FSD Beta 10.69

[AlanSubie4Life]

, one way disengagement rate goes down is when the testers feel less threatened. FSD can improve disengagement rate by gaining more trust.

Looks to me that is exactly what has happened. FSDb has gained a bit more trust from me. So, though it feels like I "lowered" my standard, in reality disengagement rate went down because it gained more trust.

Sure. I think that can happen at these extremely high disengagement rates. Get a bit more comfortable and let it ride a bit more in the jerky situations, and you don't disengage as much. Maybe only half as much!

In the end we'll have to wait until FSD Beta is doing the right thing most of the time (probably needs to be about 100x better before being at this point) and then disengagement rate will be more reflective of the actual safety level, and improvement (presumably nearly every disengagement would then be a critical safety intervention).

Right now it seems to me the measures are really noisy and it's hard to say whether any particular release is better since they all have such high disengagement rates.

 

[willow_hiller]

Also worth remembering that some relatively simple features could dramatically reduced disengagements.

Not an example of a safety-related disengagement, but I drive by multiple "No turn on red" signs and at least one "Do not enter" per day, and I have to disengage at almost every one, unless I'm lucky enough to have another car in front of me (that doesn't decide to disobey the sign themselves).

 

[aronth5]

Does anyone else get this mapping error? I have 3 routes with this problem that require a disengagement everytime. The U-turn is simply turning around in the middle of the 2 lane road you just turned right on. Last mapping update didn't resolve the problem and of course I've let Tesla's FSDb team know.

 

[KirillK]

Elon kept his promise - as soon as I increased by safety rating to 81, my Model Y received FSD Beta. I was in the beta queue since August 2021.
I'm on 10.69.2.3 now and have been driving with FSD in different parts of Seattle for few hours.
It worked OK for 90% of the time. It even tried a pretty aggressive maneuver to get in front of the bus right before turning right on the intersection (ended OK). However, it can't go through roundabouts in 50% of the time - it starts and then competely stops in the middle, requiring me to disengage and listen for honks of drivers behind me. Let's see.

 

[Reddy Kilowatt]

I've been thinking about this ...

How does FSD improve disengagements ? Obviously by making fewer mistakes .... but how does the driver know whether a mistake is going to happen or not ? What we do is - guess. The guess is based on several factors - some can be easily characterized, like we think given the speeds and trajectories of all actors on the road, an accident might happen. Or we see that FSD is very indecisive ... jittery. But it all remains .... a guesswork. May be my guess has been faulty i.e. I've been disengaging when there wasn't a need. Afterall Omar disengages quite a bit less (and AFAIK) hasn't had any accidents. So, his guess is actually better than mine ?

In other words, one way disengagement rate goes down is when the testers feel less threatened. FSD can improve disengagement rate by gaining more trust.

Looks to me that is exactly what has happened. FSDb has gained a bit more trust from me. So, though it feels like I "lowered" my standard, in reality disengagement rate went down because it gained more trust.

Now, did FSDb gain trust by improving in some ways ? I do think so ... but I don't expect my trust level to be linearly correlated to actual FSDb ability. Its probably a step function ...

EV now I would agree with your statement above. I must be a little more trusting of FSD. Probably because I leave for work real early in the AM and mostly have the road to myself. I can let it do its thing. I do have my hands on the wheel when trying new actions or movements. Then when I'm in PM rush hour I'm not so nervous and ready to take over in a heart beat.

I do get surprises when the car screws up on things it has done before flawlessly. 69.2.3 has been the worse release to date in my opinion. I'm used to improvements not regressions. I this version I was surprised with several regressions and possibly on first runs of the route only to improve with repeated trips. I constantly run to work without any disengagements because I typically don't have to look out for the other guy, and can let it screw up. I have the same turn lane I have to play with the signal stalk to over ride but that is it. I believe Tesla is activating NN functions with each release and that's where the trouble began.

I had repeatedly gone through toll booths without issues. 69.2.3 comes out and the car is now following the white line trying to center when it missed the large stationary toll booth dead ahead. I was going to hit it dead on before disengaging at the last second. Has NN been given priority over fixed images in the library or what ever you call it? After that first trip the car has continued to do the same trip fine through the plaza then goes on it white line to the right hugging maneuver. I do typically get the " tool booth detected" pop up. I don't know if the map missed it on my first trip. Rising sun blinding cameras?

A new issue is multiple turn lanes. My car is no longer staying in its lane and I don't let it waver to much. I almost plowed into the side of the adjacent car and disengaged. I believe it is because they have increased turning speeds. As EV now noted, I have a low tolerance in this type of turn which is frequent. A lot of my disengagements have happened in these types of turns. I'm holding real tight ready to take over and the car thinks I want to.

