-
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
too lazy to look it up, but a couple things would change that rule. One is that you have to obey the traffic control signs (e.g. right lane closed construction ahead). Secondly you have to obey enforcement officers directions. Not sure that you have to actually obey the construction workers. Unless they are wearing a sandwich board telling you to change lanes of course. So yeah, the bridge lanes would count under the control sign bit. I do believe you have to obey them even if they are there illegally.Does that apply when they close bridge lanes?
old pilot
Member
To be clear I was not suggesting that the Tesla autopilot drive a straight line on a curvy roadway. What I stated was that I saw no reason to cross the centerline in only one direction, because there is no savings in time or distance. I went on to say that by crossing the centerline several times in different directions you could save time and distance. I was in no way suggesting that the autopilot be programed to do this.
old pilot
Member
I notice several weeks ago when 10.12.2 was installed the distance to the virtual horizon displayed on the screen had increased from about 100 yards to what I would estimate was 300 to 400 yards. I took a long trip on interstates this past week with speeds in the 70 mph and above range and noticed the car was reacting sooner to cars ahead slowing. Do others notice this?
KArnold
Active Member
Is this implying "single stack"? Didn't think we were there yet.
Only applies to the Beta Stack. You just had some placebo or anecdotal NoA experience.
Just last week, my FSD beta braked late and hard for vehicles accelerating from a stop at a fresh green light. Caught me off guard because it probably didn't need to brake at all, but certainly could have started slowing earlier and more smoothly under the circumstances. As poorly as FSD beta generally does in terms of efficient/smooth driving in situations like this, I don't think FSD beta is using that type of prediction, but maybe the distance setting behaviors for standard autopilot changed with an update?
ETA: Oh, and then yesterday it did the same thing for a slow vehicle ahead (probably something like 50 vs my 65) when there was a vehicle coming up fast in the passing lane. FSD never signaled, so I don't think it was preparing to pass before it braked, and I think it could have slowed gently and changed lanes behind that passing vehicle if it had any concept of what was going on at all. If anything, the computer vision ought to be better at determining the speed of surrounding vehicles than I am (because I'm not that great at it), but that definitely seemed worse than I would have been.
Last edited:
jabloomf1230
Minister of Silly Walks
I have no idea whether the visualizations provide any hints to the driver of the behavior of FSD beta. However, I do know that the city streets "blue cars" represent higher priority objects in the NN prediction hierarchy. I'm guessing that pedestrians rank higher and the gray cars are lower than blue cars.
That said, there does not appear to be any obvious similar hierarchy for NOA. Nothing special happens in the visualization on highways. One would think that adjacent cars either driving faster, closer to the lane boundary or switching lanes would be higher priority, but my assumption is that the NOA algorithms are presently reactive and not predictive when it comes to nearby vehicles. Single stack may change that.
That said, there does not appear to be any obvious similar hierarchy for NOA. Nothing special happens in the visualization on highways. One would think that adjacent cars either driving faster, closer to the lane boundary or switching lanes would be higher priority, but my assumption is that the NOA algorithms are presently reactive and not predictive when it comes to nearby vehicles. Single stack may change that.
Has anyone received 10.12.2 AFTER being on the 2022.16.x branch? Or only from 2022.12.x or lower?
I received 2022.12.3.20 the day before a service visit after being on 2022.12.3.20 (yes, the exact same version). Prior to installing and looking at the version info, I asked the service center if they triggered it. I was told it had nothing to do with them, so I expected a newer version. In spite of what I was told, I doubt this information confirms anything you'd like to confirm, especially not re: Plaid gen.
andy.connor.e
Member
The most annoying part is not having FSD software and system software separate packages. I've been driving around with the dashcam bug for a while now and im getting pretty sick of unplugging my USB drive.
WilliamG
Hinge Fanatic
That bug was meeeeeega annoying. I’m happy without the BETA until that’s fixed on that branch of the updates.
everydaychris
Active Member
hybridbear
Active Member
Here is an insane behavior on 10.12 that only started a couple of weeks ago and is now 100% consistent. At this turn it now 100% of the time tries to turn into the far left opposing traffic lane. It even tried once to go on the left side of car waiting at the light. Today you can clearly see it has several lead cars to even help but it still tries for the far left lane. How Bat Shi$ Crazy is this?
Sunday morning I will get up early and try it if there is no traffic and let it play out to see what it does.
Sunday morning I will get up early and try it if there is no traffic and let it play out to see what it does.
old pilot
Member
I've been seeing similar behavior from FSD 10.12. After entering a destination and having a route displayed on the screen, when it comes to the first right turn it will sometime make the turn and other times go straight through the intersection. When the destination is entered and the blue line clearly shows the car should stay straight on the road, it will sometime try to get into a left or right turning lane at an intersection. I now keep both hands on the wheel when approaching intersections.
old pilot
Member
I've had it put on the right turn signal and then go straight so I'm not sure what to think.
It's not that surprising to me that it fails here as I used to encounter a left turn on a slope and angle where it couldn't clearly see the destination of the left turn and consistently wanted to turn into the more visible oncoming traffic lanes. This was fixed several versions ago for our intersection, but yours definitely seems tougher with different visual cues available:
The TeslaCam screenshot shows the center double yellow line isn't that clear but in this case, the lead vehicle driving there could be a hint if the neural network architecture actually shared data between moving objects and static / lane predictions. But the car still needs to drive correctly even without other vehicles around, so even ignoring Google Street View's camera positioned much higher than the front cameras in the Tesla, maybe the painted arrows on the ground could have fed into the lanes predictions?
It probably doesn't help that map data might be wrong where OpenStreetMap indicates there's 4 lanes total when there's actually 3 for oncoming and 2 for the left turn: Way: Market Street Northwest (25961138) | OpenStreetMap
Similar threads
