Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register
This site may earn commission on affiliate links.
I haven't seen anything regarding blinkers and lane selection
FSD Beta 10.12 notes have a bunch of improvements to extending the range of predictions especially for lanes, so potentially this includes predictions for lane connectivity. From my experiences with 10.11, the most common cases of unnecessary/wrong blinkers:
  1. Too much reliance on (inaccurate/incomplete) map data, e.g., car knows an upcoming intersection has a right-turn only lane not realizing the current right-most lane will have a new lane open up to become the right-turn only lane
  2. Misprediction of turn lanes, e.g., didn't have a good view of the road markings around a curve
I would guess Autopilot team is trying to have the the neural network predict turn lanes from even further with better accuracy to avoid needing to fall back to map data sooner. For example, if the neural network correctly predicts a new right-turn only lane will appear from 250m away, existing logic for "get out of right-turn only lane" that say kicks in at 200m/0.1 miles will no longer try to switch out.
 
  • Like
  • Informative
Reactions: trisk and McFlurri
I know this neighborhood very well - it's Bernal Heights in San Francisco, notorious for tight roads and two-lane streets that are barely wide enough for one car to pass at a time. Last time I had FSD Beta in late-2021 it did really poorly there, so it looks like 10.12.1 is a good deal better actually. Not sure what it stopped completely at the end though.
 
  • Like
Reactions: auger
I know this neighborhood very well - it's Bernal Heights in San Francisco, notorious for tight roads and two-lane streets that are barely wide enough for one car to pass at a time. Last time I had FSD Beta in late-2021 it did really poorly there, so it looks like 10.12.1 is a good deal better actually. Not sure what it stopped completely at the end though.
When the street disappears in the view ahead because of a convex road and hill down. It can't predict the drivable area ahead sufficiently from the visual input because the street abruptly goes away from the view.

This type of thing was also where phantom braking (a dip in the highway) was also common.

Fixing these needs mapping or 'faith'. Humans who drive the routes have confidence the roads are designed in such a way to have no weird obstructions, but you never know if there is a hidden car or obstacle.
 
  • Like
Reactions: mangrove79
Fixing these needs mapping or 'faith'. Humans who drive the routes have confidence the roads are designed in such a way to have no weird obstructions, but you never know if there is a hidden car or obstacle.
Yep, that’s why most humans slow down a fair bit when approaching such crests and drive straight over them (instead of at an angle). I would expect FSD to do a similar thing.

I remember how over 6 years ago AP1 used to veer left (sometimes onto oncoming traffic!!) when going over crests, which was scary as heck sometimes.
 
  • Like
Reactions: mangrove79
Off topic but the cabin camera is uber annoying on my MY without FSD. Trying to text and drive with AP on is unbearable with the beeping and red hands every 5 seconds.

And here I thought the apply torque nag was annoying…. This trumps it for sure 😅
Sorry for the off topic response.

I didn't realize that we can use the cabin camera instead of wheel torque for attention detection? Was that enabled recently for AP?
 
you can still see the indecision at some intersections and wheel jerking around in that video, just like the previous releases, even though he's 2 or 3x the speed. A few times it slows down to a crawl as well.
But he would still call that a zero intervention drive

Yeah. As we discussed, zero intervention drives can still have issues. Zero intervention does not mean the drive was perfect, just that the driver decided not take over.