FSD Beta 10.69

[VanFriscia]

Why on earth, that somebody would find this humorous, is beyond my comprehension. It smacks of passive-agressive trolling. It might be time for TMC to limit the post "like" choices to just upvote.

Personally, I agree with you - laughing at a post that clearly wasn't intended to be funny comes off as condescending/belittling, in my opinion. But for some of these offenders, I'm pretty certain that wasn't their intent. They legitimately found something about it to be funny in some way, without being able to explain what it was because it's just a simple button. I try hard not to do that, but I also try not to get my feelings hurt when it happens to me (and it's happened several times).

 

[jebinc]

Why on earth, that somebody would find this humorous, is beyond my comprehension. It smacks of passive-agressive trolling. It might be time for TMC to limit the post "like" choices to just upvote.

 

[berryja]

So after my couple of trips to the pharmacy to pick up the hemorrhoid cream, I have found the following:

1) Still has issues with breaking for traffic lights that are on curved roads, either brakes late and hard or brakes hard and then let’s go and slows down to a stop. Need a better smoother use of regen.
2) Some places, it’ll still get into the right hand turning lane, despite the GPS indicating to go straight.
3) Weird mapping issue with the 4 way intersection by my community is still there. The car puts on a turn signal but goes straight when it has to go straight through the intersection.
4) The delayed braking for crossing traffic is extremely annoying. When driving on a road and a vehicle up ahead making a left-hand turn from a road or a business on the right, crossing through my lane of travel, makes the car hit the brakes, despite the crossing car already being finished with their turn and out of the way.

If I remember anything else I’ll add. It’s now time to apply the cream.

I have seen the turn signal on at intersection when mapped to go straight too. The signal turns on for a bit then turns off.. then repeats 10 secs later.

 

[Mardak]

especially with no curb - just 8-12 inches of shoulder outside the right line then dirt/grass

I haven't retested this with 10.69, but previous versions did sometimes accidentally drive off the road edge where there were no painted lines into dirt, and in reviewing TeslaCam video to email to Tesla, it's visually really hard to tell whether a dirt/grass road edge is level with the road or actually a few inches drop. And dirt edges can have sudden changes where rain / water runoff happened to wash away more of one section with other vehicles splashing mud that has now dried so that the road edge is even less clear.

I would guess this is related to how confident the neural network predicts a road edge to be as well as how granular these annotations are. E.g., maybe a prediction is unsure about just a few feet of road edge, but does that cause 10.69 to drive further away from the whole road edge to avoid accidentally driving off road?

I suppose more generally, how do people know if it's safe to drive into say grass where who knows how deep it actually drops? I suppose typically grass is only so tall when maintained, so slowing driving on/off should be okay?

 

[Ramphex]

I have seen the turn signal on at intersection when mapped to go straight too. The signal turns on for a bit then turns off.. then repeats 10 secs later.

Mine stays on and the car goes straight then turns it off

 

[EVNow]

Yesterday drove for quite a bit on some errands. Probably 90 minutes on FSD Beta. We have now started counting where FSDb is bad (instead what it can do) - that is some real progress !

+ves :
- Turns are less jittery. A few turns it would earlier abort were ok - esp. smaller roads with a car waiting at the stop line on the road the car is turning in to.
- In 90 minutes of driving I didn't get any difficult unprotected lefts. Thats how rare they are here ... But in general rights & lefts have become better - esp. the creeping wall visualization is helpful.
- Unlike what some people were saying, I don't find 69.2 any more jittery when people are around. Seems same as 10.12.
- Keeps a larger distance from the front car, may be they are actually taking the follow distance into consideration. I'll have to experiment by changing the follow distance, but this is +ve if we can control it better.

-ves :
- Roundabouts are still problematic. This means my disengagement rate won't go down much. They need to spend substatial amount of time working on roundabouts - like they did with CULT.
- Wide unmarked roads might be somewhat worse, when the road is bending/winding. The path chosen optimizes for lower turning radius than keeping to the right side of the road. So, if they spend some time on this issue, they should able to fix it easily.
- Slows down late. They should change this atleast for the "chill" profile.

ps : Moving to here from the Tweet thread I posted by mistake.

