FSD Beta 10.69

[FOOLSLFDRVNG]

YO @ELONMUSK, 10.69.1.1 WAS SUPER DUPER ULTRA BEYOND HUMAN. RELEASE!!!! 100 MILES NO INTERVENTIONS!!!! SMOOTH AS BUTTER!!!! ROBOTAXI READY 4 SHIZZLE

That sounds like Whole Mars Catalog

 

[AndrewZ]

If this type of VRU sensitivity is necessary for 10.69.2 with the current pool of testers who have all long been vetted by the Safety Score system, it begs asking what will be required for a wide release to the fleet of Autosteer on City Streets newbies potentially without the safety score

 

[swedge]

While we, oh so patiently, wait for our turn for 10.69.2, I have a quick question about "forced disengagements" for other than driver inattention.

Have people seen it happen? What circumstances?

I am on 10.12.2, and I don't think I've seen FSD disengage itself. I am wondering if it ever does this.

I've seen it flash the top of the screen, beep, etc, when it thought I was not paying attention. (I've met some children who act similarly.) But I'm wondering what happens if FSD β goofs and puts itself in a situation which it does not know how to handle. Does it beep, flash the red steering wheel with a message to take control? Does it let go of steering? Slow to a stop?

Thanks,
SW

 

[momo3605]

I'm just curious at what point Tesla adds new people to the FSD Beta testing pool. Presumably those with 10.12.2 will be prioritized for updates before new people are added.

I'm very confident that 10.12.2'ers will get it before new people. Up until now, that's always been the case. Elon also hasn't said anywhere since 10.13 was announced that 10.13/10.69 will be opened to more people. I don't expect new people to get the update any time soon.

 

[Mardak]

If this type of VRU sensitivity is necessary for 10.69.2 with the current pool of testers that have all long been vetted by the Safety Score system

Maybe they're increasing the sensitivity because it's getting ready for wider release with additional testers. NHTSA does have FSD Beta in their Vehicle Research and Test Center since earlier this year, so maybe they want to see some of these types of safety improvements knowing that Tesla is aiming for wide release without safety score.

 

[HappyPanda]

Maybe they're increasing the sensitivity because it's getting ready for wider release with additional testers. NHTSA does have FSD Beta in their Vehicle Research and Test Center since earlier this year, so maybe they want to see some of these types of safety improvements knowing that Tesla is aiming for wide release without safety score.

I do feel like a bit of caution will go a long way for regulatory approval, especially looking even longer-term where the driver is no longer required to pay attention, at which point the occasional slowdowns won't be as aggravating (assuming of course it doesn't add additional risk of collision from other cars).

 

[powertoold]

Judging from the videos, it seems pedestrian sensitivity is way too high now. I wonder why because 1.1 seemed to be "good enough" around lots of pedestrians.

 

[momo3605]

I wonder why

because polish! I knew this was going to happen lol. If they had just kept 69.1 and 69.2 to safety related fixes, we'd all happily (or maybe still unhappily lol) be cruising on 10.69.2 by now

 

[EVNow]

Not sure if it's a software development issue, configuration management, NN training, etc but these builds are having a hell of a time maintaining past improvements as well as eliminating bugs. It almost never pays to rush something to delivery but the performance on the last 3 builds is a head scratcher. The responsiveness isn't uniform. It's slow in some benign scenarios (stopping in oncoming lanes) but old school laggy in others. That's not confidence inspiring to the driver.

Maybe the the design is close to reaching an optimal balance between responsiveness, phantom braking and little room for new features.

They have just done a major overhaul of the platform - it will take a while to optimize …

 

[ChaoticNeutral]

On 10.12.2 - no update available yet

While we, oh so patiently, wait for our turn for 10.69.2, I have a quick question about "forced disengagements" for other than driver inattention.

Have people seen it happen? What circumstances?

I am on 10.12.2, and I don't think I've seen FSD disengage itself. I am wondering if it ever does this.

I've seen it flash the top of the screen, beep, etc, when it thought I was not paying attention. (I've met some children who act similarly.) But I'm wondering what happens if FSD β goofs and puts itself in a situation which it does not know how to handle. Does it beep, flash the red steering wheel with a message to take control? Does it let go of steering? Slow to a stop?

