FSD Beta 10.11

[j0shm1lls]

Don't you guys on 10.11.2 worry... I'll keep rolling around farting in people's general direction on 10.10.2. Go on without me, don't look back, back, back bac...

 

[SilverString]

Valuable source of data. Get it right.

TeslaFi shows 995 (including me) at 10.10.2 and 1,441 at 10.11.2. Only 47 installs pending for 10.11.2. Not sure why the rollout of 10.11.2 seems to have paused, because it could be an intentional lets-keep-gathering-comparative-data. Or, noticing that there are 1,416 pending installs of non-FSD 2022.8.3, they might have put the 10.11.2 rollout on the back burner to get the non-FSD fleet up to 2022.8.3 as soon as possible. Or both, I suppose.

Well those are interesting and plausible reasons. I suspect it is the former and they are comparing data between the versions. To me that indicates another update after 11.2 before those of us on 10.10.2 get anything new. Do you think their is a bandwidth or some other limitation on how many OTA's can be performed at once or over a certain period of time?

 

[Supcom]

Valuable source of data. Get it right.

TeslaFi shows 995 (including me) at 10.10.2 and 1,441 at 10.11.2. Only 47 installs pending for 10.11.2. Not sure why the rollout of 10.11.2 seems to have paused, because it could be an intentional lets-keep-gathering-comparative-data. Or, noticing that there are 1,416 pending installs of non-FSD 2022.8.3, they might have put the 10.11.2 rollout on the back burner to get the non-FSD fleet up to 2022.8.3 as soon as possible. Or both, I suppose.

Given that FSD beta 10.10.2 has been out for two months, there should be little need for comparative data from that version. Most likely, evaluation of 10.11.2 telemetry shows some significant issues that need to be addressed before deploying to more cars. I see plenty of users reporting problems and saw a youtube video of one car where FSD beta software crashed (the computer, not the car!) with a red wheel of death and loss of all visualization.

The last couple of test versions has taken three, or more iterations to get out to all beta testers. I suspect that Tesla has some serious internal testing issues.

 

[LowlyOilBurner]

….but lots of YouTube/social media “influencers” say version xx.xx.x is “mind blowing”…

Curb’d wheels and vomiting wives say otherwise.

 

[sleepydoc]

Given that FSD beta 10.10.2 has been out for two months, there should be little need for comparative data from that version. Most likely, evaluation of 10.11.2 telemetry shows some significant issues that need to be addressed before deploying to more cars. I see plenty of users reporting problems and saw a youtube video of one car where FSD beta software crashed (the computer, not the car!) with a red wheel of death and loss of all visualization.

The last couple of test versions has taken three, or more iterations to get out to all beta testers. I suspect that Tesla has some serious internal testing issues.

But 11.2 has only been out a week. If they want comparative data between 10.2 and 11.2 they'll need time to get data and then analyze it. I strongly suspect this is what's going on. If you're troubleshooting an algorithm like this you identify the issues/problem scenarios, develop the fix, then compare behavior. People like me are consigned to being valuable sources of control data.

 

[wknickless]

Do you think there is a bandwidth or some other limitation on how many OTA's can be performed at once or over a certain period of time?

Could be. Could also be a general risk mitigation strategy of metering the rate of change to the fleet so as not to overwhelm their support channels if something really horrible slips through testing and (say) bricks some of the embedded sub-controllers--even if such things happen only on one out of a thousand cars.

Given that FSD beta 10.10.2 has been out for two months, there should be little need for comparative data from that version. Most likely, evaluation of 10.11.2 telemetry shows some significant issues that need to be addressed before deploying to more cars. I see plenty of users reporting problems and saw a youtube video of one car where FSD beta software crashed (the computer, not the car!) with a red wheel of death and loss of all visualization.

The last couple of test versions has taken three, or more iterations to get out to all beta testers. I suspect that Tesla has some serious internal testing issues.

