FSD Beta 10.12

[jobe]

SD Beta v10.12 Release Notes
- Upgraded decision making framework for unprotected left turns
with better modeling of objects' response to ego's actions by
adding more features that shape the go/no-go decision. This
increases robustness to noisy measurements while being more
sticky to decisions within a safety margin. The framework also
leverages median safe regions when necessary to maneuver across
large turns and accelerating harder through maneuvers when
required to safely exit the intersection.
- Improved creeping for visibility using more accurate lane
geometry and higher resolution occlusion detection.
- Reduced instances of attempting uncomfortable turns through
better integration with object future predictions during lane
selection.
- Upgraded planner to rely less on lanes to enable maneuvering
smoothly out of restricted space
- Increased safety of turns with crossing traffic by improving the
architecture of the lanes neural network which greatly boosted
recall and geometric accuracy of crossing lanes.
- Improved the recall and geometric accuracy of all lane predictions
by adding 180k video clips to the training set.
- Reduced traffic control related false slowdowns through better
integration with lane structure and improved behavior with respect
to yellow lights.
- Improved the geometric accuracy of road edge and line
predictions by adding a mixing/coupling layer with the generalized
static obstacle network.
- Improved geometric accuracy and understanding of visibility by
retraining the generalized static obstacle network with improved
data from the autolabeler and by adding 30k more videos clips.
- Improved recall of motorcycles, reduced velocity error of close-by
pedestrians and bicyclists, and reduced heading error of
pedestrians by adding new sim and autolabeled data to the training
set.
- Improved precision of the "is parked" attribute on vehicles by
adding 41k clips to the training set. Solved 48% of failure cases
captured by our telemetry of 10.11.
- Improved detection recall of far-away crossing objects by
regenerating the dataset with improved versions of the neural
networks used in the autolabeler which increased data quality.
- Improved offsetting behavior when maneuvering around cars with
open doors.
- Improved angular velocity and lane-centric velocity for non-VRU
objects by upgrading it into network predicted tasks.
- Improved comfort when lane changing behind vehicles with harsh
deceleration by tighter integration between lead vehicles future
motion estimate and planned lane change profile.
- Increased reliance on network-predicted acceleration for all
moving objects, previously only longitudinally relevant objects.
- Updated nearby vehicle assets with visualization indicating when
a vehicle has a door open.
- Improved system frame rate +1.8 frames per second by removing
three legacy neural networks.

 

[jobe]

(Guess I'm new so I can't delete or edit my previous post. Wanted to post with better formatting for readability. Forgive the double post.)

SD Beta v10.12 Release Notes

- Upgraded decision making framework for unprotected left turns with better modeling of objects' response to ego's actions by adding more features that shape the go/no-go decision. This increases robustness to noisy measurements while being more sticky to decisions within a safety margin. The framework also leverages median safe regions when necessary to maneuver across large turns and accelerating harder through maneuvers when required to safely exit the intersection.

- Improved creeping for visibility using more accurate lane geometry and higher resolution occlusion detection.

- Reduced instances of attempting uncomfortable turns through better integration with object future predictions during lane selection.

- Upgraded planner to rely less on lanes to enable maneuvering smoothly out of restricted space

- Increased safety of turns with crossing traffic by improving the architecture of the lanes neural network which greatly boosted recall and geometric accuracy of crossing lanes.

- Improved the recall and geometric accuracy of all lane predictions by adding 180k video clips to the training set.

- Reduced traffic control related false slowdowns through better integration with lane structure and improved behavior with respect to yellow lights.

- Improved the geometric accuracy of road edge and line predictions by adding a mixing/coupling layer with the generalized static obstacle network.

- Improved geometric accuracy and understanding of visibility by retraining the generalized static obstacle network with improved data from the autolabeler and by adding 30k more videos clips.

- Improved recall of motorcycles, reduced velocity error of close-by pedestrians and bicyclists, and reduced heading error of pedestrians by adding new sim and autolabeled data to the training set.

- Improved precision of the "is parked" attribute on vehicles by adding 41k clips to the training set. Solved 48% of failure cases captured by our telemetry of 10.11.

- Improved detection recall of far-away crossing objects by regenerating the dataset with improved versions of the neural networks used in the autolabeler which increased data quality.

