mtndrew1
Active Member
If I turn on FSDb without telling passengers it isn’t long until the cries of “wtf is going on with the car?!” begin.
It’s a very uncomfortable experience even when it’s working properly.
-
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
If I turn on FSDb without telling passengers it isn’t long until the cries of “wtf is going on with the car?!” begin.
It’s a very uncomfortable experience even when it’s working properly.
Actual human drivers are intelligent. They can learn pretty quickly. The car is not Intelligent. Like a cow, every time it blinks, it’s a new day. (Apologies to cows.) It is only a product of an AI.
No
The other exception is the highway where NoA works great especially since I have the change lane parameter disabled.
Similar experience here unless I'm careful about the route I take since I know which routes FSD handles really well.
The other exception is the highway where NoA works great especially since I have the change lane parameter disabled.
FSD doesn't learn right away, it takes an update and even then it might not.
Adaptive FSD and a system that can be trained by the driver would be the ultimate solution.
Any good ways to discourage FSDb from "moving out of rightmost lane". Twice today it tried to move into the left lane cutting off a car that was about to pass me. I managed to stop it the second time, but it probably startled the other driver. Oddly, it only seems to happen on one road that has a lot of at grade entrances and exits.
I agree. My family tells me the same thing. But over time one could become desensitized after miles of FSDb driving. It's amazing how the brain/body adapts.
EVNow
Well-Known Member
Traffic conditions play a big part. When the traffic is not that heavy, FSDb does well. But the traffic is heavy so you have to use judgement and experience on when to make lane changes opportunistically etc it won't do well.
I suspect it depends on the bumps and their markings. Here in Albuquerque there is a predominant style of mark (sort of wide diagonal white stripes) for which the detection rate (based solely on slowing down) of speed humps has run about 90%. For the other types of marking, and unmarked, detection is zero.
Those have both stayed the same across many versions now, including the latest.
I have not once noticed a visualization of any of them.
I'm not denying the likely possibility that there was a version change that affected the handling of a population important on your drives, but the change was far from universal.
I am not entirely clear on why anyone continues to drive on FSDb 10.X? I dutifully installed 10.69.25.2 but I haven't engaged FSDb with it once. Why would I? FSDb v11 is supposed to be a foundational rewrite of the vision stack and include a whole bunch of new neural networks. Supposedly Elon and a bunch of other employees have been driving around on it for a long time. So I don't know that any automated feedback from me "testing" 10.69.X would have any value whatsoever to Tesla, and the camera icon that gave me the ability to designate important feedback is gone. FSDb is certainly not a useful driving feature by any stretch of the imagination - I mean no where near standard AutoPilot and NoA, which I actually use. Frankly, driving around on FSDb is a lot of work and is stressful, so if I am not really contributing anything to the testing then why do it? Plus I've been through enough versions of FSDb in the last 16 months that I don't really get excited about 10ms reduction of latency on graph vector deterpolation in the latest version any more. Seriously, outside of posting on this forum, there's nothing useful to be done until v11 comes out in 2 weeks.
I find FSDb is much better than AutoPilot on two lane blacktop with the occasional traffic light. NoA and FSDb work well for about 300 miles of a 350 mile trip I take several times a month. FSDb remains not so great in small towns and on one 4 lane road. Probably need to disable FSDb before I get on that one 4 lane road since AP does work well there. Not sure v11 will be an improvement unless it somehow eliminates the bizarre "exiting rightmost lane" behavior in FSDb.
bhakan
MSLR MSM/Blk/19/FSD, Ord 6/10/21, Del 2/20/22
. That FSDb is testing when i am driving alone. Other passengers don't want to hear the screaming from the car as that is not a normal behavior passengers have accustomed to. Most of the times they are polite and say they would want to ride in my ICE car. Saves everyone the agony.Salt covered roads are definitely having issues with FSD. Lane markings are hard for a human to read, let alone a computer. I’ve noticed over the last week a lot of issues mainly due to the road conditions. Half snow covered roads are also an issue (no, really?), but now FSD detects such conditions and says “FSD degraded due to poor weather conditions”, or some such $hit. This only used to pop up in rain, now I see it when roads are half snow covered, or it’s snowing. So, I’d actually take that as an improvement, as it knows conditions have deteriorated, or will.
