Looks like it is still not completely fixed. For dirty Tesla - it worked near his home.
To me it sounds like an unhandled exception in planning code.
-
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 10.12
- Thread starter EVNow
- Start date
To me it sounds like an unhandled exception in planning code.
Regarding the left turns: supposedly the new version will anticipate the upcoming traffic gap and begin to move for the turn opportunity, like a human driver does, without waiting too long.
If done properly, this would be an excellent behavior - but it's different from the prior FSD beta behavior. So as a user, you'd have to build trust that it is only preparing/anticipating the turn, and not really going to turn into the path of the last oncoming car. It could understandably be difficult to go along with this, if you don't trust that the car knows what it's doing.
I'm absolutely not arguing with you. I don't have FSD beta yet myself, but this is one of the behaviors that I've really been hoping they would achieve. I'm simply asking exactly what the car is doing, vs your understandable concern that it could be trying to drive you into oncoming traffic.
In these close calls that you describe, is your car actually turning the wheel left as it moves forward, or is it simply beginning to move forward before the oncoming car has completely passed by? The first behavior seems wrong and could telegraph the wrong message to you and to the other driver. But the second behavior, while hard to trust at this point, is actually the correct advanced driving behavior. Maybe you could try to describe exactly what is happening in that moment, and how it's different from what a trustworthy human driver actually does in that exact situation.
This is one of the things that makes UPLs so difficult. They are nerve-wracking in everyday driving unless the traffic is very light, and people often take their biggest risks to time these turns and shoot the gap, because there aren't enough opportunities to wait for a nice big gap and a less-risky turn.
Again I'm not saying that you're incorrect to intervene. I'm saying that even if the FSD gets this exactly right, in this or in some future release, it will be hard for FSD beta users to learn to trust it. The car isn't talking to you and saying " okay I'll start to move so I can execute this turn right after this green Chevy goes by". Even if it were to put up one of those stupid tiny messages at the bottom of the screen, you can't be looking at that while you're supervising the car.
Last edited:
reporting back from my torture test mini route that usually has interventions every few hundred yards
Still tried to pull out in front of an oncoming car at the first T-junction (really a Y)
Then the right turn at the next T where it treats it like a stop, but there are no signs because we have right of way.
It still stopped, but it didn't wait forever and actually got round which is way better. Almost no wheel spinning either.
Next was the cross roads where we have right of way and turn left - still comes to a full stop and cuts the corner, but relatively smooth.
Then another T where we have right of way, still comes to full stop and see-saws through the corner.
Last in this section an actual yield where we turn right - it did this one really well, best its ever done it.
I thought it was significantly better than any of the previous versions. Still hogging the center of the road on unmarked roads, but pretty good otherwise.
That was until I hit the section which is a neighborhood road which has two lanes, one either way with a large central hiking trail area.
The drive out was uneventful, one or two unneeded slow downs, but not too bad.
Looping back was appalling with three "take over immediately" red hands of doom, one after another. No other traffic on the road, no other indication of issue.
I tagged all the issues and now the car is busy uploading gigs of data.
Based on that I doubt this release will be going wide.
The two steps forward are pretty good, but the one step back with those red hands of doom makes you never trust what its doing.
Still tried to pull out in front of an oncoming car at the first T-junction (really a Y)
Then the right turn at the next T where it treats it like a stop, but there are no signs because we have right of way.
It still stopped, but it didn't wait forever and actually got round which is way better. Almost no wheel spinning either.
Next was the cross roads where we have right of way and turn left - still comes to a full stop and cuts the corner, but relatively smooth.
Then another T where we have right of way, still comes to full stop and see-saws through the corner.
Last in this section an actual yield where we turn right - it did this one really well, best its ever done it.
I thought it was significantly better than any of the previous versions. Still hogging the center of the road on unmarked roads, but pretty good otherwise.
That was until I hit the section which is a neighborhood road which has two lanes, one either way with a large central hiking trail area.
The drive out was uneventful, one or two unneeded slow downs, but not too bad.
Looping back was appalling with three "take over immediately" red hands of doom, one after another. No other traffic on the road, no other indication of issue.
I tagged all the issues and now the car is busy uploading gigs of data.
Based on that I doubt this release will be going wide.
The two steps forward are pretty good, but the one step back with those red hands of doom makes you never trust what its doing.
The car turned the wheel completely and started to go almost causing a head on. There were multiple oncoming cars and it felt like it tried to go in between two of them even though they didn't have more than a car length between them and were going 40+. The car started making the turn and went in to their lane a bit before I emergency stopped and grabbed the yoke. I understand what you are saying about maybe moving forward a bit without turning the wheel but that isn't what happened. It was bad. Also, the car shouldn't turn the wheel like that unless it is making the turn because if you get rear ended it that could steer you right in to oncoming traffic. The car absolutely did the wrong thing and it was a very close call. I'm sure the other driver freaked out.