Two nights ago (on 10.10.2) the screen and audio system crashed after several hours of driving. It was so dead that not even the turn signal clicking was audible. A two-thumb reboot (after pulling over to the shoulder) restored everything.

Given their level of investment in data collection, model training and infrastructure, Tesla might choose to rewrite much of their critical car management and self driving stack in a memory-safe language like Rust. Traditional testing just isn't up to catching the sorts of nasty and subtle Heisenbugs that creep into high performance multithreaded C/C++ codebases. Such a move could put FSD delivery on hold for a year or more, but ultimately be the right long-term solution for high availability life-critical software like FSD.

 

[Supcom]

But 11.2 has only been out a week. If they want comparative data between 10.2 and 11.2 they'll need time to get data and then analyze it. I strongly suspect this is what's going on. If you're troubleshooting an algorithm like this you identify the issues/problem scenarios, develop the fix, then compare behavior. People like me are consigned to being valuable sources of control data.

It looks like they got plenty of data from the 20,000 or so testers and quickly determined that they had another bad release. It's getting pretty far along for there to still be FSD software crashes.

Tesla should need no more data from 10.10.2 cars. Those vehicles have been testing for months now. And, if they needed comparative data for fixes, they wouldn't have updated the 10.11.1 cars so quickly.

No, this smacks of a very problematic 10.11 release, followed by two knee-jerk fixes that have not gone well. Maybe there was pressure to get a new release out before the big Giga Texas event so the influencers would come and spread the joy.

 

[GlmnAlyAirCar]

One thing I have noticed on the new fsd is phantom braking is actually worse.

Actually, I would be disappointed if phantom braking didn't get worse from time to time. It demonstrates progress. Let me explain.

I drive a lot of rural roads, many of which are lined with mailboxes coming to the edge of the street. When I first got FSD beta (10.3 I believe?), it would consistently "phantom" brake at the majority of these mailboxes. Frequently, I could see them rendered as humans, facing toward the street. Obviously, there was much work to be done (and data collected) to reduce these false positives. Since 10.3, these false positives have reduced to zero.

Over time, one would expect FSD to start recognizing more and more object types. For instance, there is a note in the 10.11 release notes about recognizing car doors. I am certain there are other object types besides those listed in the release notes that are now being recognized as well. But, until there is sufficient data and the neural nets are trained on this data, recognition of these object types will be poor. Hence, additional phantom braking because of these false positives.

In subsequent releases, these false positives should decrease, and so should phantom braking.

If FSD had never started to recognize additional object types, we might have less phantom braking now, but the system would not be moving toward a safer system that may allow for some degree of autonomy down the road.

 

[Supcom]

Over time, one would expect FSD to start recognizing more and more object types. For instance, there is a note in the 10.11 release notes about recognizing car doors. I am certain there are other object types besides those listed in the release notes that are now being recognized as well. But, until there is sufficient data and the neural nets are trained on this data, recognition of these object types will be poor. Hence, additional phantom braking because of these false positives.

Why would there not be plenty of data to correct for false positives? Tesla should have lots of data with and without new objects for test purposes.

 

[wknickless]

Why would there not be plenty of data to correct for false positives? Tesla should have lots of data with and without new objects for test purposes.

FSD 10.9 was released on 16 January 2022, less than 90 days ago. The release notes say 10.9's "static object network" is using "10-bit photon count streams rather than 8-bit ISP tonemapped images".

Not sure anyone outside Tesla knows how long they've been collecting, labeling and training against 10-bit photon count data rather than the old 8-bit ISP tonemapped data, so that might be a limiting constraint. Maybe they don't have 10-bit photon count data going all the way back to spring of last year, so the (e.g.) weather and plant growth we see this time of year in North America is effectively new and therefore not included in the training sets.

