FSD Beta 10.69

[archae86]

I’m pretty sure it is looking at user-applied torque. <snip>

I think so too, and I think the amount is not varying all over the place. User perception, is, for sure, but I think it far more likely that the Tesla software plus hardware is employing a consistent torque threshold than that all those wildly varying user impressions are correct.

Going further, I suspect nearly all cases of reported sudden disengagement are not initiated by the car, but are circumstances in which the driver applied torque over the breakout amount.

 

[bhakan]

Try switching your screen from nighttime to daytime mode, which will brighten the cabin with some light. You can adjust the brightness so it's not as blinding, but the camera should see you better. The goal is still to replace the camera, but you may be able to use FSD at night while you work with the SC team.

The brightness is always 100%. I changed the Appearance (the option for refresh MSLR) to Day from Auto. I will see if that helps. Thx. Appearance was always on Auto in all the previous FSDb software and I did not have any problem. Not sure if software 69.2.2 tightened the cabin camera view.

 

[thewishmaster]

Funny behavior, not strictly 10.69-related - ever since they added logic for dealing with open car doors, the beta freaks out and carves a wide path around one of those Ford or Mercedes vanlife type vans with a black ladder on the side street parked in my neighborhood. I think it keeps misinterpreting the ladder as an open door, which leads to regular disengagements due to how narrow and busy that stretch of road is.

 

[nvx1977]

@sleepydoc

I think so too, and I think the amount is not varying all over the place. User perception, is, for sure, but I think it far more likely that the Tesla software plus hardware is employing a consistent torque threshold than that all those wildly varying user impressions are correct.

Going further, I suspect nearly all cases of reported sudden disengagement are not initiated by the car, but are circumstances in which the driver applied torque over the breakout amount.

There are a few left turns around here that go wider than expected and encroaches into the right-adjacent lane, also turning left. Here, I can get FSDb to disengage with a super-light torque application. It's so light that it's very hard to keep it from disengaging if you are lightly holding onto the wheel and let it do what it wants to do.

I don't think that variations in disengagement torque is really disputable at this point.

The other idea is that the required disengagement torque is tied to the confidence of FSDb at that moment. I haven't tested this enough to feel confident in this idea.

But torqueing with or counter to FSDb is very easy for anyone to test. Just find a winding, curvy road with little to no traffic. Then disengage via wheel with the curve and against the curve. You will feel the difference in resistance. If you don't have curvy roads, then you'll have to do it at a right or left turn (with little/no traffic for safety)

 

[willow_hiller]

Tried my first drive in downtown DC today. It was rough, but I think most of the issues can be boiled down to lane selection.

Most of our boulevards are 3 lanes each direction. But for the most part, you do not drive in the left or the right lane unless you really need to. The left lane piles up with drivers making unprotected lefts, and the right lane is either full of parked cars, or trucks stopped making deliveries. So my car got stuck behind left-turning drivers a few times. And it would have gotten stuck behind parked cars in the right lane multiple times if I didn't repeatedly tap to cancel the lane change.

And then if FSDb struggles with regular circles, imagine how it did on this one:

 

[legendsk]

It may be a coincidence, but with the 2022.16.8 software update a new problem appeared. In the morning, if I try to drive southbound or in the afternoon driving north, the message appears: The left door pillar camera is blinded / blocked, autosteer, autopilot, and lane change are not available. I created a service appointment to resolve this and got the response shown below.

Apparently the car only needs to work on cloudy days or when driving in a direction where the sun doesn't shine on the left door pillar? I reckon it should be okay at night and I can reschedule accordingly? The service improvement seems to be limited to improved response - they will now not fix it immediately instead of taking 3 weeks to respond?

 

[sleepydoc]

Tried my first drive in downtown DC today. It was rough, but I think most of the issues can be boiled down to lane selection.

Most of our boulevards are 3 lanes each direction. But for the most part, you do not drive in the left or the right lane unless you really need to. The left lane piles up with drivers making unprotected lefts, and the right lane is either full of parked cars, or trucks stopped making deliveries. So my car got stuck behind left-turning drivers a few times. And it would have gotten stuck behind parked cars in the right lane multiple times if I didn't repeatedly tap to cancel the lane change.

And then if FSDb struggles with regular circles, imagine how it did on this one:

View attachment 858072

Yay! Minnesota Ave!

Driving issues like you describe are hard. They can be very location-specific and what FSD did wasn’t technicallywrong, it was just suboptimal. Most urban areas have spots like that where you learn the best way to navigate after you’re more familiar with the area.