Thanks for the clear reply, even though it isn't what I was hoping to hear. It's extremely important for them to get this right if they're going to attempt UPLs in moderate to heavy traffic.
everydaychris
Active Member
CaptUAL
Member
Same for me. Could it be what model cars we have. Mine is a 2021 MSP. Or how much we use FSD BETA? Who knows…
clydeiii
Member
FOOLSLFDRVNG
Member
Let me get it first.
We ALL used to get them more quickly but since 8.0 Tesla has become more cautious. Seems with each update they send to fewer cars (except influencers which always get first) to test before going to most. So your odds of getting it sooner are generally going down since fewer people are getting it to begin with.
Also looks like 10.12.2 may have been pulled.
jabloomf1230
Minister of Silly Walks
Most likely. The only non-random component might be the firmware server that's used. I've sometimes been in the tier right after the Tesla employees/videographers and sometimes in the final push.
Or, maybe it has something to do with where you drive and whether you use FSD beta in certain situations where the FSD wants to get early test data.
jebinc
Well-Known Member
What an abomination. Seems like progress has been limited since first public release last fall? Not only that, but the period between releases is longer than back then and the quality of release (issues/bugs wise, seems to have increased.
But if I compared 10.12.2 to the first one I got last year (10.3), it is significantly improved in terms of usability. It obviously still not done, but 10.12.2 can easily navigate roads and routes that 10.3 would find impossible. Also, 10.12.2 is noticeably more comfortable to be driven by.
In terms of quality - remember the 10.3 -> 10.3.1 chaos?
The wider issue I think is that public testers are looking for noticeable changes from release to release, whereas timed releases are great for technical testers but rarely work with people who are only looking for the finished product. Gradual changes are rarely noticed as improvements, which is why I think they've switched to goal based release schedules.
I also think that the language used to describe these changes really needs to change, certainly before a wider release.
Public testers need a non-technical translation of the release notes. Realistically, most aren't going to notice a 10% improvement on how well some feature works, nor do they notice that frame rates have been increased but 1.2 fps, even if many technically oriented folks appreciate knowing that.
The release note changes really are significant technical improvements - they just don't directly translate into large changes on the road, which is how most of the public perceives progress.
But from 10.3 -> 10.12.2 lots and lots of usable improvements.
What continues to annoy me, is the pathetic lack of communication on pushing out the release (or not) to existing beta testers. It's also clear that there is preferential treatment for schillers in release availability. I don't quite understand what differentiates me, from the guy on youtube with respect to identifying issues with beta releases. Doesn't really seem like we are in a "beta testing group" does it?
There’s the beta testing group and then there’s the beta testing group beta
Even before the days of FSD beta, there was a cohort of people part of something called the "Early Access Program" that got new firmware to test before the rest of the fleet. Lately we've been calling this group the "shillers." Some of them do indeed shill, but most of them don't. I'd say the majority of them don't have youtube channels and generally keep silent about what they've received. The original EAP people signed NDAs and weren't allowed to discuss or post videos. More recently, Elon ditched the NDA since Tesla wasn't doing anything to enforce leaks.
Employees get new firmware first, followed by the EAP crowd, and then some subset of the safety score test cohort. Tesla is likely minimizing risk by releasing smaller-to-larger cohorts. They took a lot of stupid FUD heat for the 10.3 beta release. Tons of articles reporting that Tesla had to "recall" FSD due to catastrophic failures (false AEB on the highway). They patched it within a day via an OTA. Rather than deal with that kind of fallout again, they're doing the right thing by trying to fix severe issues while it's still on the employee cohort.
I personally don't need any advanced communication. It comes when it comes.
oh please. Don’t include everyone in that **** group.
Most of us call them YouTubers or influencers.
Absolutely nobody shills for Tesla I.e. nobody gets paid to say positive things about Tesla. There are a lot of fans of Tesla, though - who have made a lot of money from stock.
GlmnAlyAirCar
Active Member
FSDtester#1
Craves Electrolytes
Out of curiosity, how many tmc members in this thread have received 10.12.2?
If so, are you usually first to get the updates? I was an original 100% safety score and seemed to always get it as soon at released to the wider rollout. However over the last few updates, it seems really random...
If so, are you usually first to get the updates? I was an original 100% safety score and seemed to always get it as soon at released to the wider rollout. However over the last few updates, it seems really random...
Similar threads