 

[tmoz]

Have any of the 10.x releases not been labeled problematic at one point or another?
It sounds like Phantom Braking is a problem every release has had.
What about 4-way stop signs? Are there still problems there, or did one of the releases fix that?
Roundabouts are still problematic from what I'm reading from you guys.
Speaking of guys, are there any female beta testers on this thread?

 

[antchan]

I don’t like the new map UI update, the route is more towards the passenger side now, looks weird

View attachment 791299

Please excuse my ignorance. Which sw version is this?

 

[Skione65]

I don't know what is going on with the newest update. My last 2 FSD trips (with the latest version 10.11.2) missed the required exit to my destination each time!

This new update is Trash. I honestly don’t know what’s going on. I went for a drive to run errands today with it and it’s downright unusable. I can’t even list everything that went wrong. Just plain got tired of the manual disconnects and overrides and constantly hitting the camera report. Holy hell. What a regression from 10.2. From turns at T intersections, to 4 ways, to missing exits, wrong turn lanes, going straight in turns. I just plain turned FSD off. I was so excited after watching Cyber Rodeo (though he didn’t say anything new) then after this drive I came back down to earth. LOL.

Ski

 

[Supcom]

FSD 10.9 was released on 16 January 2022, less than 90 days ago. The release notes say 10.9's "static object network" is using "10-bit photon count streams rather than 8-bit ISP tonemapped images".

Not sure anyone outside Tesla knows how long they've been collecting, labeling and training against 10-bit photon count data rather than the old 8-bit ISP tonemapped data, so that might be a limiting constraint. Maybe they don't have 10-bit photon count data going all the way back to spring of last year, so the (e.g.) weather and plant growth we see this time of year in North America is effectively new and therefore not included in the training sets.

60,000 beta testers should be generating huge amounts of training data every day. My Tesla has uploaded 10 GB on WiFi today alone and I drove it less than 2 miles without even using FSD beta!

You can try to think everything is going great, but sadly, it's not. Tesla's AI director has taken a 4+ month sabbatical and beta releases are moving backward in reliability. What progress that has been released in the last six months has been very minor. Certainly not the big improvements required to meet any of the CEO's overstatements.

Just look at Chuck Cook's unprotected left turn series on youtube. Unprotected lefts are a big deal for self driving. Really big. Yet, there's been no real improvement on these for a long time. This certainly should be a big priority for development - as big as solving phantom braking. So, why no sunstantial change from version to version? Even dealing with going straight through from a stop sign is problematic when there is NO other traffic.

Don't get me wrong. I love my Tesla and love what Tesla has been able to accomplish with FSD beta so far and look forward to further progress. But I believe that any objective observer would conclude that progress has stalled.

 

[jebinc]

60,000 beta testers should be generating huge amounts of training data every day. My Tesla has uploaded 10 GB on WiFi today alone and I drove it less than 2 miles without even using FSD beta!

You can try to think everything is going great, but sadly, it's not. Tesla's AI director has taken a 4+ month sabbatical and beta releases are moving backward in reliability. What progress that has been released in the last six months has been very minor. Certainly not the big improvements required to meet any of the CEO's overstatements.

Just look at Chuck Cook's unprotected left turn series on youtube. Unprotected lefts are a big deal for self driving. Really big. Yet, there's been no real improvement on these for a long time. This certainly should be a big priority for development - as big as solving phantom braking. So, why no sunstantial change from version to version? Even dealing with going straight through from a stop sign is problematic when there is NO other traffic.

Don't get me wrong. I love my Tesla and love what Tesla has been able to accomplish with FSD beta so far and look forward to further progress. But I believe that any objective observer would conclude that progress has stalled.

Yep, stalled since 10.3, when I opted out. May hold off getting back I until a useable v11 releases next year (or the year after).

 

[nvx1977]

Have any of the 10.x releases not been labeled problematic at one point or another?
It sounds like Phantom Braking is a problem every release has had.
What about 4-way stop signs? Are there still problems there, or did one of the releases fix that?
Roundabouts are still problematic from what I'm reading from you guys.
Speaking of guys, are there any female beta testers on this thread?

