FSD Beta 10.69

[pa28pilot]

Couple observations from my drives:

1. Poor lane selection in congested areas: when making a left or right turn at an intersection, if the lane has heavy/backed up traffic, it will attempt to go around them and fail the turn. it will literally stop and try to cut in front of all the cars waiting in the lane. seems like an easy fix if they can predict the lane they need to be in sooner(say 1000-1500ft before the intersection), and ease up on the passing slow traffic logic.

2. I haven’t had one pass in double/multiple intersections. If they’re both green or both red, the car navigates correctly. if the first light is green, and the second is red, the car will stop at the first intersectio. Seems as if the ap team has some stop light depth perception work to on stop lights.

3. right turn on red with an intersection that has double left hand turn lanes across from the car. It will not wait for the cars on the outer left turn lane.

4. poor lane selection and positioning on circular overpasses where merging is important(see image below). The car will get way to close to the right, super close to the wall and fail the merge.

all in all I’m very impressed with FSP beta 10.69.2.2 and the errors I found seem like easy fixes for the ap team.

 

[AlanSubie4Life]

That's actually pretty advanced. I'll typically gauge my braking so I hit the light as it turns green (or so I approach the cars in front of me right as they're speeding up.) Doing so requires looking at the light, looking at the number of cars waiting, estimating how long the light will take before it turns and also knowing the pattern of the light (does the turn lane go first or last? When does your lane turn green relative to the other lanes?)

Doesn't seem all that advanced. But anyway it's a side issue from the stopping thing. Having this would just slightly improve efficiency and make for more logical behavior. But having it would not guarantee smoother stops, since the issue there isn't the stopping distance.

I had a case yesterday where it kept driving up to a bunch of cars waiting at a light and waited far too long to brake, necessitating a pretty aggressive stop. I'd be happy with a more reasonable braking distance.

For sure. Having a longer braking distance and not staying on the accelerator when the light is obviously red (and I'm pretty sure the car can see it...) would be nice!

Still, a separate issue is the smoothness of the stop, which has been discussed a lot. A quite short and fast stop can still be quite smooth. It's all about the execution.

Right now it is both failing on the distance (which would help with optimizing efficiency), and also failing on the execution of the slowdown (which would be less of an issue if it could slow in a greater distance - but would still be an issue since the body is so sensitive to jerk).

Two distinct and separate problems.

I'm pretty sure this was being used in some form before too. FSDb would jog around unvisualized street poles for instance, and also avoid those outdoor eating setups (also not visualized)

I agree that it's not clear exactly what was being used. However, some of the evidence from failures on Chuck's turn suggested that it really didn't understand what it couldn't see. It would proceed when a turning car obscured oncoming traffic, for example. There may be other examples too.

There's a difference between detecting obstacles and "understanding" that they obscure what exists in a region behind the object.

Another capability that I didn't add above which I don't think exists is a substantial memory. Hopefully we can get up to at least 5-10 seconds, and possibly a lot lot longer in stationary situations. I'm not sure what is required; it could be substantial, to cover all the cases! Anyway that's a fundamental capability we'll need before we should expect anything close to reasonable behavior. So wait for all those capabilities to be demonstrated, and then maybe start expecting good behavior. Before that, expect it to be pretty poor.

If it doesn't have capabilities that humans have, it's not going to be very good. Even with excellent reaction time. That's another thing to look out for - it's very unclear what the reaction time of FSD is right now. Some instances would suggest it's quite poor, but it's really very difficult to say since we don't have any way to see what the car sees. Just really hard to judge without all the information.

Really the only advantage it has right now over humans is that sometimes humans are compromised or are not paying any attention to what they're doing. Right now it seems extremely clear that any advantage it has does not outweigh its deficiencies.

 

[Reddy Kilowatt]

I got to drive about 200 miles with the new 69.2.2, and am very impressed. It has wiped out about 95% of my issues. I'm very impressed on how refined it is now.

The biggest issue I have left is FSD jumping to quick into a right hand turn lane. It always seems to be a block to soon. I know to just over ride with turn stalk. I little to no disengagements. A tap or two of the gas pedal or turn stalk is about all I need to do or assists. These are routes I have done endlessly through the program.. I've started letting the car drive in parking lots if traffic is lite. School buses and School zones are ignored. That would be the next library they should build along with speed zone reductions and exit ramp speeds.

My large u turn has been hit or miss. I think it is more me not wanting to get rear ended and overriding it. I has done it when I give it time to fidget.

Nav still need to be fixed.

 

[wknickless]

It seems challenging to reproduce the exact same bugs in a new architecture though.
Unless most of the bugs are intentional...