I have a nasty 45MPH to 55MPH a tight right turn on red that I deal with every morning. No exit and or merge lane. I can have a wall of cars coming or nothing. One morning I had a wall of cars coming and very close to me, and the car went for it after creeping out. I slammed on the brakes and stopped. It had been doing this properly. This was a total surprise but I was ready. Its real fun to do when the light is green. I don't know if were going around at 15 or 45 MPH, they need to tune the speed when turning.

The last surprise I got was sitting at a signal on red. the turn lane signal went to a flashing yellow and we were red. The car just drove off like the light was green, but it was red. No sun light issue either. It flat screwed up and drove right towards opposing traffic turning left.

So after all these trying events I feel the system took a step backwards, and I lost some trust. I still use it daily thought. I have been running it for almost 30K miles.

Funny side note on how used to FSD I am. I went to Indy last weekend and had a brand new 2022 rental car and had to use Waze and Drive the car, I didn't know the area at all. I had to focus and listen to Waze, and drive the car! man o man I'm not used to such antiquated machinery. my driving skills have really dropped off. I couldn't even get used to the gas and brake pedals for the first day. The wife is bitching at me to drive normal, reverse case to FSD complaints. What can I say, I have been driving one pedal electric for nine years now. I didn't know you HAD to use the brake pedal.

 

[PACEMD]

I didn't know you HAD to use the brake pedal.

A what? What in the ham sandwich is a brake pedal?

 

[willow_hiller]

Well, this has my hopes up that 10.69.3 will behave differently from 10.62.2.3. Time will tell whether differently is better or worse:

https://twitter.com/x/status/1582729657350180865

 

[Silicon Desert]

Well, this has my hopes up that 10.69.3 will behave differently from 10.62.2.3. Time will tell whether differently is better or worse:

https://twitter.com/x/status/1582729657350180865

Yea, I am hoping for some improvement on a few things. While on the subject, here is one of my nagging items. Yes, it is the traffic circle again (roundabouts for our English friends). Although the latest FSDb version seems to finally handle them well in my area, it insists on putting on the right turn signal light as it goes around the circle, even when the intent is to go straight ahead and not turn right.

You know what might happen in this case..... if a car is at the street on there right (where the black pickup truck is in my photo), that driver might think I am going to turn right and into his road instead of going straight. Some people make bad assumptions about what a signal means and like to start pulling out in front of my car because they are too impatient to wait 3 seconds more to see if my car actually goes straight or turns. First time that happened, I stopped using FSD on that traffic circle in traffic. If no other cars are in the area, I turn on FSD to confirm it does the same thing every time with turning on the right turn signal for no reason but to confuse other drivers.

 

[Silicon Desert]

A what? What in the ham sandwich is a brake pedal?

Hmmm, good question. I think it might be that big square thing on the floor board where my wife put a "Do Not Touch" sign on it

 

[Dewg]

Yea, I am hoping for some improvement on a few things. While on the subject, here is one of my nagging items. Yes, it is the traffic circle again (roundabouts for our English friends). Although the latest FSDb version seems to finally handle them well in my area, it insists on putting on the right turn signal light as it goes around the circle, even when the intent is to go straight ahead and not turn right.

You know what might happen in this case..... if a car is at the street on there right (where the black pickup truck is in my photo), that driver might think I am going to turn right and into his road instead of going straight. Some people make bad assumptions about what a signal means and like to start pulling out in front of my car because they are too impatient to wait 3 seconds more to see if my car actually goes straight or turns. First time that happened, I stopped using FSD on that traffic circle in traffic. If no other cars are in the area, I turn on FSD to confirm it does the same thing every time with turning on the right turn signal for no reason but to confuse other drivers.

I have the same issue with people who want turn signals to come on much earlier than 100 feet before an intersection. There are many retail centers around me, some with entrance/exits around 100 feet from the intersection. If the car turns on the signals at 200 or 300 feet from the intersection, cars who are coming out of the retail center may think I'm going to turn into the center and not at the intersection and proceed in front of me. Just confuses people.

 

[AlanSubie4Life]

Well, this has my hopes up that 10.69.3 will behave differently from 10.62.2.3. Time will tell whether differently is better or worse:

https://twitter.com/x/status/1582729657350180865

I predict it will be better. However, to the end user, there is a decent chance it will seem worse even though it is better.

 

[AlanSubie4Life]

I have the same issue with people who want turn signals to come on much earlier than 100 feet before an intersection. There are many retail centers around me, some with entrance/exits around 100 feet from the intersection. If the car turns on the signals at 200 or 300 feet from the intersection, cars who are coming out of the retail center may think I'm going to turn into the center and not at the intersection and proceed in front of me. Just confuses people.

This is weird. Obviously people are not asking for that. Obviously the car only signals early when appropriate. Just like a good human driver. Signaling early doesn’t mean you always signal early!!! Not complicated to implement - if there is going to be ambiguity, you do not create ambiguity.

 

[Silicon Desert]

This is weird. Obviously people are not asking for that. Obviously the car only signals early when appropriate. Just like a good human driver. Signaling early doesn’t mean you always signal early!!! Not complicated to implement - if there is going to be ambiguity, you do not create ambiguity.