Thanks,
SW

I’ve had FSDb disengage when my gps location wasn’t accurate. In this case I was driving along a 30mph main road, but the location was off to the left of the road by about 50-100ft so FSD would keep dropping to 20mph max cause it thought I was on a side street every time my location marker went over the side street. I constantly had to increase the max to 30mph and eventually I got a giant “take control immediately” red wheel of death. This has happened a few times at this same place, usually after a car wash but I don’t know if that was a factor in messing up gps accuracy? I know car washing will always make FSD force a window clean whenever I’m behind someone at a red light, which is absolutely frustrating every time (and I report it every time with snapshot)

 

[Dreadnought]

You got me to thinking..... I guess this is mainly true because some people will always complain about anything and everything. I know a few of them

<snip/>

This whole argument about Shannon limits and information processing is why I tend to disbelieve the occasional poster who shows up and says, without much evidence other than Gut Feeling, that it'll be years or never before self-driving cars roam the highways. That's a human-centric view that Humans are Best At Driving, Period. I don't think so: There's plenty of hardware and other living entities out there that are better at various tasks than we are, so, where's the proof that self-driving isn't a task that nothing else can do? (Strawman argument, I realize, but, still.)

Unfortunately, as European I won't get the chance to get FSD Beta for the foreseeable future and have to settle for other users' reviews to get a feel for the progress of the program. On the one hand, I am thankful for the ever increasing number of testers and the wider range of individual perspectives. On the other hand, it has become a lot more difficult to filter out the useful feedback from the "it's still not behaving like a perfect clone of my immaculate own driving style" complaints, often combined with scepticism whether it's ever going to be as good as a human.

Someone said (during AI day? Can't remember the source right now) that driving, be it automated or by humans, is an optimization task with three competing goals:

1. Safety
2. Speed
3. Jerk / acceleration (equals passenger comfort)

There are a few exceptions like UPLs across six or more lanes (unthinkable in EU btw) but otherwise it's a good model. It basically says that driving is easy if you're very slow or if you have unlimited acceleration (positive and negative) or if you're willing to compromise on safety. The tricky part is to stay away from these extremes. Or, to put it differently, stay somewhere in the middle of the triangle: be safe enough at an acceptable speed while keeping passengers reasonably comfortable.

Everyone, please keep in mind that for FSD to go out of Beta and consider being "released" doesn't imply that it has to be as good as a veteran driver with a million miles experience. No! Once it's consistently driving at the level of a beginner who just got his/her license, it's certainly fit for a wide release to all licensees. To become an L3/L4 system, it could even compromise on the speed factor quite a bit because people care less about a couple extra minutes when they don't have to drive but can work or watch movies instead.

I appreciate all the firsthand impressions that you share and they are most helpful when it's clear whether findings relate to the car leaving the acceptable safety / speed / comfort zone or complaints are rather about personal preferences.

 

[powertoold]

If I had 1.1, I wouldn't update to .2 lol

 

[dl_jehu]

While we, oh so patiently, wait for our turn for 10.69.2, I have a quick question about "forced disengagements" for other than driver inattention.

Have people seen it happen? What circumstances?

I am on 10.12.2, and I don't think I've seen FSD disengage itself. I am wondering if it ever does this.

I've seen it flash the top of the screen, beep, etc, when it thought I was not paying attention. (I've met some children who act similarly.) But I'm wondering what happens if FSD β goofs and puts itself in a situation which it does not know how to handle. Does it beep, flash the red steering wheel with a message to take control? Does it let go of steering? Slow to a stop?

Thanks,
SW

There is a soft ding like when you disengage with the stalk. Easy to try out if you set up a u-turn - I think it flashes small orange message along the lines of “unable to complete maneuver “.

It can be very dangerous - Fsd beta entered a left turn on a median at one turn too early. I was confident it would take the unprotected left and then reroute. As the Tesla realized it was at the wrong turn it made the soft ding, disengaged and *gradually came to a stop in front of oncoming 50mph traffic*.

 

[ChaoticNeutral]

Unfortunately, as European I won't get the chance to get FSD Beta for the foreseeable future and have to settle for other users' reviews to get a feel for the progress of the program. On the one hand, I am thankful for the ever increasing number of testers and the wider range of individual perspectives. On the other hand, it has become a lot more difficult to filter out the useful feedback from the "it's still not behaving like a perfect clone of my immaculate own driving style" complaints, often combined with scepticism whether it's ever going to be as good as a human.

