First FSD Beta accident?

[nvx1977]

I can see where someone might believe FSD was at fault. Twice when I have needed to immediately take control in an intersection the force required to turn the wheel was much much harder. To the point I almost failed. Definitely scared me but I never thought FSD was trying to retake control just that the force to turn the wheel was extremely hard. Both times using 10.4

is it published somewhere the torque required to disengage AP/FSD? I believe I've seen it somewhere. I've definitely experienced varying torque forces required to disengage. Usually what gets my attention is when it's super light and I barely apply any torque. On the firm side, this is quite subjective, but I haven't felt anything resist harder than disengaging on the highway.

I wonder if the direction of the driver's torque matters. For example, if FSD is jerking right and the driver applies torque to the left to undo the turn, does that require more torque than say, turning right in addition to FSD's turn because FSD isn't turning sharp enough. I feel my light-torque disengagements have always been turning same direction as FSD.

 

[Dan D.]

I can see where someone might believe FSD was at fault. Twice when I have needed to immediately take control in an intersection the force required to turn the wheel was much much harder. To the point I almost failed. Definitely scared me but I never thought FSD was trying to retake control just that the force to turn the wheel was extremely hard. Both times using 10.4

I have had to take over control from FSD Beta, only to have the Lane Departure take control, forcing me to wrest control from the car once again. Dropping out of FSD Beta takes a very light touch, but overriding the Lane Departure control takes more force. Since TACC stays active, and if it is set to a relatively high speed, I could see how the car could unexpectedly accelerate and exert steering control at the worst possible time.

If the report is true, the driver may not be as technically aware as you two. You've offered similar experiences to the driver's, and he/she may not be able to understand what happened, but you've given a possible way of their accident happening.

I see no need to ridicule the driver, we've seen many FSD beta videos that were close calls, some are bound to crash eventually, it even happens to good drivers.

In fact @FloridaJohn's account is quite concerning.

 

[qdeathstar]

I have had the lane assist kick in after an fsd disengagement but I don’t think it requires more force, but even if I am wrong the difference is minuscule and certain not enough yo render the car uncontrollable, assuming your hands are gripping the wheel at that point.

 

[aronth5]

After going back to the intersection where the I needed extra force to take over I found this. I'm not saying this was related to the report filed with the NHSTA just that stuff happens that's difficult to explain at the moment. The planner indicated the car should turn left but the navigation steps showed a right turn then a u turn. Crazy but the mapping error may explain the contention between FSD and me. Reported to FSD beta team right after it happened.
Interestingly even though I've been a critic of 10.4 it does take this turn much better then 10.2 or 10.3.

View attachment 726052

 

[S4WRXTTCS]

Then you aren’t a very good beta tester. Hands should be at 3 and 9, which would be right next to the stalk.

You want the car to coast after adjusting the direction of the steering wheel in FSD?

You might be better off with Ford Blue Cruise

I'm not a good beta tester for other reasons, but there is thread for that.

I do want the car to coast after DISABLING fsd beta with the control my hand is already on, and not control right next to it. Certainly not a control that has multiple functions (a bad UX design).

There is no adjusting the direction with the steering wheel. The implementation would be significantly better if steering adjustments didn't cancel anything, and all they did was acted as feedback to FSD beta.

Like if I'm approach something in the road that I want to avoid I'd love to be able to nudge it around it without having it cancel.

Ford Blue Cruise ruined their systems because they dumbed it down when its hands free. So it can't take curves in the freeway like it can while not in the hands free mode.

 

[S4WRXTTCS]

I have had to take over control from FSD Beta, only to have the Lane Departure take control, forcing me to wrest control from the car once again. Dropping out of FSD Beta takes a very light touch, but overriding the Lane Departure control takes more force. Since TACC stays active, and if it is set to a relatively high speed, I could see how the car could unexpectedly accelerate and exert steering control at the worst possible time.

I guess I wrongly assumed everyone turned that off. But, that is an excellent observation.

 

[Fernand]

The guy who had the accident reported to the NTSB (?) was supposed to keep control, hand(s) on wheel. Basically he's admitting he wasn't in control and doesn't even understand the auto-steer system. His tone is litigious. That's all we need. Newbies on an advanced beta.

Regarding the continued engagement of TACC, it's tolerable on the freeway, but I've found it VERY troublesome on the streets, both previously in AP and in BetaFSD. Since I "can't trust it to let go" I've developed the reflex to immediately push up on the stalk to fully disengage, even while steering, braking, cursing.

 

[AWDtsla]

There is no adjusting the direction with the steering wheel. The implementation would be significantly better if steering adjustments didn't cancel anything, and all they did was acted as feedback to FSD beta.

Like if I'm approach something in the road that I want to avoid I'd love to be able to nudge it around it without having it cancel.

Ford Blue Cruise ruined their systems because they dumbed it down when its hands free. So it can't take curves in the freeway like it can while not in the hands free mode.

I suggested this ~5 years ago on the first release of AP v1, and people looked at me like I was crazy. It makes a lot more sense to allow this for training FSD, where you can make the appropriate arc through a turn even though FSD is engaged. Oh, and it needs the help. I'm surprised how bad the path planner is compared to inference.

 

[EVNow]

I suggested this ~5 years ago on the first release of AP v1, and people looked at me like I was crazy. It makes a lot more sense to allow this for training FSD, where you can make the appropriate arc through a turn even though FSD is engaged. Oh, and it needs the help. I'm surprised how bad the path planner is compared to inference.

