Wiki FSD Beta 10.4

[K5TRX]

I don't know how useful it is to report duplicate bugs.

In beta testing, you isolate bugs by developing sufficient data points to see trends and isolate causes. It is by seeing duplicate bugs that you can see that a seemingly random bug is not random.

For instance, right now we all need to be reporting every Phantom Braking error. I am seeing it every time I drive to work at exactly the same places. For me, these seem to be the same shadows every morning and same repaired potholes at night.

One person reporting one bug once may be almost impossible to isolate. And probably should be a lower priority unless a critical.

Duplicate bug reports about the same bug in similar situations may make the causes more clear.

 

[TravelFree]

Waiting on 10.5 now before spending more time testing 10.4.

What I saw was near perfect in 10.4 is speed bumps, speed limit tracking, red and green light detection and control. Yellow light decision control was hit of miss ( no pun jokes please), stop sign stopping at the line was 50:50. Slow speed steering continues to be erratic. Pedestrian stoppage was too cautious. stopping in the middle of a right or left turn was failure requiring 8 out of 10 interventions. Construction zone impossible to really dangerous.

No point in continuing more testing drives until the next update.


I did not do the camera recalibration that was recently suggested but I may do that later this week after 10.5 arrives and I have a chance to do some driving on my test route first.

 

[TravelFree]

In beta testing, you isolate bugs by developing sufficient data points to see trends and isolate causes. It is by seeing duplicate bugs that you can see that a seemingly random bug is not random.

For instance, right now we all need to be reporting every Phantom Braking error. I am seeing it every time I drive to work at exactly the same places. For me, these seem to be the same shadows every morning and same repaired potholes at night.

One person reporting one bug once may be almost impossible to isolate. And probably should be a lower priority unless a critical.

Duplicate bug reports about the same bug in similar situations may make the causes more clear.

Very well stated.
On common traffic control detections like stop signs, I usually run those 50-100 times to see any pattern of pass or failure. But on those random phantom braking incidents, just report every one because I have yet to see any pattern except for one: When a car's brake lights come on in an adjacent lane and has no intention of coming into my lane. This is common on the highway going past an entrance ramp. All the others happen without obvious pattern. Might be shadows, but there are plenty and I haven't seen that here.

On speed bumps, I have a test route that has all types of speed bumps from no warning or markings to well marked. It does excellent on all of them. 10.3 it would miss all of them as in no detection. There is even a dip in the road at an intersection with no stop signs. Purpose is to channel heavy rain water to storm drain. The car now even slows for that dip to 15 mph. It's opposite a bump.

 

[TravelFree]

A friend posted this in another private forum:

FSD Beta 10.5 is expected to be released tonight. Here are the release notes:

• Improved VRU (pedestrians, bicyclists, motorcycles) crossing velocity error by 20% from improved quality in our auto-labeling.
• Improved static world predictions (road lines, edges, and lane connectivity) by up to 13% using a new static world auto-labeler and adding 165K auto-labeled videos.
• Improved cone and sign detections by upreving the generalized static object network with 15K more video clips and adjusting oversampling and overweighting strategies (+4.5% precision, +10.4% recall).
• Improved cut-in detection network by 5.5% to help reduce false slowdowns.
• Enabled “emergency collision avoidance maneuvering” in shadow mode.
• Enabled behavior to lane change away from merges when safe to do so.
• Improved merge object detection recall by using multi-modal object prediction at intersections.
• Improved control for merges by increasing smoothness of arrival time constraints and considering possible merging objects beyond visibility.
• Improved lane changes by allowing larger deceleration limit in short-deadline situations.
• Improved lateral control for creeping forward to get more visibility.
• Improved modeling of road boundaries on high curvature roads for finer maneuvers.
• Improved logic to stay on-route and avoid unnecessary detours/rerouting.

 

[Mardak]

right now we all need to be reporting every Phantom Braking error

My usage of video report button has mostly focused on incorrect understanding of the road as I believe that's the most fundamental aspect to get right and other issues are generally more comfort or preference that I can override by pressing the accelerator. So my reported issues mostly are for not understanding odd intersections or lanes/road shape, and with the 10.5 release notes, looks like it's also one of Tesla's higher priorities with adding autolabeling capabilities to get "⁠Improved static world predictions."

Tesla's first autolabeling focus (at least according to release notes) was for Vulnerable Road Users like pedestrians, and that makes sense to prioritize human safety and avoid hurting (or confusing) people walking or riding on the road.

Tesla is collecting a lot of data including Phantom Braking and is still improving the training infrastructure and neural network capabilities, so they do focus their engineering efforts on specific types of problem areas. Phantom Braking can be caused by many different aspects, so that also means some cases are resolved with better VRU and static world autolabeling, e.g., better biker recall means not slamming brakes for a non-existent/phantom human.