 

[Ben W]

Is your experience so bad that it feels only like 20% done?

FSD is simultaneously incredibly impressive technology, and yet not remotely close to anything that could be considered L4. It's a situation where the last 10% of the project takes 90% of the effort. In city-streets driving with 10.69.2, I estimate that I have to make a moderate intervention (e.g. tapping the accelerator or changing lanes) about once per minute, and a more serious intervention (disabling FSD, or FSD disables itself) about every 10 minutes. This 10 minutes will need to become 10 years to reach L4. Put another way, on a minute-by-minute basis, the current system is about 90% reliable. It will need to approach something like 99.99999% minute-over-minute reliability to achieve L4. So in one sense (reliability per minute) you could simplistically say that they're already 90% of the way there, but in a more important sense (number of 9's of reliability required to attain L4), they are only ~20% of the way there, if that.

A more important and shorter-term milestone will be when they reach a point where it's less mentally taxing to drive with the system turned on than turned off, for equivalent levels of safety. ("Useful L2.") This is already very much the case on the highway, and has been so for a few years. But on city streets, at the rate of current progress, I'd guess this tipping point is probably still 2-4 years away. So in terms of reaching useful city-streets L2, given the Oct 2020 start date for FSD Beta, I'd say they are optimistically about 50% there.

 

[momo3605]

- Roundabouts are still problematic. This means my disengagement rate won't go down much. They need to spend substatial amount of time working on roundabouts - like they did with CULT.

What is your issue from them? What's interesting is that if you look at Chucks videos recently where they introduced roundabouts, in 69.1 it would completely disregard traffic and drive full speed through (and you can see Chuck has to slam the brakes a couple times for this) but in 69.2 it reverts back to just coming to a full stop before entering the round about, and then he can't get it to go unless using accelerator intervention. So it's kind of strange that the behavior is a total 180 between minor revisions. Seems like they are working on it, but are not yet sure what behavior they want to be acceptable.

 

[Dewg]

So after my couple of trips to the pharmacy to pick up the hemorrhoid cream, I have found the following:

1) Still has issues with breaking for traffic lights that are on curved roads, either brakes late and hard or brakes hard and then let’s go and slows down to a stop. Need a better smoother use of regen.
2) Some places, it’ll still get into the right hand turning lane, despite the GPS indicating to go straight.
3) Weird mapping issue with the 4 way intersection by my community is still there. The car puts on a turn signal but goes straight when it has to go straight through the intersection.
4) The delayed braking for crossing traffic is extremely annoying. When driving on a road and a vehicle up ahead making a left-hand turn from a road or a business on the right, crossing through my lane of travel, makes the car hit the brakes, despite the crossing car already being finished with their turn and out of the way.

If I remember anything else I’ll add. It’s now time to apply the cream.

2) I've seen some pathing issues too, I was hopeful they would have improved the planner - so this is a little disappointing.
4) This feels the same from previous versions. In AP, the lead car is darker in color than surrounding cars. In FSD Beta the cars that are being focused on by the NNs are in blue. I've noticed that speed changes are based on when the cars change color. In AP, once the dark car changes lanes, my car speeds up the moment the dark car changes to grey. In FSD Beta, my car changes speed the moment the blue car changes to grey. The car feels like it's being extra cautious. We, as humans, infer from the behavior of the car crossing our path, and know from experience that it will likely continue its trajectory and we can safely proceed without slowing down. FSD Beta is being cautious, assuming the crossing car could, at any moment in the lane, slam on its brakes and stop right in front of you. The odds are low that will happen, but not 0. Humans would either slam on the brakes, or make a jerky turn to the right to avoid hitting them, possibly causing other cars to react. Hopefully this behavior from FSD Beta will improve, but I understand where it's coming from here.

