Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

FSD Beta 10.69

This site may earn commission on affiliate links.
His reasoning was that "we've made some pretty significant architectural improvements, so it's really going to be more than a 10.12 to 10.13 release… it might quality for 10.69."

For reference, 10.13 release notes:
  • Improved decision making for unprotected left turns using better estimation of ego's interaction with other objects through the maneuver.
  • Improved stopping pose while yielding for crossing objects at "Chuck Cook style" unprotected left turns by utilizing the median safety regions.
  • Made speed profile more comfortable when creeping for visibility, to allow for smoother stops when protecting for potentially occluded objects.
  • Enabled creeping for visibility at any intersection where objects might cross ego's path, regardless of presence of traffic controls.
  • Improved lane position error by 5% and lane recall by 12% with a full vector lanes neural network update. Information bottlenecks in the network architecture were abated by increasing the size of the per-camera feature extractors, video modules, and internals of the GPT attention modules.
  • Improved lane position error of crossing and merging lanes by 22% by adding long-range skip connections and a more powerful trunk to the network architecture.
  • Improved pedestrian and bicyclist velocity error by 17%, especially when ego is making a turn, by improving the onboard trajectory estimation used as input to the neural network.
  • Improved animal detection recall by 34% and decreased false positives by 8% by doubling the size of the auto-labeled training set.
  • Improved detection recall of far away crossing vehicles by 4% by tuning the loss function used during training and improving label quality.
  • Improved the "is parked" attribute for vehicles by 5% by adding 20% more examples to the training set.
  • Upgraded the occupancy network to detect dynamic objects and improved performance by adding a video module, tuning the loss function, and adding 37k new clips to the training set.
  • Reduced false slowdowns around crosswalks by better classification of pedestrians and bicyclists as not intending to interact with ego.
  • Reduced false lane changes for cones or blockages by preferring gentle offsetting in-lane where appropriate.
  • Improved in-lane positioning on wide residential roads.
  • Improved object future path prediction in scenarios with high yaw rate.
  • Improved speed limit sign accuracy on digital speed limits by 29%, on signs with difficult relevance by 23%, on 3-digit speeds by 39%, and on speed limit end signs by 62%. Neural network was trained with 84% more examples in the training set and with architectural changes which allocated more compute in the network head.
Unclear if there are additional neural network architecture improvements not listed or since those notes were created, but probably the most important one is the first one for "full vector lanes neural network update" where that's likely the bleeding edge of their current perception stack. Seems like once that architecture works for lanes, Tesla can further improve lane predictions with more auto-labeled training data as well as apply those architecture learnings to other predictions.
 
I'm still slightly concerned that the release notes don't mention any improvements in lane choices, i.e. picking the correct lane for an upcoming turn, as opposed to the current software which frequently seems to pick the worst lane for upcoming turns.

But it's possible that's handled in one of the more opaque descriptions like:
  • Improved lane position error of crossing and merging lanes by 22% by adding long-range skip connections and a more powerful trunk to the network architecture.
I live in hope.
 
I'm still slightly concerned that the release notes don't mention any improvements in lane choices
Some lane selection issues are newer / recent 10.12 regressions such as suddenly switching the wrong way just before an intersection probably from relying on vector space predictions introduced with 10.11. There's also issues such as unnecessarily switching out of a lane due to inaccurate map data for an upcoming intersection.

The former could be partially addressed in 10.69 by various network architecture improvements to better learn from the training data. The latter probably needs the network show it's more often correct than map data to remove the dependency, and that probably is always evaluated with each release and could happen with 10.69 for a later version to make the switch.
 
I'm still slightly concerned that the release notes don't mention any improvements in lane choices, i.e. picking the correct lane for an upcoming turn, as opposed to the current software which frequently seems to pick the worst lane for upcoming turns.

But it's possible that's handled in one of the more opaque descriptions like:
  • Improved lane position error of crossing and merging lanes by 22% by adding long-range skip connections and a more powerful trunk to the network architecture.
I live in hope.
I would hope that after a month or more of additional work they’re not just rolling out what amounts to 10.13.1
 
  • Like
Reactions: edseloh and dramsey
Good enough to skip 56 minor revisions but not good enough to call v11
I believe V11 is reserved for the "single stack", which integrates highway driving (currently the Navigate On Autopilot that all FSD subscribers have) with the beta Navigate On City Streets. The V13.69 software won't have this and will still switch to NoA on highways.
 