Yo Alan, help me out on that one. I read it about 4 times and still missed the point of the point. Not your fault. Guess I am just having a brain dead logic day.

 

[AlanSubie4Life]

Yo Alan, help me out on that one. I read it about 4 times and still missed the point of the point. Not your fault. Guess I am just having a brain dead logic day.

This was a prior discussion so you probably missed the context.

So if signaling early creates ambiguity (because there is another intersection in advance of your intended turning point that people may interpret as you signaling for - and then they can turn out in front of you), then don’t do that, or do it at the last second as you pass that intersection.

“Advanced” technique is you can signal early if you know no one is at the non-intended intersection and no one will arrive there until after you pass.

Just standard stuff that every good driver thinks about as they drive.

Additional caution and procedures might be followed if you know someone is following you too closely. (You might signal early and watch turning traffic very carefully, moving your car to the left of your lane when turning right, to signal to people that you are not turning where they might think you are. Have to use that body language.)

 

[Silicon Desert]

This was a prior discussion so you probably missed the context.

So if signaling early creates ambiguity (because there is another intersection in advance of your intended turning point that people may interpret as you signaling for - and then they can turn out in front of you), then don’t do that, or do it at the last second as you pass that intersection.

“Advanced” technique is you can signal early if you know no one is at the non-intended intersection and no one will arrive there until after you pass.

Just standard stuff that every good driver thinks about as they drive.

Additional caution and procedures might be followed if you know someone is following you too closely.

Ah thanks. now I get it. I do as you describe, which is why I am annoyed at the Tesla putting on the signal light in a traffic circle. Even worse than signaling early, there is no intent for the car to actually turn at all. About a hundred feet later after the circle, it turns off the signal

 

[kabin]

This morning FSDb choked going straight through an intersection from a red to green light change. I was lead vehicle when the light turned green. FSDb slowly started, stopped hard, dawdled for a couple of seconds, started again, and then stopped hard, dawdled again for several seconds until I took over. Fortunately no one was following.

It might have something to do with the relatively deep dip (rain/water channel) perpendicular to vehicle travel. I'm imagining camera processing not knowing how to proceed after the center of camera view rapidly changes from viewing pavement and then sky but who knows. I bet there aren't many training sets with this scenario.

The google cat-eye lens image below doesn't do the dip justice. Moving vehicles might need to slow to 15mph to avoid bottoming out suspension although not an issue when you are a lead vehicle with a red light.

Butherus Dr/Scottsdale Rd

 

[Dewg]

This morning FSDb choked going straight through an intersection from a red to green light change. I was lead vehicle when the light turned green. FSDb slowly started, stopped hard, dawdled for a couple of seconds, started again, and then stopped hard, dawdled again for several seconds until I took over. Fortunately no one was following.

It might have something to do with the relatively deep dip (rain/water channel) perpendicular to vehicle travel. I'm imagining camera processing not knowing how to proceed after the center of camera view rapidly changes from viewing pavement and then sky but who knows. I bet there aren't many training sets with this scenario.

The image below doesn't do the dip justice. Cars might need to slow to 15mph to avoid bottoming out suspension.

Butherus Dr/Scottsdale Rd

Was the stopping and dawdling happening while transitioning over the dip? Like the front tires when into the dip and the car stopped, then when the back tires when into the dip, it stopped again?

 

[kabin]

Was the stopping and dawdling happening while transitioning over the dip? Like the front tires when into the dip and the car stopped, then when the back tires when into the dip, it stopped again?

I'll have to verify that but I suspect so as I wasn't too far away from the dip when I needed to take over.

 

[willow_hiller]

It might have something to do with the relatively deep dip (rain/water channel) perpendicular to vehicle travel. I'm imagining camera processing not knowing how to proceed after the center of camera view rapidly changes from viewing pavement and then sky but who knows. I bet there aren't many training sets with this scenario.

FSDb does seem to struggle a bit with reflections at the moment. One day I happened to have a puddle placed exactly between my vehicle and a large metal pole, such that the reflection of the pole was visible in the puddle. While waiting at the light, FSDb rendered a grey obstruction in the middle of the intersection, right where the puddle was reflecting the pole. It disappeared when the vehicle proceeded through the intersection, as the angle sufficiently changed to reflect something else.

 

[aronth5]

I have the same issue with people who want turn signals to come on much earlier than 100 feet before an intersection. There are many retail centers around me, some with entrance/exits around 100 feet from the intersection. If the car turns on the signals at 200 or 300 feet from the intersection, cars who are coming out of the retail center may think I'm going to turn into the center and not at the intersection and proceed in front of me. Just confuses people.

I am one who believes blinkers come on too late on many turns especially when someone is behind you. However, I agree when you have entrances/exits close together that would be incorrect behavior. Humans adjust for this difference so FSD should as well.