 

[jebinc]

https://www.reddit.com/r/teslamotors/comments/qymvvc

 

[d21mike]

A friend posted this in another private forum:

FSD Beta 10.5 is expected to be released tonight. Here are the release notes:

• Improved VRU (pedestrians, bicyclists, motorcycles) crossing velocity error by 20% from improved quality in our auto-labeling.
• Improved static world predictions (road lines, edges, and lane connectivity) by up to 13% using a new static world auto-labeler and adding 165K auto-labeled videos.
• Improved cone and sign detections by upreving the generalized static object network with 15K more video clips and adjusting oversampling and overweighting strategies (+4.5% precision, +10.4% recall).
• Improved cut-in detection network by 5.5% to help reduce false slowdowns.
• Enabled “emergency collision avoidance maneuvering” in shadow mode.
• Enabled behavior to lane change away from merges when safe to do so.
• Improved merge object detection recall by using multi-modal object prediction at intersections.
• Improved control for merges by increasing smoothness of arrival time constraints and considering possible merging objects beyond visibility.
• Improved lane changes by allowing larger deceleration limit in short-deadline situations.
• Improved lateral control for creeping forward to get more visibility.
• Improved modeling of road boundaries on high curvature roads for finer maneuvers.
• Improved logic to stay on-route and avoid unnecessary detours/rerouting.

All looks good. However, one of my biggest complaints at this point is completely unnecessary lane changes. They need option like NOA to approve lane changes not related to NAV.

 

[TravelFree]

All looks good. However, one of my biggest complaints at this point is completely unnecessary lane changes. They need option like NOA to approve lane changes not related to NAV.

I don't care for unnecessary lane changes either, but I think those represent an inconvenience. Of more concern to me is not making the right and left turns smoothly as opposed to FSD stopping in the oncoming traffic which is extremely dangerous. Priority should be focused on correcting real safety issues before those of missing a turn due to being in the wrong lane.

 

[lolder]

Did camera calibration today on 2018 M3 LR RWD. On a straight 4 lane divided highway it took 5 miles to enable AP/NOA and only another 5 to enable FSD. Car will still not make turns from my unmarked residential st. onto lane marked 2 lane thru st. Turned FSD off again to await improvement.

 

[d21mike]

I don't care for unnecessary lane changes either, but I think those represent an inconvenience. Of more concern to me is not making the right and left turns smoothly as opposed to FSD stopping in the oncoming traffic which is extremely dangerous. Priority should be focused on correcting real safety issues before those of missing a turn due to being in the wrong lane.

Agreed that your concern is by far the most important but could be much hard to solve (maybe mid next year). Having a option to approve lane changes like with FSD Freeway NOA could be much easier and help train the car for when and when not to change lanes for no reason. And of course changing lanes and then trying to get back into the correct lane to make a turn can be a safety issue. In my case < 1/2 mile from home it will move into the left lane before a light that is just maybe 200 yards before needing to get back into the right lane to make a right turn. And by the time to get back in the right lane the lane can be full of cars that it needs to get in between. Generally just need to take it out of FSD Beta.

 

[EVNow]

Of more concern to me is not making the right and left turns smoothly as opposed to FSD stopping in the oncoming traffic which is extremely dangerous.

Agreed that your concern is by far the most important but could be much hard to solve (maybe mid next year).

They need those clips. They should be able to figure out why the car stops in the middle of the lane - this would be a planner issue that they should be able to solve - I don't think it would be that hard.

 

[powertoold]

They need those clips. They should be able to figure out why the car stops in the middle of the lane - this would be a planner issue that they should be able to solve - I don't think it would be that hard.

Solving any single issue is easy, but solving it for all locales and intersection geometries is difficult.

 

[d21mike]

They need those clips. They should be able to figure out why the car stops in the middle of the lane - this would be a planner issue that they should be able to solve - I don't think it would be that hard.

Well it has be > 1 year of 2000 FSD Beta testers before we were even invited and they have not solved it yet. I have right and left turns that fail a lot and have reported them multiple times. It must be hard because these are kind of important. Two right turns I use all the time. They have a right turn lane where one of them is well marked with an arrow but the car will try to make the right turn from the regular lane and be very jerky about it so I generally just disengage now since it has failed so many times and I have reported. A left turn with slight up hill which I understand my be a little more difficult fails every time by just disengaging on it's own. I thought it was just me holding on too tight but I do not think it is me. Now I just disengage myself. So, yes turns need a lot of work and after > a year seems like pretty hard to solve. I assume they will fix it eventually. Every new version I will continue to test these multiple times to see if there is an improvement. BTW. 10.4 did improve over 10.3.1 with now seeing the Gates coming and going from my community. Coming back works well but going out could be smoother since the gate opens at me.