I believe V11 is reserved for the "single stack", which integrates highway driving (currently the Navigate On Autopilot that all FSD subscribers have) with the beta Navigate On City Streets. The V13.69 software won't have this and will still switch to NoA on highways.
interestingly, I've found I can stop it make stupid lane choices when driving on road with a >65mph limit by disabling NoA.
Here in Texas there are lots of non-highway roads that have a 65 or higher limit, so the old AP stack kicks in quite a lot and NoA is much less useful on roads that aren't divided highways.
 
  • Informative
Reactions: SalisburySam
I believe V11 is reserved for the "single stack", which integrates highway driving (currently the Navigate On Autopilot that all FSD subscribers have) with the beta Navigate On City Streets. The V13.69 software won't have this and will still switch to NoA on highways.
It's strange the single stack has been delayed because the new code isn't as good as the old code. Not confidence inspiring.
 
  • Like
Reactions: MTOman and jebinc
IIRC that was fixed in 2022.16.3. The previous 10.13 was a 2022.16.5 build so 10.69 should on whatever branch it's based on.
Yep, I think that is a good bet. I was on 2022.16.3 with a good safety score, and I think they were holding me back to get the next FSD beta release. I opted out of the FSD beta and almost immediately got offered 2022.20.8.
 
IIRC that was fixed in 2022.16.3. The previous 10.13 was a 2022.16.5 build so 10.69 should on whatever branch it's based on.
certainly seems to have fixed it for me. The only time the dash cam quits working for me now is when the inside of the hits >130F and it exceeds working temp of my SSD and it quits working. Have to unplug it to restart it.
 
It's strange the single stack has been delayed because the new code isn't as good as the old code. Not confidence inspiring.

I would guess that the current FSDb code on highways is exhibiting unacceptable levels of phantom braking and would be completely unacceptable to anyone currently using regular AP/NoA. So they would be correct to withhold a single-stack release until it performs better than current AP/NoA.

I also think there would be highway PB situations that overlap with local street PBs so we could see the number of PB occurrences drop on local roads once single stack is good enough for release.

Also likely is that single stack will feel great for some people and crappy for others :) This has been the case since AP1.
 
Yep, I think that is a good bet. I was on 2022.16.3 with a good safety score, and I think they were holding me back to get the next FSD beta release. I opted out of the FSD beta and almost immediately got offered 2022.20.8.
I'm opted into beta and stuck on 16.3 but my roomie just got the 20.8 and it doesn't seem worth opting out of beta for.. Hoping the new release expands beta users and I'm granted access.
 
I am currently on 2022.16.3 and just got an update for 2022.20.7, by updating will that eliminate my chances of getting into Beta 10.69? I'm afraid I will be ahead of the beta update branch and have to wait again.

I've been opted in for 2 months now with an avg of 96 with 3k miles in haha... I really want to ensure higher chances of getting into beta this time.
 
I am currently on 2022.16.3 and just got an update for 2022.20.7, by updating will that eliminate my chances of getting into Beta 10.69? I'm afraid I will be ahead of the beta update branch and have to wait again.

I've been opted in for 2 months now with an avg of 96 with 3k miles in haha... I really want to ensure higher chances of getting into beta this time.

I’m in virtually the same situation. 2022.20.7 just popped up for me after being on 2022.16.3 and having not budged even with all the 2022.20.x versions coming out. I had theorized that I had not been upgraded yet because I was opted into the beta, and that would have pushed us that were opted in past the branch point.

If everyone that is currently opted in but hasn’t received fsd beta yet is being pushed to 2022.20.7, that might indicate that they have moved the branch point of 10.69 to that version. I’m tempted to go for it if enough others who are opted in report that they're getting the update.


That might also be what this cryptic tweet is referring to (I have no idea who this guy is though):

 
  • Like
Reactions: Prairie