 

[kriskros91]

FSD is simultaneously incredibly impressive technology, and yet not remotely close to anything that could be considered L4. It's a situation where the last 10% of the project takes 90% of the effort. In city-streets driving with 10.69.2, I estimate that I have to make a moderate intervention (e.g. tapping the accelerator or changing lanes) about once per minute, and a more serious intervention (disabling FSD, or FSD disables itself) about every 10 minutes. This 10 minutes will need to become 10 years to reach L4. Put another way, on a minute-by-minute basis, the current system is about 90% reliable. It will need to approach something like 99.99999% minute-over-minute reliability to achieve L4. So in one sense (reliability per minute) you could simplistically say that they're already 90% of the way there, but in a more important sense (number of 9's of reliability required to attain L4), they are only ~20% of the way there, if that.

A more important and shorter-term milestone will be when they reach a point where it requires less mental effort to drive with the system turned on than turned off, for equivalent levels of safety. ("Useful L2.") This is already very much the case on the highway, and has been so for a few years. But on city streets, at the rate of current progress, I'd guess this tipping point is probably still 2-4 years away.

I understand what you're saying and I agree with the overall sentiment. I disagree about the time frame it will take to achieve this. I got in ~11/2021 with 10.3.1 and the leaps and bounds of improvements that has been made is extraordinary. There are no doubt many extremely talented engineers behind the scenes that have transformed the way this system perceives and reacts to the world around it.

I know the pace of improvement is going to significantly change now. But even over the 100+ days between updates I notice such a remarkable amount of improvement that I am so hopeful that we will get to near flawless driving reasonably quickly. The disengagements in the future will be more preference based (I want to pass this slow car, I want to get around this car turning into this store parking lot, etc) and less "safety" problems.

That being said I drove with FSDb today with my wife who is a non-believer at this time excited to show her how great everything was. On every single drive (each about 15-20 mins a lot of which were on city streets with FSD) required disengagements and some actual safety issues. I should have put chill mode on (was on aggressive to see if there was a noticeable difference) and the car went for unprotected windows where no reasonable human would go for and would often make wide turns that were not needed resulting in discomfort. Some mapping issues have really crept up also - car was supposed to turn left at a light and correctly chose the outside turning lane. Put the turn signal on at the light and when it went green went forward which I thought was preparing for a wide turn and just slowed down to almost a stop and then just decided to go forward. Total fail resulting in me having to go 1+ mile forward to do a U-turn. Not good especially with the wife watching.

 

[Ramphex]

2) I've seen some pathing issues too, I was hopeful they would have improved the planner - so this is a little disappointing.
4) This feels the same from previous versions. In AP, the lead car is darker in color than surrounding cars. In FSD Beta the cars that are being focused on by the NNs are in blue. I've noticed that speed changes are based on when the cars change color. In AP, once the dark car changes lanes, my car speeds up the moment the dark car changes to grey. In FSD Beta, my car changes speed the moment the blue car changes to grey. The car feels like it's being extra cautious. We, as humans, infer from the behavior of the car crossing our path, and know from experience that it will likely continue its trajectory and we can safely proceed without slowing down. FSD Beta is being cautious, assuming the crossing car could, at any moment in the lane, slam on its brakes and stop right in front of you. The odds are low that will happen, but not 0. Humans would either slam on the brakes, or make a jerky turn to the right to avoid hitting them, possibly causing other cars to react. Hopefully this behavior from FSD Beta will improve, but I understand where it's coming from here.

I do too. But the braking is delayed and happens after the car has already crossed the path. That’s the annoying part.

 

[jebinc]

Haven’t been following all the posts. As anyone reported any new testers being added yet?

 

[EVNow]

Haven’t been following all the posts. As anyone reported any new testers being added yet?

Just one new tester out of ~2,900 on Teslafi.

 

[MrTemple]

I laughed when I read it. I guess I'm just easily amused.
The idea that miles driven with FSD beta is a metric for progress is comical. Sorry, not sorry.