- Improved offsetting behavior when maneuvering around cars with open doors.

- Improved angular velocity and lane-centric velocity for non-VRU objects by upgrading it into network predicted tasks.

- Improved comfort when lane changing behind vehicles with harsh deceleration by tighter integration between lead vehicles future motion estimate and planned lane change profile.

- Increased reliance on network-predicted acceleration for all moving objects, previously only longitudinally relevant objects.

- Updated nearby vehicle assets with visualization indicating when a vehicle has a door open.

- Improved system frame rate +1.8 frames per second by removing three legacy neural networks.

 

[mhan00]

Very curious to see how well the car uses medians to complete left turns.

 

[AlanSubie4Life]

“- Improved the geometric accuracy of road edge and line
predictions by adding a mixing/coupling layer with the generalized
static obstacle network.”

Never gonna hit a curb again.

 

[JHCCAZ]

Very curious to see how well the car uses medians to complete left turns.

I'm also curious to see if it knows not to do this in areas where it is not permitted. If the road and median design is not appropriately configured, then stopping at the median makes things more dangerous for everyone.

 

[JHCCAZ]

(Guess I'm new so I can't delete or edit my previous post. Wanted to post with better formatting for readability. Forgive the double post.)

SD Beta v10.12 Release Notes

- Upgraded decision making framework for unprotected left turns with better modeling of objects' response to ego's actions by adding more features that shape the go/no-go decision. This increases robustness to noisy measurements while being more sticky to decisions within a safety margin. The framework also leverages median safe regions when necessary to maneuver across large turns and accelerating harder through maneuvers when required to safely exit the intersection.

- Improved creeping for visibility using more accurate lane geometry and higher resolution occlusion detection.

- Reduced instances of attempting uncomfortable turns through better integration with object future predictions during lane selection.

- Upgraded planner to rely less on lanes to enable maneuvering smoothly out of restricted space

- Increased safety of turns with crossing traffic by improving the architecture of the lanes neural network which greatly boosted recall and geometric accuracy of crossing lanes.

- Improved the recall and geometric accuracy of all lane predictions by adding 180k video clips to the training set.

- Reduced traffic control related false slowdowns through better integration with lane structure and improved behavior with respect to yellow lights.

- Improved the geometric accuracy of road edge and line predictions by adding a mixing/coupling layer with the generalized static obstacle network.

- Improved geometric accuracy and understanding of visibility by retraining the generalized static obstacle network with improved data from the autolabeler and by adding 30k more videos clips.

- Improved recall of motorcycles, reduced velocity error of close-by pedestrians and bicyclists, and reduced heading error of pedestrians by adding new sim and autolabeled data to the training set.

- Improved precision of the "is parked" attribute on vehicles by adding 41k clips to the training set. Solved 48% of failure cases captured by our telemetry of 10.11.

- Improved detection recall of far-away crossing objects by regenerating the dataset with improved versions of the neural networks used in the autolabeler which increased data quality.

- Improved offsetting behavior when maneuvering around cars with open doors.

- Improved angular velocity and lane-centric velocity for non-VRU objects by upgrading it into network predicted tasks.

- Improved comfort when lane changing behind vehicles with harsh deceleration by tighter integration between lead vehicles future motion estimate and planned lane change profile.

- Increased reliance on network-predicted acceleration for all moving objects, previously only longitudinally relevant objects.

- Updated nearby vehicle assets with visualization indicating when a vehicle has a door open.

- Improved system frame rate +1.8 frames per second by removing three legacy neural networks.

These all sound very encouraging, but as always we'll have to see how they work out for various people and locations.

Our favorite UPL tester Chuck Cook has been out of the FSD beta action with his car in the shop after the very unfortunate accident he had (tragic situation but nothing to do with FSD/Autopilot nor any fault on his part). Though I'm not a Twitter user, I'm able to see his recent update tweet indicating that the shop has the parts and his car should be getting back together soon.

 

[drtimhill]

(Guess I'm new so I can't delete or edit my previous post. Wanted to post with better formatting for readability. Forgive the double post.)