Edit: i also just noticed ‘wibble wobble child development center’ - made me laugh!

 

[sleepydoc]

There are a few left turns around here that go wider than expected and encroaches into the right-adjacent lane, also turning left. Here, I can get FSDb to disengage with a super-light torque application. It's so light that it's very hard to keep it from disengaging if you are lightly holding onto the wheel and let it do what it wants to do.

I don't think that variations in disengagement torque is really disputable at this point.

But torqueing with or counter to FSDb is very easy for anyone to test. Just find a winding, curvy road with little to no traffic. Then disengage via wheel with the curve and against the curve. You will feel the difference in resistance. If you don't have curvy roads, then you'll have to do it at a right or left turn (with little/no traffic for safety)

Without actual testing and making some quantitative measurements I don’t think anyone can say for sure. I’ve noticed that 69.2.2 seems like it takes less force in general but again, that’s totally subjective.

The other idea is that the required disengagement torque is tied to the confidence of FSDb at that moment. I haven't tested this enough to feel confident in this idea.

That’s an interesting thought. I haven’t specifically notice it but up until the latest version most of the times FSD was really unsure the steering wheel was jerking wildly back and forth so it was really hard to gauge force. It would certainly make sense to design the system that way.

Like I said above, I’ll often let the steering wheel ’drag’ between my thumb and fingers when FSD is turning. That puts friction resistance/torque on it without me needing to try and guess what the wheel will do. I haven’t noticed that technique causing more disengagements, for whatever that’s worth.

 

[VanFriscia]

It may be a coincidence, but with the 2022.16.8 software update a new problem appeared. In the morning, if I try to drive southbound or in the afternoon driving north, the message appears: The left door pillar camera is blinded / blocked, autosteer, autopilot, and lane change are not available. I created a service appointment to resolve this and got the response shown below.

Apparently the car only needs to work on cloudy days or when driving in a direction where the sun doesn't shine on the left door pillar? I reckon it should be okay at night and I can reschedule accordingly? The service improvement seems to be limited to improved response - they will now not fix it immediately instead of taking 3 weeks to respond?

Blinding sun is one of several issues that Tesla has yet to address. Add to that the inability of the cameras to clean themselves, and also possibly the very placement of the cameras being sub-optimal in the first place. I really don't know how they could say all their vehicles of the last several years have the required hardware for self-driving robotaxis when these very basic and common items are still an issue on every single Tesla vehicle out there.

 

[willow_hiller]

Blinding sun is one of several issues that Tesla has yet to address. Add to that the inability of the cameras to clean themselves, and also possibly the very placement of the cameras being sub-optimal in the first place.

I was actually pleasantly surprised how FSDb handles blinding sunlight on my vehicle. I've never seen it cease to drive while being blinded, but I have seen it slow down ~5 MPH below my set speed when driving directly towards the sun.

And the front cameras can clean themselves. That's probably sufficient for a level 4 system that can detect when the other cameras are sufficiently dirty and find a safe location to pull over before alerting the owner it needs help.

 

[VanFriscia]

I was actually pleasantly surprised how FSDb handles blinding sunlight on my vehicle. I've never seen it cease to drive while being blinded, but I have seen it slow down ~5 MPH below my set speed when driving directly towards the sun.

And the front cameras can clean themselves. That's probably sufficient for a level 4 system that can detect when the other cameras are sufficiently dirty and find a safe location to pull over before alerting the owner it needs help.

In my experience, it will USUALLY drive in blinding sun, but on occasion it gives me a takeover message or refuses to engage. I've never seen it slow down for this though, so that's interesting to hear.

Also interesting - the front cameras clean themselves? How/when do they do this??? I suppose the only one I've ever had to clean myself is the rear camera. I just assumed that's because it's in a position to get especially dirty, not because some of the others took care of their own business!

 

[sleepydoc]

I was actually pleasantly surprised how FSDb handles blinding sunlight on my vehicle. I've never seen it cease to drive while being blinded, but I have seen it slow down ~5 MPH below my set speed when driving directly towards the sun.

And the front cameras can clean themselves. That's probably sufficient for a level 4 system that can detect when the other cameras are sufficiently dirty and find a safe location to pull over before alerting the owner it needs help.

Ditto. Of course, if the sun causes horrible glare on the window of one of the B pillar cameras then it will be an issue but in general it seems to handle the sun much better than I’d expect.

