old pilot
Member
I think you're looking for an adjective here not a noun.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Wiki MASTER THREAD: Actual FSD Beta downloads and experiences
- Thread starter jebinc
- Start date
Main reason is probably because the perception is currently not good/sure enough until the vehicle gets closer with a better view of the turn. Theoretically the neural networks with more training should be able to make the prediction earlier and more confidently resulting in a smooth turn, but at least right now, it's actually very important that FSD Beta makes these "real time" jerky corrections to its path.
There's a left turn where FSD Beta 10.x doesn't see the destination past the median until it is in the middle of the intersection, so it actually starts the turn aiming towards oncoming traffic to the left of the median. Here's a Google Street View of the turn except it's from a higher vantage point, which I believe also helps the Y take the turn better than the 3:
Similarly, right turns can get confused by bike lanes and/or parking lanes such as this turn that has both (and wife actually messaged me complaining about this just yesterday
). And of course faded lines or those covered up by dirt/tire tracks are likely to confuse the perception:
CharleyBC
Active Member
We had an odd one a few days ago. I was on the leg of a T intersection, coming to a stop sign, from which I would turn left. There was a line of 8 or 10 cars in front, all waiting to do the same thing. It's a neighborhood street, with plenty of space for on-street parking, though it happened no one was parked there at the time. FSD apparently became impatient with the line of traffic, and started to pull to the right around the car in front (into where parked cars would have been).
I interrupted. Aside from being a jerk maneuver, it wouldn't have accomplished anything useful. It could have passed the cars ahead, but then what? Turn left from the right? Try to force its way back in line? Oh well!
I interrupted. Aside from being a jerk maneuver, it wouldn't have accomplished anything useful. It could have passed the cars ahead, but then what? Turn left from the right? Try to force its way back in line? Oh well!
Yes, what your first paragraph explains is exactly what’s going on.
Last edited:
So, FSD is pretty much completely useless in light to moderate rain or snow, at least that’s what I’ve come to the conclusion of. Just now, while on FSD, coming up the rural road I live off of, it’s like it was almost trying to dodge snowflakes, swerving aggressively (moderate snowfall this evening on my way home). During the day in rain, FSD was pretty much at a 50/50 on the confidence level of completing a manure. Mostly comical at best. Lulz. Still interesting to see how conditions/lighting changes things.
Ben W
Chess Grandmaster (Supervised)
I've heard comparisons between horses and autonomous vehicles before, but this takes the cak.
Last edited:
AlanSubie4Life
Efficiency Obsessed Member
Good catch. Completing a manure does pretty much describe FSD’s driving style.
Snow especially when the road is covered will pretty much prevent FSD from being used at least for the next couple of winters so that Tesla can get real data. Not sure why this doesn't get discussed more here but suspect it will once the larger Beta tester group actually uses FSD with an a inch or more snow on the ground and/or during a snow/ice storm. If you're actually considering using robotaxi and live in a snowy climate just assume there will be downtime for customers. Then again IMO actual robotax service is at best 2+ years away. There is a reason Waymo selected Arizona.
I would never think of using FSD on snow covered roads. What I’m getting at, the fact it was snowing (the roads were just wet) and it didn’t work.
Drove 150 miles yesterday on the highway and 10.4 performed pretty well. During rush hour traffic the ride was smoother then earlier versions. Used chill setting just to see how it worked and it worked pretty well. Fewer lane changes and acceleration on a lane change was pokey but that was expected. All in all a positive drive. Per usual the following still need to be addressed and have been widely known since forever.
- Don't move into a faster lane when approaching an exit
- Don't move into a faster lane it that causes the oncoming car to slow down. This drives me crazy and I sometimes cancel the lane change. Chill was better since it did fewer lane changes but chill still wants to move lanes when cars are coming up too fast. Will try more aggressive settings next time.
- In heavy traffic just move over to the right lane when approaching an exit. Don't wait forcing me to take over or miss the exit.
- Merging traffic entering the highway. Sometimes it cannot make up its mind whether to slow down or speed up. Usually caused at the last minute by the merging driver deciding to accelerate or slow down at the last minute.
- FSD integration has definitely improved exit ramps. Not as many "hot" exits requiring manually slowing or disengaging. Easy to tell when my wife grabs the door handle as we approach a tight exit ramp at 60mph and the exit ramp speed sign is 25mph.
Just got 10.4 for as my first FSD beta. I live a heavily wooded area with a ton of 2 lane roads. Main concerns: 10.4 hugs the left of the lane too close. As a human I hug the right side more, without FSD it centers in the road, FSD hugs the danger side - left side. Roads that are heavily covered in leaves like we have really drive it bonkers.
EVNow
Well-Known Member
Yesterday I decided to just use AP, instead of FSD. Mainly because I got the booster on Wednesday and had mild headache. A few things I learnt.
- AP works better than production AP. Lane keeping is better and doesn't get confused when the lane splits.
- AP worked like before in rain. None of that "severe weather detected" stuff. This is really a great discovery for me since I need AP to drive at night when raining. AP definitely sees the lines and can make out lanes better than me. So, from now on - until FSD improves in these conditions - I'm just going to switch to the AP profile and use AP when it is raining (esp. at night)
- The blue lines it shows are rock solid in rain at night and I've no idea how it can figure out the lines that well. With FSD the lines were all over the place.
- AP works better than production AP. Lane keeping is better and doesn't get confused when the lane splits.
- AP worked like before in rain. None of that "severe weather detected" stuff. This is really a great discovery for me since I need AP to drive at night when raining. AP definitely sees the lines and can make out lanes better than me. So, from now on - until FSD improves in these conditions - I'm just going to switch to the AP profile and use AP when it is raining (esp. at night)
- The blue lines it shows are rock solid in rain at night and I've no idea how it can figure out the lines that well. With FSD the lines were all over the place.
How can AP work better than production version when that stack remains production? Single stack doesn't exist yet, does it?
Vision only vs Vision + radar? Maybe more recent effort put into the newer vision only code.
I would think Radar would still be a head of Vision, no?
Over time, vision should win, but not yet - as widely reported by many.
Radar shouldn't have anything to do with line detection. I'm speculating that the vision improvements are going into vision only first.
But according to Musk radar just adds noise and no improvement in signal compared to what they get using their pseudo-lidar with vision only.
Terminator857
Active Member
I wouldn't trust a biased opinion that has billions at stake because of a radar chip shortage.
Similar threads
