for a couple of seconds. Sometimes the car that cuts you off is long gone to the next lane and that's finally when the Tesla starts reacting by braking (there's no longer any obstruction in front of it whatsoever, obstacle is long gone).
This is a very repeatable behavior of the car.
I wonder where the delay is and how Tesla could sort this out purely by software (with the existing camera / computer hardware).
Is it a delay in the cameras sending the data to the FSD computer, or is it the FSD computer not processing and reacting to the data fast enough, or is it that the car signals commanded by the FDS computer take too long to get executed?
Where is the delay in reaction time? And how could they fix this?
I recall Elon recently discussing camera photon counting (instead of processing full camera frames) to shave a few milliseconds on the data capturing.
What are your thoughts on this?
For those with FDS BETA (which can be disabled on demand), which stack is faster at reacting to the environment, BETA or non-BETA auto-steer? Is there a perceivable difference?
Tam
Well-Known Member
That's why there as HW2 then HW3 and soon HW4. It's uncertain what number of HW will be able to catch up with the human brain's level.
I don't see any difference but it's only subjective.
Daniel in SD
(supervised)
Older versions of Autopilot would react quickly to detection of someone changing lanes in front of you and it was super annoying. You could see that the car was going to change lanes but Autopilot was oblivious until the car was more than a third in the lane. It seems they added the slow reaction to compensate for the poor perception of when cars are changing lanes. I prefer the new tuning.
The release notes are full of subjective terms, "improved", "smoother", "better". Even 4% of a made-up property is likely unmeasurable. I wouldn't call them release notes except where something definitely changed like "disabled rolling stop".Does the last paragraph of the 10.10.2 deal with green bollards?
"- Improved generalized static object network by 4% using improved ground truth trajectories."
Does that mean its recognition ability of green bollards is now increased to 4%?
This user summarized the 10.10.2 release with his own subjective notes. "Hopefully it sucks less". That's basically what Tesla said too.
People wanted better release notes with each new version; well they are getting longer-winded feel-good ones anyway. I guess that's progress.
My 2022 MYP non-beta doesn’t have that delay with anything. it actually anticipates people changing into my lane even if they don’t lol. I’ve driven 8000 Bay Area miles and probably 60% of it is AP.
How do you know the property is meaningless? Or that 4% is made up? "I dont understand it therefore its meaningless" isnt a valid argument. Tesla undoubtedly have a significant number of measurements they use to judge the algorithms the car uses, how else could they measure progress? (and no, I don't think "Dan D appreciation level" is one of those measurements).
Have you spent any time looking at the massive technical effort underway? Custom CPUs for the cars, and for the NN training. Massive simulation efforts with camera-accurate rendering. Super-computer scale NN training. But nooooo!!!, they make stuff up because its all just a silly game in your mind.
EVNow
Well-Known Member
Those are good release notes. The appropriate expectation would be 4% is essentially not noticeable for us.
ps : Someone asked Elon whether they can get detailed release notes and he said yes. The detailed notes we see is a result of that i.e. more or less internal change description being made public.
The numbers mostly don't mean a lot without a basis for comparison, and no basis was ever provided -- the release notes hit and started talking about % improvements.
It's like saying "I'm going 4% faster" -- 4% faster than what? 4% is a very different number if you're going 100mph vs 2mph. I'm sure the numbers mean something to someone, but they're not particularly helpful without a baseline.
It's like saying "I'm going 4% faster" -- 4% faster than what? 4% is a very different number if you're going 100mph vs 2mph. I'm sure the numbers mean something to someone, but they're not particularly helpful without a baseline.
Last edited by a moderator:
Similar threads