 

[willow_hiller]

Also interesting - the front cameras clean themselves? How/when do they do this??? I suppose the only one I've ever had to clean myself is the rear camera. I just assumed that's because it's in a position to get especially dirty, not because some of the others took care of their own business!

Yes, since 10.2 or so, I believe. FSD Beta must have a "dirty-windscreen" parameter it estimates.

https://www.reddit.com/r/teslamotors/comments/q697zk

 

[JulienW]

Another great day for 69 drives. It is sooooooo strange for me. So far 69 either drives like a near champ all day every drive or fails over and over and over.....all day every drive. it is almost like Tesla is flipping switches in 69 testing different parameters OR f$$$ing with me.

Now that I'm home what will tomorrow be? Please Elon leave it like it is today. I know you are entertained watching me get frustrated.

 

[Dewg]

Another great day for 69 drives. It is sooooooo strange for me. So far 69 either drives like a near champ all day every drive or fails over and over and over.....all day every drive. it is almost like Tesla is flipping switches in 69 testing different parameters OR f$$$ing with me.

Now that I'm home what will tomorrow be? Please Elon leave it like it is today. I know you are entertained watching me get frustrated.

A/B testing?

 

[JulienW]

A/B testing?

Then I just want to be an A tester. Let others do the B thing.

 

[bobsmy]

After more testing on FB I have the following thoughts. it appears that a large number of the serious FB incidents occur when the car encounters an road, driveway,parking lot entrance exit from the right even if there are no vehicles on the intersecting roadway. The braking varies from a slight slowing to a full stop in the travel lane. There seems to be no criteria for when FSD ignores the intersecting road and when it causes PB. The other instance of FB occurs when on a winging road and making an inside turn with oncoming traffic from the other direction even though the oncoming vehicle is clearly on the proper side of the double yellow lines, In one instance the lines were painted only two days ago. Both of these appear to be FSD being overly cautious.

 

[2101Guy]

This one is on the way for me to test with the yoke. Seller claims it will work and if it doesn’t I can return it. Will report back

https://www.ebay.com/itm/234610491237

I can confirm, that this one DOES work with the Yoke. Its one of those bendy/twisty type of designs. Wrap around the lower right corner of the yoke snugly and it stays in place/doesnt move. Now admitedly, not the best design but all we have until this version comes out for the yoke (Used this on my prior S for 2 years straight and worked flawlessly):

Tailgate Car Backup Reverse Handle Camera For Dodge Ram 1500 2500 3500 | eBay

Find many great new & used options and get the best deals for Tailgate Car Backup Reverse Handle Camera For Dodge Ram 1500 2500 3500 at the best online prices at eBay! Free shipping for many products!

www.ebay.com

As always, you, the driver, are 100% accountable/responsible for the actions of your car whether on AP, off of AP, on FSD, FSD(B), FSDb, and any of the other names of the Tesla driver aids that require full attention at all times. You kill someone with one of these on your wheel? Same penalty as killing someone without one.

Please feel free to reply back with how horrid and dangerous this is (while you quietly and secretly order one).

Your Welcome.

 

[willow_hiller]

Building off my data upload post from yesterday, the amount uploaded is surprisingly inconsistent given the length of the drive or the number of disengagements.

Yesterday, my Model 3 uploaded 2.699 GB after a 4 mile, 16 minute drive with one tap of the report button after one disengagement.

Today, I did a total of 24 miles over 1 hour and 12 minutes, with at least two taps of the report button and many many disengagements. And... 2.3 GB upload.

I figured the downtown disengagement data would be a lot larger than a quick trip in the suburbs, but maybe not?

 

[nvx1977]

Building off my data upload post from yesterday, the amount uploaded is surprisingly inconsistent given the length of the drive or the number of disengagements.

Yesterday, my Model 3 uploaded 2.699 GB after a 4 mile, 16 minute drive with one tap of the report button after one disengagement.

Today, I did a total of 24 miles over 1 hour and 12 minutes, with at least two taps of the report button and many many disengagements. And... 2.3 GB upload.

View attachment 858274

I figured the downtown disengagement data would be a lot larger than a quick trip in the suburbs, but maybe not?

The bulk of the data is from the video, and the compressibility of the video will vary widely on sunny vs cloudy, day vs night, low vs high traffic, trees vs no trees, city vs rural, etc. As an example, night time videos compress well because there's way fewer important pixels to preserve.