Someone said (during AI day? Can't remember the source right now) that driving, be it automated or by humans, is an optimization task with three competing goals:

1. Safety
2. Speed
3. Jerk / acceleration (equals passenger comfort)

There are a few exceptions like UPLs across six or more lanes (unthinkable in EU btw) but otherwise it's a good model. It basically says that driving is easy if you're very slow or if you have unlimited acceleration (positive and negative) or if you're willing to compromise on safety. The tricky part is to stay away from these extremes. Or, to put it differently, stay somewhere in the middle of the triangle: be safe enough at an acceptable speed while keeping passengers reasonably comfortable.

Everyone, please keep in mind that for FSD to go out of Beta and consider being "released" doesn't imply that it has to be as good as a veteran driver with a million miles experience. No! Once it's consistently driving at the level of a beginner who just got his/her license, it's certainly fit for a wide release to all licensees. To become an L3/L4 system, it could even compromise on the speed factor quite a bit because people care less about a couple extra minutes when they don't have to drive but can work or watch movies instead.

I appreciate all the firsthand impressions that you share and they are most helpful when it's clear whether findings relate to the car leaving the acceptable safety / speed / comfort zone or complaints are rather about personal preferences.

I have been in FSDb for close to a year now, and the improvements have been great, however the reason I feel it’ll be at least another year before full US fleet release is that they haven’t spent time implementing some super important driving instructions, such as:

• don’t turn into/or use bus only/light rail only lanes (I have to take control every time it does this)
• left turns on red when the turn is onto a one way street in the flow of traffic (this is legal when turning from one way and two way streets in my state)
• roundabouts are still too complicated for FSD
• needs better confidence maintaining set speeds under 25mph (presently it’ll slow for every intersection when on a main road)
• better freeway onramp speed gradients when meters aren’t active
• on ramp meters (and choosing which lane to be in) are presently not done correctly - and sometimes when I am off the ramp max speed will reset to the speed that the street was at for five seconds before jumping back up to 55mph
• school zone speed limit awareness (which can be either a set timespan during school days, when children are present, or only when the school zone speed limit has yellow lights flash)

 

[Mardak]

They have just done a major overhaul of the platform - it will take a while to optimize

Apparently the new video occupancy network needs to be optimized for water and learn about something called reflections.

Although I have noticed with 10.69.1.1, the occupancy network does get confused about how far certain things are, e.g., I was directly behind a motorcyclist, and I guess his clothing confused the occupancy network to think something was in the road ahead of us resulting in the gray blob visualizations well ahead of the motorcycle visualization and an unnecessary "path blockage" lane change.

 

[sleepydoc]

No, definitely not!

I just watched the video in real time, then counted frames in youtube:
0 sec - head barely visible (only identifiable if you know what to look for)
0.65 sec - entire head visible
0.9 sec - shoulder visible
1.2 sec - torso visible
1.6 sec - braking starts

Even knowing what to expect, I don't see the man as a pedestrian until about 1 second in and don't recognize that he might be coming into traffic until about 1.2 sec. Assuming you seem him as soon as possible (a big assumption on a busy street like this), you would only have about half a second to recognize him identify that he's coming into traffic and react. The vast majority of drivers would have missed him completely or only started braking very late, so Yeah, I count this is a score for Tesla.

 

[kabin]

They have just done a major overhaul of the platform - it will take a while to optimize …

NN architecture change would be major. 10.69 looked so promising to start over.

 

[sleepydoc]

You trust a plug at 150 mph?

you drive 150 MPH?

 

[AlanSubie4Life]

Looks like it's currently a 6% rollout of 10.69.2 so far only to those who had received 10.69.1.1, which went up to a 21% rollout. So roughly 1 in 3 people who had 10.69.1.1 should be seeing the update right now.

TeslaFi.com Firmware Tracker

Perhaps. However, I choose to believe I have been put on the blacklist.

Even knowing what to expect, I don't see the man as a pedestrian until about 1 second in and don't recognize that he might be coming into traffic until about 1.2 sec.

When do you see the hazard lights/turn signals? What do they indicate?

 

[Gilliland]

Anyone on 10.12.2 get it or are we still on 10.69.x users?

No, not yet.

Teslascope says the same thing: