FSD Beta 10.11

[Discoducky]

I'm hopeful for this release to solve my two blocking turns. Other than these two turns, all my other turns are varying degrees of performance, but not for safety.

This one: It doesn't see well enough in both directions for this unprotected left. Picture doesn't capture just how steep it is coming to a stop (4% grade). It comes to a complete stop at the black line from the stop sign. This means the cameras are completely occluded from seeing oncoming traffic from the left or right which is 55 mph. It starts creeping and halfway to the oncoming traffic lane it decides to go, but it still doesn't have a good view left or right. Proper behavior at this turn would be to come to a complete stop 10 feet short of the oncoming traffic lane and then start creeping when it is clear.



and this roundabout: It starts and stops (aka lots of jerk in motion and rapid steering wheel movements) taking at least 4 times as long to navigate to the 3rd exit, even without traffic.

 

[EVNow]

FSD Beta v10.11 Release Notes - Build 2022.4.5.15

1. Upgraded modeling of lane geometry from dense rasters (“bag of points”) to an autoregressive decoder that directly predicts and connects “vector space” lanes point by point using a transformer neural network. This enables us to predict crossing lanes, allows computationally cheaper and less error prone post-processing, and paves the way for predicting many other signals and their relationships jointly and end-to-end.
2. Use more accurate predictions of where vehicles are turning or merging to reduce unnecessary slowdowns for vehicles that will not cross our path.
3. Improved right-of-way understanding if the map is inaccurate or the car cannot follow the navigation. In particular, modeling intersection extents is now entirely based on network predictions and no longer uses map-based heuristics.
4. Improved the precision of VRU detections by 44.9%, dramatically reducing spurious false positive pedestrians and bicycles (especially around tar seams, skid marks, and rain drops). This was accomplished by increasing the data size of the next-gen autolabeler, training network parameters that were previously frozen, and modifying the network loss functions. We find that this decreases the incidence of VRU-related false slowdowns.
5. Reduced the predicted velocity error of very close-bymotorcycles, scooters, wheelchairs, and pedestrians by 63.6%. To do this, we introduced a new dataset of simulated adversarial high speed VI interactions. This update improves autopilot control around fast-moving and cutting-in VRUs.
6. Improved creeping profile with higher jerk when creeping staand ends.
7. Improved control for nearby obstacles by predicting continuous distance to static geometry with the general static obstacle network.
8. Reduced vehicle “parked” attribute error rate by 17%, achieve increasing the dataset size by 14%. Also improved brake light accuracy.
9. Improved clear-to-go scenario velocity error by 5% and highway scenario velocity error by 10%, achieved by tuning loss functions targeted at improving performance in difficult scenarios.
10. Improved detection and control for open car doors.
11. Improved smoothness through turns by using an optimization based approach to decide which road lines are irrelevant for c ? given lateral and longitudinal acceleration and jerk limits as w? vehicle kinematics.
12. Improved stability of the FSD Ul visualizations by optimizing ethernet data transfer pipeline by 15%.
13. Improved recall for vehicles directly behind ego, and improved precision for vehicle detection network.

 

[McFlurri]

Wait.. i thought the next release was 10.12....

 

[EVNow]

Wait.. i thought the next release was 10.12....

I think it started with Elon tweeting 10.12.

Anyway,I’ve requested mods to change the title. @Krash FYI.

 

[McFlurri]

I think it started with Elon tweeting 10.12.

Anyway,I’ve requested mods to change the title. @Krash FYI.

Ok so there wasnt actually a 10.11 then? And the next release is actually supposed to be 10.11 and not 10.12?

 

[MP3Mike]

Ok so there wasnt actually a 10.11 then? And the next release is actually supposed to be 10.11 and not 10.12?

We don't know. Those could be the release notes from 10.11 and it didn't work out well enough to release to early FSD beta participants, and that the next release will be 10.12 with those changes and more. (Or some of the changes could even have been pulled out of 10.12 if they were too problematic and couldn't be fixed quickly.)

Or Elon could have misspoken and the next release will be 10.11.

 

[GranularHail]

FSD Beta v10.11 Release Notes - Build 2022.4.5.15

1. Upgraded modeling of lane geometry from dense rasters (“bag of points”) to an autoregressive decoder that directly predicts and connects “vector space” lanes point by point using a transformer neural network. This enables us to predict crossing lanes, allows computationally cheaper and less error prone post-processing, and paves the way for predicting many other signals and their relationships jointly and end-to-end.
2. Use more accurate predictions of where vehicles are turning or merging to reduce unnecessary slowdowns for vehicles that will not cross our path.
3. Improved right-of-way understanding if the map is inaccurate or the car cannot follow the navigation. In particular, modeling intersection extents is now entirely based on network predictions and no longer uses map-based heuristics.
4. Improved the precision of VRU detections by 44.9%, dramatically reducing spurious false positive pedestrians and bicycles (especially around tar seams, skid marks, and rain drops). This was accomplished by increasing the data size of the next-gen autolabeler, training network parameters that were previously frozen, and modifying the network loss functions. We find that this decreases the incidence of VRU-related false slowdowns.
5. Reduced the predicted velocity error of very close-bymotorcycles, scooters, wheelchairs, and pedestrians by 63.6%. To do this, we introduced a new dataset of simulated adversarial high speed VI interactions. This update improves autopilot control around fast-moving and cutting-in VRUs.
6. Improved creeping profile with higher jerk when creeping staand ends.
7. Improved control for nearby obstacles by predicting continuous distance to static geometry with the general static obstacle network.
8. Reduced vehicle “parked” attribute error rate by 17%, achieve increasing the dataset size by 14%. Also improved brake light accuracy.
9. Improved clear-to-go scenario velocity error by 5% and highway scenario velocity error by 10%, achieved by tuning loss functions targeted at improving performance in difficult scenarios.
10. Improved detection and control for open car doors.
11. Improved smoothness through turns by using an optimization based approach to decide which road lines are irrelevant for c ? given lateral and longitudinal acceleration and jerk limits as w? vehicle kinematics.
12. Improved stability of the FSD Ul visualizations by optimizing ethernet data transfer pipeline by 15%.
13. Improved recall for vehicles directly behind ego, and improved precision for vehicle detection network.

@WholeMarsBlog posted another picture that shows the right part of the second page that was cut off on the first picture:

https://twitter.com/x/status/1503090278286036992

 

[macmee]

https://twitter.com/x/status/1500336079458627585

Really? I definitely have to intervene quite a lot in LA:

1. It often doesn't know which lane to go into for a turn
2. It often doesn't know when to go at a four way
3. it often gets confused at left turns in intersections with no turning signal
4. (most alarming) it often tries to turn right onto a road when there's an oncoming car!!!
5. (most annoying) in LA it prefers to stay in the right lane sometimes... even though there's cars parked in that lane on the side of the road. So when this happens I just have to turn FSD off completely or it immediately gets stuck.

 

[sleepydoc]

"more accurate predictions of where vehicles are turning or merging to reduce unnecessary slowdowns" - hopefully less phantom braking!

"modeling intersection extents is now entirely based on network predictions and no longer uses map-based heuristics" - sounds like they are moving more towards the car 'seeing' the intersection and away from depending on maps. another good sign!

"Improved smoothness through turns" - hopefully it there will be less whiplash in turns. Jerky turns make you look drunk on the road and scare passengers, so this will be another welcome improvement.

I'm getting excited for V11 but I need to remind myself that it's a major rebuild so there will likely be a lot of bugs in the initial release.

 

[EVNow]

@WholeMarsBlog posted another picture that shows the right part of the second page that was cut off on the first picture:

https://twitter.com/x/status/1503090278286036992

I think it’s all in there.

BTW, from Reddit. Not my transcription.

 

[EVNow]

Even if 10.11 comes out tomorrow, it will be the longest time between releases since 10 came out in Sep ...

 

[CaptUAL]

Even if 10.11 comes out tomorrow, it will be the longest time between releases since 10 came out in Sep ...

View attachment 780481

Now what might that mean? Great breakthrough and improvements? Or they have run out of options on trying to make the crappy cameras work, and there will be little to no change?

 

[EVNow]

Now what might that mean? Great breakthrough and improvements? Or they have run out of options on trying to make the crappy cameras work, and there will be little to no change?

OT - pls take your Lidar love elsewhere.

 

[jabloomf1230]

OT - pls take your Lidar love elsewhere.

Isn't that a tune by Golden Earring?

 

[CaptUAL]

OT - pls take your Lidar love elsewhere.

It will be interesting to see if there are any notable improvements, since it seems to have had very little progress lately.

 

[M109Rider]

It will be interesting to see if there are any notable improvements, since it seems to have had very little progress lately.

Hmm.
Are you hoping for improvements, or hoping for little progress ?

 

[CaptUAL]

Hmm.
Are you hoping for improvements, or hoping for little progress ?

Both.

 

[sleepydoc]

It will be interesting to see if there are any notable improvements, since it seems to have had very little progress lately.

I can think of a couple reasons for that -
1. They've been responding to NHTSA inquiries
2. They're working on a major rewrite, combining the previous AP software with the FSD software (i.e. it's more than just tweaking the existing algorithms.)
3. Further improvements are getting harder to achieve.
4. A combination of the above

you can draw your own conclusions.

 

[azred]

Really? I definitely have to intervene quite a lot in LA:

1. It often doesn't know which lane to go into for a turn
2. It often doesn't know when to go at a four way
3. it often gets confused at left turns in intersections with no turning signal
4. (most alarming) it often tries to turn right onto a road when there's an oncoming car!!!
5. (most annoying) in LA it prefers to stay in the right lane sometimes... even though there's cars parked in that lane on the side of the road. So when this happens I just have to turn FSD off completely or it immediately gets stuck.

Nothing new. Musk’s comments are frequently so far removed from reality that one has to wonder if he just makes up crap for the hell of it. Meanwhile I am busy developing a marketing plan for my two future robotaxis. Obviously I don’t have much time to complete preparations, right?

 

[FelixORiley]

For any of you like me hoping to be added…

https://twitter.com/x/status/1503171794932543494