 

[Twiglett]

I did a drive yesterday on a new route. The car did mostly really well. Still some panic wheel twirling on a few corners but no interventions needed.
Felt like a day two learner now rather than the day one learner I get on my usual routes.
Not had any change in version etc, but yesterday it decided on the drive home that it would navigate down my driveway which was new.
While it may have set the nav to go down the driveway (we have a longish driveway) it overshot by 10-15 feet and gave up
overall, better than normal

 

[d21mike]

Solving any single issue is easy, but solving it for all locales and intersection geometries is difficult.

Agreed, and this is why I suggested a fairly simply option to require the driver to authorized lane changes for reasons other then lane changes for NAV. I have seen a lot of complaints on here and on video of lane changes for no good reason. They already do it for FSD NOA on Freeways. Seems pretty easy to add that to FSD Beta on surface streets. You would even thing it would really help in the training if a lot of drivers are saying which lane changes are good and which are bad and at the same time make the FSD Beta much more enjoyable to use.

 

[powertoold]

Agreed, and this is why I suggested a fairly simply option to require the driver to authorized lane changes for reasons other then lane changes for NAV. I have seen a lot of complaints on here and on video of lane changes for no good reason. They already do it for FSD NOA on Freeways. Seems pretty easy to add that to FSD Beta on surface streets. You would even thing it would really help in the training if a lot of drivers are saying which lane changes are good and which are bad and at the same time make the FSD Beta much more enjoyable to use.

I think the main reasons they don't have this option right now are:

1) Based on Autonomy Day, the car is constantly calculating when / where / how to make lane changes to complete the route in a safe / comfortable / efficient manner. If the car requests authorization, then the delay in the driver authorization may change the Monte Carlo decision tree, so the car might end up canceling the lane change, re-request it, etc. This would create an odd and perhaps confusing user experience.

2) FSD beta is still rough, and they're working at break-neck speed to improve all aspects of it, so building some feature / option that has pros / cons may not be worth it at this point.

 

[EVNow]

I have right and left turns that fail a lot and have reported them multiple times. It must be hard because these are kind of important. Two right turns I use all the time.

Check the map. Usually when the turns fail a lot - there is a mapping issue (unless the intersection is inherently complex).

Solving any single issue is easy, but solving it for all locales and intersection geometries is difficult.

Its not about solving all turns - because that would mean FSD Beta is ready for general release. I'm talking about the specific case (as OP alluded to) of the car stopping in the middle of a lane with side / oncoming traffic.

 

[Knightshade]

Well it has be > 1 year of 2000 FSD Beta testers before we were even invited and they have not solved it yet.

To be fair, something like 95%+ of those 2000 were Tesla employees-and then a tiny handful of youtubers and such...which means vast bulk of incoming data was likely only located in a very few places.

See again Elons remarks about overfitting to the bay area.

 

[d21mike]

I think the main reasons they don't have this option right now are:

1) Based on Autonomy Day, the car is constantly calculating when / where / how to make lane changes to complete the route in a safe / comfortable / efficient manner. If the car requests authorization, then the delay in the driver authorization may change the Monte Carlo decision tree, so the car might end up canceling the lane change, re-request it, etc. This would create an odd and perhaps confusing user experience.

2) FSD beta is still rough, and they're working at break-neck speed to improve all aspects of it, so building some feature / option that has pros / cons may not be worth it at this point.

I mostly agree with you #1 because sometimes when in the mood to test FSD NOA (freeway) I turn on auto because I may be to slow to respond. So, I was really hoping that Chill, Average, Assertive would take care of it but it did not. There is no difference. I would think Chill would not want to change lanes so much. Just chill and drive down the highway. In my case sometimes it is changing lanes and there is NO CARS in front of me. It makes no sense after all this time.

 

[powertoold]

Its not about solving all turns - because that would mean FSD Beta is ready for general release. I'm talking about the specific case (as OP alluded to) of the car stopping in the middle of a lane with side / oncoming traffic.

Yup, we want fsd beta to creep into oncoming lanes in certain situations (like unprotected left at a light, the car should creep into the intersection), but if Tesla creates code that says (for example) "never get in front of any cars," this would lead to undesirable behavior.

Many decisions that may seem obvious to us as experienced drivers with human brains is difficult to re-create using procedural code, as the variables may change according to our understanding of driving behavior and semantics among humans and road regulations.