Phantom braking is not a singular problem/bug. It's the result of false positives from thousands (or more) situations. So it's likely going to exist to some degree even when FSD is declared safe for L4/5 categorization.

And I don't think any of us can objectively say with certainty if PB is getting better/worse. Improvements have been expressed as percentages in the release notes. But a percentage of what? 50% improvement from a success rate of 0.01% is hardly noticeable (exaggerated numbers to illustrate the point).

Yeah, I've only see people mentioning wives' reactions in the passenger seat. This part of the forum skews heavily male in an already male-skewed overall Tesla demographic.

 

[JulienW]

.....Just look at Chuck Cook's unprotected left turn series on youtube. Unprotected lefts are a big deal for self driving. Really big. Yet, there's been no real improvement on these for a long time. This certainly should be a big priority for development - as big as solving phantom braking. So, why no sunstantial change from version to version? Even dealing with going straight through from a stop sign is problematic when there is NO other traffic.

Don't get me wrong. I love my Tesla and love what Tesla has been able to accomplish with FSD beta so far and look forward to further progress. But I believe that any objective observer would conclude that progress has stalled.

Perfect example and you could watch one of his videos from a year ago and probably not be able to tell which is new and which is old. While we do see some improvements all the time we are NO WHERE near reaching the hard part of chasing the 0.999%. Hell we still seem far away from having Feb's integrated stack. We need a GIGANTIC leap in order to have any chance of L3 (L5, no way) before Dec 31. Just can't see L3 (don't think we will get L4/5 with hardware) before about Jun 22. Hope I'm wrong but just go back and watch any FSD videos from a year or more ago and you can see they are incremental worse at the most.

Of course I'm probably more pessimistic since I'm still on 10.10.2.

 

[jabloomf1230]

Perfect example and you could watch one of his videos from a year ago and probably not be able to tell which is new and which is old.

Chuck's ULT has been discussed over and over here. It is a reasonable test if one is interested in how FSD behaves under extreme conditions. Without additional forward-placed side facing cameras, Tesla's FSD is never going to handle that specific edge case with a 0.01% failure rate. FSD can't negotiate what it can't sense.

I have no idea why Chuck persists in testing that same ULT. He already knows the reasons why the car can't safely make that maneuver.

 

[JulienW]

Chuck's ULT has been discussed over and over here. It is a reasonable test if one is interested in how FSD behaves under extreme conditions. Without additional forward-placed side facing cameras, Tesla's FSD is never going to handle that specific edge case with a 0.01% failure rate. FSD can't negotiate what it can't sense.

I have no idea why Chuck persists in testing that same ULT. He already knows the reasons why the car can't safely make that maneuver.

...and why I said I don't think we will ever have L4/5 with current hardware. I will be happy with L3 but believe it is "pie in the sky" aspirational believing our cars will be safely driving around with no one behind the wheel. Living in the middle of a city I know first hand how often the view is occluded, especially from the B-pillars.

Just to add is another BIG occlusion. The rear camera is wide angle and can't do as good a job sensing depth. Also when reversing the vision is EXTREMELY limited and dangerous. Repeater cameras are even more likely to be blocked than B-pillar's. You can't have L4/5 and avoid all occluded B-pillar turns and never have to reverse in a dangerous situation.

 

[jabloomf1230]

...and why I said I don't think we will ever have L4/5 with current hardware. I will be happy with L3 but believe it is "pie in the sky" aspirational believing our cars will be safely driving around with no one behind the wheel. Living in the middle of a city I know first hand how often the view is occluded, especially from the B-pillars.

You're right, but one way to achieve something greater than L3 is for the car to not make any high traffic unprotected left turns. Then we have to hope that the robotaxi doesn't charge by either distance traveled or elapsed time.