The problem is - the turning is not a neural network that can be trained with that kind of data. Its a procedural code - using cost optimization based strategy to figure out the right path. Once the desired path & speed are figured out the procedural code knows what instructions to send to the wheel, accelerator, brake etc.

Apparently they use a simulator to test the planner - and they could use that kind of information when testing, I guess.

 

[AWDtsla]

The problem is - the turning is not a neural network that can be trained with that kind of data. Its a procedural code - using cost optimization based strategy to figure out the right path. Once the desired path & speed are figured out the procedural code knows what instructions to send to the wheel, accelerator, brake etc.

Apparently they use a simulator to test the planner - and they could use that kind of information when testing, I guess.

Procedural just means you using a grey matter neural net to tune the algorithm. In either case, a difference between actual steering angle and commanded steering angle could be used to identify when the path planner is doing really badly. Right now all they ahve is "user disengaged".

 

[EVNow]

Procedural just means you using a grey matter neural net to tune the algorithm.

If only AI was that simple.

In either case, a difference between actual steering angle and commanded steering angle could be used to identify when the path planner is doing really badly. Right now all they ahve is "user disengaged".

Yes - they can use it for testing the planner in the simulator. Basically, the question is how does the difference in steering angle translate to various costs they use in their Monte Carlo simulation to find the best cost optimized path.

 

[nvx1977]

Regarding the continued engagement of TACC, it's tolerable on the freeway, but I've found it VERY troublesome on the streets, both previously in AP and in BetaFSD. Since I "can't trust it to let go" I've developed the reflex to immediately push up on the stalk to fully disengage, even while steering, braking, cursing.

not to mention when you disengage in mid turn, the turn signal gets stuck. Lots of extra unnecessary cognitive load to turn off TACC and/or the turn signal just as you're manually steering the car from a potentially bad situation.

 

[Thp3]

I guess I wrongly assumed everyone turned that off. But, that is an excellent observation.

Yep, my learning is to turn off lane departure. Gotta go, I’m doing it now…

 

[TresLA]

Re turning torque causing disengagement vs brake pedal vs hitting up on the stalk, I’ve learned to hold the wheel lighter when traveling slower and at times FSD beta changes velocity (speed and/or direction) more quickly. In those moments, I also hold my right hand even lighter with fingers touching the bottom of the right stalk, ready to disengage FSD beta very quickly and smoothly. I have videos of this I sent to a friend with FSD beta. I don’t know how to easily post vids here without creating a YouTube account for it, so I’ll just post screen shots from that vid:

 

[Fernand]

FWIW keeping the right hand at 3:00 seems like an ideal way to have access to "everything", including tapping AP on and off. I keep a small 100g counter-weight on that wheel spoke, making a good spot to rest fingers on. The counter-weight reduces the amount of anti-nag torque that needs to be applied, so resting fingers on it is often sufficient during freeway NOA.

 

[nvx1977]

Re turning torque causing disengagement vs brake pedal vs hitting up on the stalk, I’ve learned to hold the wheel lighter when traveling slower and at times FSD beta changes velocity (speed and/or direction) more quickly. In those moments, I also hold my right hand even lighter with fingers touching the bottom of the right stalk, ready to disengage FSD beta very quickly and smoothly. I have videos of this I sent to a friend with FSD beta. I don’t know how to easily post vids here without creating a YouTube account for it, so I’ll just post screen shots from that vid:

View attachment 732593View attachment 732594

I remember someone saying (maybe chuck cook) that disengaging with stalk won't trigger an upload to the mothership. Whereas steering wheel and brake have a chance of that event getting uploaded. so the stalk should be used only to cancel FSD when you want to stop using it and no bad behavior was exhibited.

 

[TresLA]

I remember someone saying (maybe chuck cook) that disengaging with stalk won't trigger an upload to the mothership. Whereas steering wheel and brake have a chance of that event getting uploaded. so the stalk should be used only to cancel FSD when you want to stop using it and no bad behavior was exhibited.

I’m just in the habit of tapping the camera icon every time anyway, just to make sure the snippet is saved for uploading.

 

[JulienW]

I’m just in the habit of tapping the camera icon every time anyway, just to make sure the snippet is saved for uploading.

in the past it was reported (by YouTuber's I think) that Tesla only uploads 5 snapshots per drive and if you push it a 6the time it starts overwriting.

No word if there as a limit or what criteria on Brake/Steering Wheel disengagements.

Too bad Tesla dosn't spell this out so we know what HELPS the best. That would make us more productive as Beta testers.

 

[CyberGus]

Yea, at first he was saying beta steered into an adjacent lane and he didn’t have time to react. Which is 100 percent possible even if you are paying attention. But if he is saying the car overrides his inputs that is highly impossible.

I'm sure Consumer Reports will find a 15-step method to recreate it with toothpicks and duct tape

 

[TresLA]

in the past it was reported (by YouTuber's I think) that Tesla only uploads 5 snapshots per drive and if you push it a 6the time it starts overwriting.

No word if there as a limit or what criteria on Brake/Steering Wheel disengagements.

Too bad Tesla dosn't spell this out so we know what HELPS the best. That would make us more productive as Beta testers.

I've heard similarly it was a space issue, but I figure I just keep reporting and hope that useful data gets to them!