I have found this to be a common problem for me with the current version. It was better in the previous versions, so hopefully a fix will come out soon.
-
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
FSDtester#1
Craves Electrolytes
At least 500 times over the last year for me. Good example of why to keep a foot near the go pedal and the brake. It is way better since last October, but not perfect.
EVNow
Well-Known Member
That is neither here, nor there.
The idea is more cars will drive more miles in more varied conditions.
There is no particular reason why first 100k cars will be very different from next 100k.
Last edited:
EVNow
Well-Known Member
Yes - best to keep the expectations low. Obviously its nowhere close to being fully autonomous - as I keep writing, it has to improve atleast 100x ...
That doesn't sound right. May be you are disengaging too fast ? Would be useful if you can upload videos for us to see ..
Yes - the reasons are mostly around how much risk you are willing to take (and how far do you let FSDb go without intervening). It looks indecisive, slow etc. If you intervene in every one of those occasions, obviously the experience will be negative. You have to slowly understand how it turns etc - hopefully first when there is no traffic around. You will start understanding how FSDb drives ...
I agree with this. Some behaviors take getting used to before you can trust it. Creeping at a busy intersection is very hard to trust since the jerk is strong and it feels like the car is going into the intersection. I expect the vast majority hit the brake needlessly - though understandable.
AlanSubie4Life
Efficiency Obsessed Member
If you don’t like turns now, you should’ve seen them one year ago!
They may have deemed it good enough, but we will see with the next releases. Hard to know until it becomes a simple flag, which most certainly would qualify as wide release.
Just too early to say, though it does seem likely that they think it is good enough, somehow. Who knows what their metric is, though.
I’m wrong about stuff all the time. I’ll admit when I am. So “winning” is possible! But… It’s not about winning or losing arguments. There is just the way things are, and establishing that truth. It’s important to be matter-of-fact about this, especially when it comes to FSD, which can be a matter of life or death. There’s a lot of nonsense about it that gets promulgated (pro and con).
Yeah, it seems like pre-FSD-Beta visualizations were shown before during reverse, so it might also be a hint of which stack is currently active. Indeed, it makes sense for FSD Beta with its objects and occupancy networks to be used for parking and improving from ultrasonic sensors. Although, were the other vehicle position predictions stable for you when reversing into a spot?
I've also noticed 10.69.3.1 reverting to the old visualizations for lane departure warning and plain TACC without Autosteer.
KArnold
Active Member
I don't believe there is a 2022.40 or 2022.44 version of FSDb yet. Those appear to be the regular software branch
Something is wrong here. If you could post a video of what you're describing that would be helpful.
DarkForest
10.69 made me a shill
Most of the time, yes. Smaller cars seem to be more stable on the visualization than larger vehicles as usual.
The visualizations of the other cars and what the USSs were saying didn’t agree. I’m guessing that’ll get better over time.
Another interesting thing is that one time when FSD booted up when I started a drive, it momentarily flickered the layout of when I had turned off the car (noticeable because the surroundings were significantly different). I’m not sure if it was just a coincidental cache thing, or if the intention to remember the last scene is there now (for the USS code or for the parking stack).
On the other hand, there might be a noticeable step up in richness of both auto labeling and the occupancy network when there's a more even distribution of Tesla's driving our roads versus the past higher concentration in California.
FSD can't decide which path to take. It's likely calculating similar probabilities to both paths. I assume you were using navigation for this test? Even with navigation it can fail.
The logic is suppose to piece together known sections of roadway with the intermediate unknown section of roadway - in this case the straight-ahead dead-end. That's when FSD dithers left/right. They probably don't have enough of those scenarios to train the network to piece together a valid driving sequence.
This is an example of why FSD needs an almost endless set of real-world training data. Edge cases are key. And as all these edge cases are introduced it will be a can of worms to train, test, and release versions without adding regressions.
Last edited:
Here in New England, there's no shortage of edge cases
We just call them roads. Not sure if Elon has ever visited the area, but I'm pretty sure he'd change his tune on the readiness of FSDb if he saw the quantity of weird intersections here. The one in question isn't even all that complicated. Seriously, get the AP team here and get our roads into the sims ASAP.Here in New England, there's no shortage of edge cases
The old five nine reliability might be more like five sevens for FSD level 2 and that might be tolerable as long as FSD can fail gracefully with driver takeover versus this herky jerky uncertainty scaring and upsetting pedestrians and drivers on the road. However it might substantially drop demand at a $15k asking price.
Random detour this morning, thought it handled it pretty well (Massachusetts)
sleepydoc
Well-Known Member
I’ve been on FSDb for about a year and there’s been a lot of improvement in that time spattered with some regression.
I use FSD for the majority of my drives and I’d say I have a 90+% success rate. As others have noted, it used to be a remarkable event to have a 0 intervention drive whereas now it’s almost the norm for me. I drive to work on a mixture of residential, trunk/feeder roads and county roads with stoplights and turns. My trip home is almost always intervention free. On the way there there are a couple areas that can be dicey.
Having said that, I’ll also note that Tesla software can be aggravatingly inconsistent. Before I got FSD I had a ton of problems with phantom braking, on the order of every 10 minutes. Meanwhile, others claim never to have any problems. Why is that? Were they just unobservant? Different roads? Different expectations? Different definitions of ‘good?’
Even for me I’ll have times where FSD struggles when it handled the same section of road seamlessly the day before.
Similar threads
- Article