Elon said the old architecture had reached a point beyond diminishing returns for more training. However, the same training set was likely used for both the old and new architecture. So it's reasonable to anticipate similar behaviors for the first rollout of the new architecture, except (possibly) for the use cases that motivated the new architecture in the first place

For example, this morning the new architecture did a great job handling an unprotected high speed cross traffic intersection, where earlier versions routinely bailed and said "hit the accelerator when you think it's safe". But sadly it's still making other mistakes when the map data appears to be misleading.

 

[AlanSubie4Life]

Elon said the old architecture had reached a point beyond diminishing returns for more training. However, the same training set was likely used for both the old and new architecture. So it's reasonable to anticipate similar behaviors for the first rollout of the new architecture, except (possibly) for the use cases that motivated the new architecture in the first place

For example, this morning the new architecture did a great job handling an unprotected high speed cross traffic intersection, where earlier versions routinely bailed and said "hit the accelerator when you think it's safe". But sadly it's still making other mistakes when the map data appears to be misleading.

I don't think there's much evidence they use training for stopping and going. It's hard to know for sure of course; there are probably gray areas.

 

[STUtoday]

Do you work in construction? I'd be interested in seeing a milestone payment schedule or SOV that has 50% completion for a slab on grade and rough framing. You don't claim significant progress for submitting permit applications and then waiting for approval, and design costs scale with the complexity of construction that follows. Physical work in the field is what consumes the biggest portion of $$$, and detailed finishing is what takes the most time.

There are plethora parallels to draw between FSD's progress and this comic

Tangentially, yes.
No-one (you/me/comic) was talking progress payments, just overall timeframe. And from frame to completion would be several months out of a process of likely a couple of years. Your thinking here seems quite narrow/ignorant.
Either way you don't need the extra bluster to try to cover for your mistakes. I didn't mean you any particular malice. Your comment was probably even just a quick, throw away comment. It's just that it was sloppy and inaccurate.
If you're going to make brash/bold statements best just stick to topics you (think you) know about which I presume/hope isn't construction.

 

[LowlyOilBurner]

The last three days, I’ve had “Whole Mars Catalog” drives. No interventions or disengagements. Even did a “Chuck style” unprotected left turn, used the center turn lane, did great. A couple odd things that need some work, but all in all, pretty decent.

 

[dckiwi]

Still stuck on .15 (69.2). Less than 5% of beta testers still on this version. Anyone else been left on a prior FSD release? Curious if/when I will get it.

 

[aronth5]

The last three days, I’ve had “Whole Mars Catalog” drives. No interventions or disengagements. Even did a “Chuck style” unprotected left turn, used the center turn lane, did great. A couple odd things that need some work, but all in all, pretty decent.

This is great.
I'm finding 1069.2.2 is one step forward and one step back. Some nice improvements but also some significant regressions.
Just a couple of examples, FSD cannot make the turn into my street which was fixed 10 months ago. Starts the turn then realizes it won't succeed and aborts by stopping 100 feet past the turn. FSD also fails to stop properly 50% of the time at a particular stop sign. Slow down then stops in the middle of the crossing road. (when I let it to see what would happen!).

 

[LowlyOilBurner]

This is great.
I'm finding 1069.2.2 is one step forward and one step back. Some nice improvements but also some significant regressions.
Just a couple of examples, FSD cannot make the turn into my street which was fixed 10 months ago. Starts the turn then realizes it won't succeed and aborts by stopping 100 feet past the turn. FSD also fails to stop properly 50% of the time at a particular stop sign. Slow down then stops in the middle of the crossing road. (when I let it to see what would happen!).

See, I haven’t been on FSD since 10.11, so 10.69.2.2 is a leap forward for me. Like I said, still some oddball issues here and there, things that did work better on 10.11 (like interacting with pedestrians I thought was better on 10.11), are worse on 10.69.2.2.

Big plus for me living in a rural area:
-unmarked roads can now be used at the correct speed limit, before, max speed was 27mph, and lane keeping on unmarked roads is way better than before. Almost better than some roads that have lane markings.

 

[diplomat33]

Almost got rear ended on FSD Beta today (10.69.2.2). FSD Beta did not move over to the left turn lane when it should have. When it was almost at the stop line, it realized that it needed to move over so it slammed on the brakes. It almost caused a rear end collision as the truck behind me did not expect me to brake so hard. I disengaged immediately and moved over to the left lane. I don't know why FSD Beta does not seem to detect the left turn lane anymore. On previous versions, it used to move over very quickly as soon as the left turn lane starts. Now, it seemingly does not recognize the left turn lane anymore and will lane change at the last second.