A new morning begets a new 25.2 snafu... I was in the left turn lane of a protected left turn, first in line, and red turn signal. I initiated FSD. Recently I've noticed a slight jerk forward with 0 velocity when initiating FSDb at intersection red lights but this time it jumped and began moving forward - I quickly disengaged. Thinking it was an anomaly, I tried initiating again and it jumped out again with forward movement until I disengaged.
Maybe the only good reason to run a current V10.x release is to have a baseline for comparison with future releases. Once I get a feel for the current release I try to avoid scenarios I know will have poor outcomes. But I agree there's a lot of stress and extra work nursing V10.x along. It's no rose garden.
I think that has been normal operation for a long time. It treats you engaging FSD as a signal that it is OK to proceed through the light/stop sign. (Chuck has run into that on his UPLT, he has to enable FSD far enough back from the stop sign or it just goes right through.)
FOOLSLFDRVNG
Member
I noticed them appear for earlier versions when it was a new feature and later disappear at Costco. They have no paint markings since it is a private lot and most people drive under 20mph.
While I agree FSDb has plenty of shortcomings that can sometimes frustrate me, overall my experience has been one of steady improvements with few regressions over the past year. Even in these latest 10.69.25.x releases I've seen subtle improvements. Really the key issues that frustrate me on a daily basis boil down to just a few:
1) Poor lane planning
2) Random blinkers
3) Terrible behavior on some narrow-ish two way neighborhood roads with no painted center line. My guess is it doesn't see/infer the centerline at all, and this causes a lot of bad behavior: phantom braking for normal oncoming traffic, hogging the center until way too late for courtesy with oncoming cars, etc.
However, in spite of all this, I often use it and often find it useful as an L2 ADAS in many circumstances, and I do often have lengthy and/or complex disengagement-free drives, and even some intervention-free drives.
1) Poor lane planning
2) Random blinkers
3) Terrible behavior on some narrow-ish two way neighborhood roads with no painted center line. My guess is it doesn't see/infer the centerline at all, and this causes a lot of bad behavior: phantom braking for normal oncoming traffic, hogging the center until way too late for courtesy with oncoming cars, etc.
However, in spite of all this, I often use it and often find it useful as an L2 ADAS in many circumstances, and I do often have lengthy and/or complex disengagement-free drives, and even some intervention-free drives.
Both are connected, as the planner still makes incorrect choices. The difference with the current version is that we can now start to see the planner's thinking process as indicated by blue shaded lane choices on visualization. When you're in a left lane, and a left turn lane appears, but the car needs to ignore it and continue in the thru lane, you can see it shade the current lane as if it was going to change lanes. This shows us the thinking the planner has when a new lane appears. The problem is that it sometimes signals its intent to stay in the existing lane, which is a mistake.
Unmarked roads are a tough one. I've read people complain about how it drives more towards the center when it should stay more towards the right. However, in residential unmarked roads, I've noticed that I (driving manually) and most other people tend to drive more center when there are no other cars around. I think that's just normal human behavior. The difference is that when we see an oncoming car up ahead, we do start to pull right to make room. In slower residential road speeds, my car does this behavior as well - pulling right to make room. I'm guessing that at higher speeds, it's not anticipating far enough in advance and waiting until much closer to the oncoming vehicle before starting to move right. This seems like something that can be adjusted in the neural nets in the future.
I've definitely noticed an improvement on slower residential unmarked roads where the car stays further to the right then on previous builds. Still a problem though since FSD will sometimes stop or slow to a crawl with an approaching car, even when FSD has moved the car further to the right when it sees the approaching car.
Last edited:
I'm wondering if this is a perceptual issue. The car might think it doesn't have enough room to maneuver past the other car with a safety buffer, so it slows or stops to allow the other car to handle the maneuver instead. A case of being overly cautious?
Similar threads
- Article