SD Beta v10.12 Release Notes

- Upgraded decision making framework for unprotected left turns with better modeling of objects' response to ego's actions by adding more features that shape the go/no-go decision. This increases robustness to noisy measurements while being more sticky to decisions within a safety margin. The framework also leverages median safe regions when necessary to maneuver across large turns and accelerating harder through maneuvers when required to safely exit the intersection.

- Improved creeping for visibility using more accurate lane geometry and higher resolution occlusion detection.

- Reduced instances of attempting uncomfortable turns through better integration with object future predictions during lane selection.

- Upgraded planner to rely less on lanes to enable maneuvering smoothly out of restricted space

- Increased safety of turns with crossing traffic by improving the architecture of the lanes neural network which greatly boosted recall and geometric accuracy of crossing lanes.

- Improved the recall and geometric accuracy of all lane predictions by adding 180k video clips to the training set.

- Reduced traffic control related false slowdowns through better integration with lane structure and improved behavior with respect to yellow lights.

- Improved the geometric accuracy of road edge and line predictions by adding a mixing/coupling layer with the generalized static obstacle network.

- Improved geometric accuracy and understanding of visibility by retraining the generalized static obstacle network with improved data from the autolabeler and by adding 30k more videos clips.

- Improved recall of motorcycles, reduced velocity error of close-by pedestrians and bicyclists, and reduced heading error of pedestrians by adding new sim and autolabeled data to the training set.

- Improved precision of the "is parked" attribute on vehicles by adding 41k clips to the training set. Solved 48% of failure cases captured by our telemetry of 10.11.

- Improved detection recall of far-away crossing objects by regenerating the dataset with improved versions of the neural networks used in the autolabeler which increased data quality.

- Improved offsetting behavior when maneuvering around cars with open doors.

- Improved angular velocity and lane-centric velocity for non-VRU objects by upgrading it into network predicted tasks.

- Improved comfort when lane changing behind vehicles with harsh deceleration by tighter integration between lead vehicles future motion estimate and planned lane change profile.

- Increased reliance on network-predicted acceleration for all moving objects, previously only longitudinally relevant objects.

- Updated nearby vehicle assets with visualization indicating when a vehicle has a door open.

- Improved system frame rate +1.8 frames per second by removing three legacy neural networks.

wow that is a LOT of changes compared to most beta releases .. will be interesting to see which bits get better and which bits get worse

 

[drtimhill]

These all sound very encouraging, but as always we'll have to see how they work out for various people and locations.

Our favorite UPL tester Chuck Cook has been out of the FSD beta action with his car in the shop after the very unfortunate accident he had (tragic situation but nothing to do with FSD/Autopilot nor any fault on his part). Though I'm not a Twitter user, I'm able to see his recent update tweet indicating that the shop has the parts and his car should be getting back together soon.

wow I didn't hear about that I hope he is ok.

 

[JHCCAZ]

wow I didn't hear about that I hope he is ok.

He posted a rather sad YouTube video explaining things. He was on a road trip on the highway and a modified motorcycle or motor trike lost a wheel and careened out of control. I think it struck the rear quarter of his Tesla but it's not even clear the rider would have survived the event had Chuck's Tesla not even been nearby. Just the kind of thing we all dread when we head out on the highway.

 

[mhan00]

wow I didn't hear about that I hope he is ok.

Yeah, a guy driving one of those three wheeled motorcycles blew a tire/lost a wheel and careened across the highway directly into Chuck’s rear quarter panel. Unfortunately the cyclist died. Nothing Chuck could have done (the dash cam vid shows that) and he is physically fine, but being involved in an accident where someone loses their life in any capacity has to be traumatizing to some degree.

Edit - sheesh, I got ninja’ed, lol.

 

[EVNow]

(Guess I'm new so I can't delete or edit my previous post. Wanted to post with better formatting for readability. Forgive the double post.)

SD Beta v10.12 Release Notes

- Upgraded decision making framework for unprotected left turns with better modeling of objects' response to ego's actions by adding more features that shape the go/no-go decision. This increases robustness to noisy measurements while being more sticky to decisions within a safety margin. The framework also leverages median safe regions when necessary to maneuver across large turns and accelerating harder through maneuvers when required to safely exit the intersection.