 

[Dewg]

Almost got rear ended on FSD Beta today (10.69.2.2). FSD Beta did not move over to the left turn lane when it should have. When it was almost at the stop line, it realized that it needed to move over so it slammed on the brakes. It almost caused a rear end collision as the truck behind me did not expect me to brake so hard. I disengaged immediately and moved over to the left lane. I don't know why FSD Beta does not seem to detect the left turn lane anymore. On previous versions, it used to move over very quickly as soon as the left turn lane starts. Now, it seemingly does not recognize the left turn lane anymore and will lane change at the last second.

I had a similar experience on a drive on Thursday - yesterday it drove fine. The planner needs some love from Telsa quickly.

 

[Charliek]

10.69.2.2 looks good for me in VA.. 3 items that were bugging me on unlined suburban streets are fixed and clean. I am working to tweak my "speed above the Speed limit" to match the prevailing traffic just a little bit more. still a few spots it goes a little faster than I would like heading into a neighbor hood turn in from a larger road. also switching from FSD b to FSD NOA high way stack slows down to the "switch oversetting" and I have to notch it back up to prevailing traffic ( 59 to >68MPH .people go too fast on the toll road in reston va ). That switch over catches me like it like it is a phantom brake but it isnt.. no significant hesitations or phantom breaking under any light conditions. GUYS KEEP THESE SETTINGS!.

I have taken to not adjusting speed- except the mode switchover from the onramp to big highway _( quick plus ups to notch downs) at all to see how the car "naturally" selects the speed and the lane and I must say this version is very impressive.. I also have taken to starting it in parking lots for a complete door to door experience ( it picks a nice speed in the parking lot and slowly goes over speed bumps - saving my suspension, Thankyou!) and it does it well 99 percent of the time with other cars. I think we may be heading the 3 9's percent reliability shortly!. not ready to take a nap ( thats 5 9's), but it is quite enjoyable just tooling around point for errands in FSDb. I haven't really " driven" the car since this load!. Best version yet! and that is coming from this 2017 MS owner ( fully upgraded of course).

 

[Mardak]

FSD Beta 10.69.2.2 definitely wasn't slowing down for the yellow and potentially not even for the red light, so I needed to disengage around here as the visualization was showing bright green traffic lights:

As you can probably tell from the car's own shadow, the sun was pretty low, and FSD Beta was already getting confused at unlit yellow school lights but brightened by the sun angle. Here's a view from the back at the traffic light, and even then it's hard to see the next car:

Definitely a corner case as the sun setting only lines up with this road only for a few days of the year and happens only for a few minutes each of those days. On the flip side, I've noticed that FSD Beta actually does fine even with the sun directly behind a traffic light as there's multiple traffic lights to look for and doesn't mind "staring at the sun."

 

[Charliek]

FSD Beta 10.69.2.2 definitely wasn't slowing down for the yellow and potentially not even for the red light, so I needed to disengage around here as the visualization was showing bright green traffic lights:
View attachment 856442

As you can probably tell from the car's own shadow, the sun was pretty low, and FSD Beta was already getting confused at unlit yellow school lights but brightened by the sun angle. Here's a view from the back at the traffic light, and even then it's hard to see the next car:
View attachment 856443

Definitely a corner case as the sun setting only lines up with this road only for a few days of the year and happens only for a few minutes each of those days. On the flip side, I've noticed that FSD Beta actually does fine even with the sun directly behind a traffic light as there's multiple traffic lights to look for and doesn't mind "staring at the sun."

they switched in Virginia to bright yellow reflective traffic light backing edges to provide more contrast in cases like this . this time of year we have the same problem with our roads being very east west facing ( bad in the morning sun up and the evening commute sun down) I did notice it did not complain or down select FSDb in either case on this load. much better adn safer than my squinting!

 

[JulienW]

Beta kerfuffled this intersection twice from different directions. Was heading west on 8th St (blue) that “ends” and merges with Marietta St. Nav said “right Turn on Marietta St" but the Nav line was straight. Light red and Beta stopped and turned on right signal. It then proceeded to make a STRAIGHT “right turn” on red. About ½ way through I yanked the wheel right and turned on Brady Ave to keep from running the red light.

Made a U-turn and Nav was to then turn right on Marietta St. Had signal on but proceeded to drive straight across (green).

Seems the angled off of Marietta St was too much for it to understand. Odd since this is not an unusual edge case.