Maybe you can explain why the accelerating slope of those particular miles driven isn’t ”illustrative of the accelerating pace of capability”?

Without glossing over the facts that:

1. The TYPE of driver has changed fairly dramatically (from paid internal tester with very high systems knowledge, to testers who demonstrated extreme test pilot capability in maintaining 100% on the arbitrary safety score, eventually to fairly average test-pilot capability in maintaining only 93% on the safety score).

2. The WILDLY increasing situational variety of roads and traffic patterns in all the variety of US/Can regions. (A critical challenge for generalized driving AI).

3. There has been one low-speed FSD Beta bollard-boop in all those millions of miles.

I’m sorry, but if you can’t see how that does illustrate accelerating pace of capability, you simply don’t understand what’s happening.

 

[Ben W]

But even over the 100+ days between updates I notice such a remarkable amount of improvement that I am so hopeful that we will get to near flawless driving reasonably quickly.

An analogy: think of AI and Moore's Law. Since the 1950's, computers have been doubling in power every two years. That's a remarkable amount of improvement in any given short time-frame. Computers in 1986 felt significantly faster than computers in 1984. Yet here we are, 70 years later, and we still don't have intelligent computers, in the general AI sense.

Suppose FSD is on a 4x faster pace, getting twice as good every 6 months. There still could be many years before it achieves enough "intelligence" to substitute for a human driver in an L4 capacity, or several years for a useful L2 capacity. Useful L2 will obviously come much sooner than L4, but I still believe (as a FSD Beta tester since last October) that it's at least 2 years off, probably more.

 

[Dewg]

I do too. But the braking is delayed and happens after the car has already crossed the path. That’s the annoying part.

That has not been my experience, but I will pay more attention when the situation occurs next - to see when the car reacts vs the color of the car in the visualization.

 

[mhan00]

FSD handles my suburban area really well now (outside of parking lots). It always did, but it’s to the point where if I for some reason had to drive while disabled in some way I’d have confidence in the car making it to the destination without any issues besides maybe annoying other drivers by hesitating at stop signs every so often. There is one interesting regression from early versions when entering a particular parking lot from one particular entrance thati think is due to the supposed switch to more dependence on vision over map data. There is a slight hump on the parking lot entrance so you can’t see the parking lot road until you crest that hump. Before, presumably when it used map data more, it would just drive. Now, it meanders and panics and you can see the blue driving space on the screen go a little crazy trying to figure out what to do.

 

[EVNow]

What is your issue from them? What's interesting is that if you look at Chucks videos recently where they introduced roundabouts, in 69.1 it would completely disregard traffic and drive full speed through (and you can see Chuck has to slam the brakes a couple times for this) but in 69.2 it reverts back to just coming to a full stop before entering the round about, and then he can't get it to go unless using accelerator intervention. So it's kind of strange that the behavior is a total 180 between minor revisions. Seems like they are working on it, but are not yet sure what behavior they want to be acceptable.

To properly handle roundabouts FSDb has to
- understand entering roundabouts is like an unprotected right, with an implicit yield sign
- understand which lane to use in multi-lane roundabouts depending on the exit
- understand we have right of the way when it comes to vehicles coming from the right
- correctly predict the future location of the vehicles coming from left since they will be traveling in a circle and straight line

Roundabouts don't have any of these logic. The stop / not stop behavior is probably a flag they set about how to handle yield sign.

 

[2101Guy]

Useful L2 will obviously come much sooner than L4, but I still believe (as a FSD Beta tester since last October) that it's at least 2 years off, probably more.

Elon confirms that Tesla will reach L4 before 12/31/2022.

1:06 mark here

 

[Mardak]

Just one new tester out of ~2,900 on Teslafi

Even then, unclear if that TeslaFi vehicle just wasn't connected to the service since January when it last indicated it was using 2021.44.30.2. That vehicle could have been added to FSD Beta back in June but wasn't actively using TeslaFi then and only now resumed to show it's currently on 2022.20.15 / FSD Beta 10.69.2.