- Improved creeping for visibility using more accurate lane geometry and higher resolution occlusion detection.

- Reduced instances of attempting uncomfortable turns through better integration with object future predictions during lane selection.

- Upgraded planner to rely less on lanes to enable maneuvering smoothly out of restricted space

- Increased safety of turns with crossing traffic by improving the architecture of the lanes neural network which greatly boosted recall and geometric accuracy of crossing lanes.

- Improved the recall and geometric accuracy of all lane predictions by adding 180k video clips to the training set.

- Reduced traffic control related false slowdowns through better integration with lane structure and improved behavior with respect to yellow lights.

- Improved the geometric accuracy of road edge and line predictions by adding a mixing/coupling layer with the generalized static obstacle network.

- Improved geometric accuracy and understanding of visibility by retraining the generalized static obstacle network with improved data from the autolabeler and by adding 30k more videos clips.

- Improved recall of motorcycles, reduced velocity error of close-by pedestrians and bicyclists, and reduced heading error of pedestrians by adding new sim and autolabeled data to the training set.

- Improved precision of the "is parked" attribute on vehicles by adding 41k clips to the training set. Solved 48% of failure cases captured by our telemetry of 10.11.

- Improved detection recall of far-away crossing objects by regenerating the dataset with improved versions of the neural networks used in the autolabeler which increased data quality.

- Improved offsetting behavior when maneuvering around cars with open doors.

- Improved angular velocity and lane-centric velocity for non-VRU objects by upgrading it into network predicted tasks.

- Improved comfort when lane changing behind vehicles with harsh deceleration by tighter integration between lead vehicles future motion estimate and planned lane change profile.

- Increased reliance on network-predicted acceleration for all moving objects, previously only longitudinally relevant objects.

- Updated nearby vehicle assets with visualization indicating when a vehicle has a door open.

- Improved system frame rate +1.8 frames per second by removing three legacy neural networks.

So a lot of changes related to ULT and lane geometry. They are using a lot of video clips to train, moving to NN etc.

May be this does pull them out of a local maximum and helps get to a newer level.

 

[EVNow]

I think s/he meant that Omar doesn’t get totally smooth, no intervention trips, either. He just finds the token one to post and brown nose Elon.

Considering the number of trips he posts, no way he is selecting a small % that are zero disengagement.

I’ve given better explanation for his zero engagement trips.

Afterall even dirty Tesla has posted no disengagement trips and I’d have zero disengagement trips too if we didn’t have so many roundabouts here everywhere.

Actually, if we all ignore driving etiquette and don’t care about fellow drivers, we can all have zero disengagement drives.

 

[momo3605]

Considering the number of trips he posts, no way he is selecting a small % that are zero disengagement.

I’ve given better explanation for his zero engagement trips.

Afterall even dirty Tesla has posted no disengagement trips and I’d have zero disengagement trips too if we didn’t have so many roundabouts here everywhere.

Actually, if we all ignore driving etiquette and don’t care about fellow drivers, we can all have zero disengagement drives.

I don’t know…I’m also in the SF Bay Area. Supposedly “over fit”. But I can’t do anything close to Omar. If I pick very specific routes with very light traffic yes, it can do 0 disengagements, but most of the time there will be at least one or two times in the drive that the car will try to jump into the wrong lane before a turn(even with clear lane markings) or try take an unprotected turn while it’s still unsafe and cars are coming.

 

[EVNow]

I don’t know…I’m also in the SF Bay Area. Supposedly “over fit”. But I can’t do anything close to Omar. If I pick very specific routes with very light traffic yes, it can do 0 disengagements, but most of the time there will be at least one or two times in the drive that the car will try to jump into the wrong lane before a turn(even with clear lane markings) or try take an unprotected turn while it’s still unsafe and cars are coming.

I guess Omar runs routes he knows can be done with no disengagements. He also pushes limits in terms of letting the car keep going until he has to disengage. He doesn’t disengage for wrong lanes or turns. He lets the car reroute.

 

[ammulder]

I don’t know…I’m also in the SF Bay Area. Supposedly “over fit”. But I can’t do anything close to Omar. If I pick very specific routes with very light traffic yes, it can do 0 disengagements, but most of the time there will be at least one or two times in the drive that the car will try to jump into the wrong lane before a turn(even with clear lane markings) or try take an unprotected turn while it’s still unsafe and cars are coming.