 

[Charles Guerin]

FSD Beta 10.69.2.2 definitely wasn't slowing down for the yellow and potentially not even for the red light, so I needed to disengage around here as the visualization was showing bright green traffic lights:
View attachment 856442

As you can probably tell from the car's own shadow, the sun was pretty low, and FSD Beta was already getting confused at unlit yellow school lights but brightened by the sun angle. Here's a view from the back at the traffic light, and even then it's hard to see the next car:
View attachment 856443

Definitely a corner case as the sun setting only lines up with this road only for a few days of the year and happens only for a few minutes each of those days. On the flip side, I've noticed that FSD Beta actually does fine even with the sun directly behind a traffic light as there's multiple traffic lights to look for and doesn't mind "staring at the sun."

I had the same thing happen, same conditions.

Most of my drives have been uneventful but I may have discovered another edge case — anyone are try this? :

When pulling out of a parking lot into a 2-4 lane street, position the car at a right angle to the street and turn on FSD.

Twice now the car has gone in a straight line rather than turning into the street, the first time directly into a bus that was passing. 2nd time it was aiming at the opposite curb.

Both times I disengaged quickly. I think it couldn’t tell I was entering a street and it need to turn either right or left.

FSD Beta 10.69.2.2 definitely wasn't slowing down for the yellow and potentially not even for the red light, so I needed to disengage around here as the visualization was showing bright green traffic lights:
View attachment 856442

As you can probably tell from the car's own shadow, the sun was pretty low, and FSD Beta was already getting confused at unlit yellow school lights but brightened by the sun angle. Here's a view from the back at the traffic light, and even then it's hard to see the next car:
View attachment 856443

Definitely a corner case as the sun setting only lines up with this road only for a few days of the year and happens only for a few minutes each of those days. On the flip side, I've noticed that FSD Beta actually does fine even with the sun directly behind a traffic light as there's multiple traffic lights to look for and doesn't mind "staring at the sun."

I had the same experience with the same conditions. Most of my drives were flawless, but I think I found another edge case.

If you exit a parking lot into a two or 4 lane street and set up the car perpendicular to the street, then turn on FSD.

Two times doing this, the car lurched straightforward into traffic aiming at the opposite curb. The first time towards a moving bus. Both times I caught it quickly because it wasn’t making a turn like it was supposed to, but I do think it’s an edge case that’s repeatable because it’s happened a few times.

My guess is that the car doesn’t realize it’s direction or compass heading at that time and thinks it’s on the street.

 

[kabin]

Almost got rear ended on FSD Beta today (10.69.2.2). FSD Beta did not move over to the left turn lane when it should have. When it was almost at the stop line, it realized that it needed to move over so it slammed on the brakes. It almost caused a rear end collision as the truck behind me did not expect me to brake so hard. I disengaged immediately and moved over to the left lane. I don't know why FSD Beta does not seem to detect the left turn lane anymore. On previous versions, it used to move over very quickly as soon as the left turn lane starts. Now, it seemingly does not recognize the left turn lane anymore and will lane change at the last second.

It's a bit random for me but it seems to want to carry the road's set speed velocity to the bitter-end and then make a hasty turn decision while braking in that last 100 ft area. Acts a bit like a distracted driver with no time to transition from driving straight to prepping for a left turn.

 

[aronth5]

FSD completely disengaged going around a 90 degree bend and headed straight for the sidewalk where the orange cones are. This has happened twice. Yesterday I figured I must have grabbed the wheel causing the disengagement. 10.69.2.2 doesn't handle this curve as well as prior versions so I pay extra close attention taking the curve now. Today I specifically had both hands an inch off the wheel and FSD simply disengaged and drove straight towards the sidewalk again. Braked and grabbed the wheel to make it thru the turn safely. Another example of a regression in 10.69.2.2

Note street view doesn't do the curve justice. Much sharper than it appears.

Google Maps

 

[Charliek]

FSD completely disengaged going around a 90 degree bend and headed straight for the sidewalk where the orange cones are. This has happened twice. Yesterday I figured I must have grabbed the wheel causing the disengagement. 10.69.2.2 doesn't handle this curve as well as prior versions so I pay extra close attention taking the curve now. Today I specifically had both hands an inch off the wheel and FSD simply disengaged and drove straight towards the sidewalk again. Braked and grabbed the wheel to make it thru the turn safely. Another example of a regression in 10.69.2.2

Note street view doesn't do the curve justice. Much sharper than it appears.

Google Maps

definitely send them a clip of that one may even email them too! it should still be able to see the edge of the street( maybe it thinks it is all one lane and is taking its half out of the middle. Then maybe map doesnt have enough info or some thing