I'm not sure zero disengagements is the right metric. I'd be pleased with zero unexpected disengagements and zero unexpected speed changes, I guess.

For now, I'm happy to let FSD Beta do its thing and I'll take over in the places I know I will want to. For instance, there's a stop sign where I need to turn right and immediately cross two lanes to the left to make a left turn... within maybe 1/8 mile? FSD so far isn't aggressive enough. It doesn't pull away from the stop fast enough to blend with traffic speed, so it can't make the lane changes if there are any other cars around. I don't mind taking over for that bit and then letting it go again afterward. I don't mind taking over for some left turns where I want to go quickly and not wait for an overly large gap in oncoming traffic (usually when there's only the one lane my way so cars are piling up behind me while I wait to turn). We also have a couple super-narrow bridges over the Delaware river. I'm sort of impressed that FSD can handle them at all (there are no lane markings and they're narrow enough that people fold mirrors to avoid tagging oncoming traffic) -- but it will slow to maybe 3-5mph at the time an oncoming car actually passes, which doesn't work for me.

It also hasn't quite nailed max speeds for me yet. There are plenty of rural/suburban roads where the speed limit goes up and down, and sometimes it behaves pretty oddly; the speed limit will go up 5 and the set speed will go down 5 at the same time. It doesn't seem to me like the set speed should *ever* go down when the speed limit goes up. There's a 1/4 mile stretch where the map data incorrectly drops the speed limit by 10 and it doesn't recover until the next speed limit sign -- I just wish it would remember my set speed at that location like it can remember certain other location-based stuff.

Still, I feel like I'm in "autopilot jail" on our new X because I don't have the beta yet. Really hoping they'll add it in the 10.12 cycle!

 

[Reddy Kilowatt]

I swear the second to last release note one was a direct result of the incident I had. I was ready to put a guy through his Car door. He swung his door open, jumped out, never looked and was focused on yelling at someone on his cell phone. Closest call I ever had, and instantly swung out of the way, no time to think. . I sent the clip to Tesla and reported it as a bug.

I had posted this on 10.11 a while back, to remain Vigilant when using the system. we know how Tesla is trying to see how close the car's rear view mirror to parked car mirrors.

Any Teslafi hits yet??

 

[diplomat33]

I'm not sure zero disengagements is the right metric. I'd be pleased with zero unexpected disengagements and zero unexpected speed changes, I guess.

Zero disengagements just means that the driver never disengaged. It does not necessarily mean that the drive was good. Assuming zero accidents, a zero disengagement drive could have still lots of issues, like near misses, jerky maneuvers, bad lane change, phantom braking, etc...

IMO, there are 4 stages to autonomous driving:
1) Basic driving
2) Safe driving
3) Smooth driving
4) Defensive driving

 

[sleepydoc]

The biggest single flaw with 11.2 has been lane selection and the propensity to dive into turn lanes without going straight. There is a conspicuous absence of anything regarding lane selection in the release notes Omar posted. Really hoping that’s improved since it’s made FSD almost unusable at times. (The turn signals are a huge issue, too but I suspect that's related to the issues with lane selection)

 

[sleepydoc]

Zero disengagements just means that the driver never disengaged. It does not necessarily mean that the drive was good. Assuming zero accidents, a zero disengagement drive could have still lots of issues, like near misses, jerky maneuvers, bad lane change, phantom braking, etc...

IMO, there are 4 stages to autonomous driving:
1) Basic driving
2) Safe driving
3) Smooth driving
4) Defensive driving

you forgot to add courteous driving. I suppose that could be lumped into safe driving, but there are plenty of time FSD will do something that’s not necessarily unsafe but is enough to aggravate other drivers and/or make them do something unsafe.

 

[diplomat33]

you forgot to add courteous driving. I suppose that could be lumped into safe driving, but there are plenty of time FSD will do something that’s not necessarily unsafe but is enough to aggravate other drivers and/or make them do something unsafe.

Thanks. Yeah, courteous driving is not something that will necessarily be recorded as a disengagement but it is important too.