FSD Beta : Steering Wheel Jerking

[AlanSubie4Life]

Elon: Tesla Autopilot spends a lot of effort on minimizing jerk.

https://twitter.com/x/status/1457083243933601802

Haha! Looking forward to the day when they expend the same effort on minimizing FSD jerk. Hopefully with better success than they've had with Autopilot. They may want to consider a different approach! At least they're aware of the issue (it's well known to be a high priority in control systems engineering of course). It must be an inherently difficult problem to solve for AI, even though it is so easy for a human. Though maybe it has nothing to do with AI in this case if it's the procedural part of the code that's the issue. Seems impossible to know, though some cases are certainly related to perception (which would be in the AI category), so that clearly has to improve.

 

[EVNow]

Sounds like it is a high priority to smooth it out.

He is talking about a different jerk - not the wheel jerk I mentioned (and the one @AlanSubie4Life mentions above).

Speaking of which - the jerky behavior of the steering wheel just before a turn is a headscratcher. I don't see what kind of cost optimization would come up with that ... what would get minimized if you move the steering wheel left & right rapidly ?

Seems to me there is high uncertainty at the beginning of a turn when the planner still doesn't know what to do. The result is no forward movement - because the planner is still trying to decide what to do. But, it decides to move the steering wheel in all directions ? Seems like something they should be able to fix easily by not moving the steering wheel until the car starts moving forward.

 

[Ramphex]

He is talking about a different jerk - not the wheel jerk I mentioned (and the one @AlanSubie4Life mentions above).

Speaking of which - the jerky behavior of the steering wheel just before a turn is a headscratcher. I don't see what kind of cost optimization would come up with that ... what would get minimized if you move the steering wheel left & right rapidly ?

Seems to me there is high uncertainty at the beginning of a turn when the planner still doesn't know what to do. The result is no forward movement - because the planner is still trying to decide what to do. But, it decides to move the steering wheel in all directions ? Seems like something they should be able to fix easily by not moving the steering wheel until the car starts moving forward.

I don’t know why the turns can’t be treated the same as the curves in the road, smooth, once FSD is committed and there’s no cross traffic.

 

[AlanSubie4Life]

But, it decides to move the steering wheel in all directions ? Seems like something they should be able to fix easily by not moving the steering wheel until the car starts moving forward.

I don't like the conclusion, but I've decided that this is likely intentional, as it communicates the state of uncertainty of FSD, and helps keep the driver alert. Hopefully on average it results in more disengagements and trains the driver to keep their hands on the wheel (which can be important in those rare situations where FSD actually turns into other traffic or towards parked cars). Not sure whether it will actually accomplish that, but they may have made a considered decision and finally just let the Scrub Bot have its way with the steering rack. Seems like there are other better ways to communicate the uncertainty, but this was easy pickings.

It's not great when it turns the wheels before a left turn, though.

 

[EVNow]

I don't like the conclusion, but I've decided that this is likely intentional, as it communicates the state of uncertainty of FSD, and helps keep the driver alert. Hopefully on average it results in more disengagements and trains the driver to keep their hands on the wheel (which can be important in those rare situations where FSD actually turns into other traffic or towards parked cars). Not sure whether it actually will actually accomplish that, but they may have made a considered decision and finally just let the Scrub Bot have its way with the steering rack. Seems like there are other better ways to communicate the uncertainty, but this was easy pickings.

It's not great when it turns the wheels before a left turn, though.

That’s sounds extremely unlikely to me. Infact this makes it unlikely people will interrupt a bad turn. After all until the turn happens, we can’t make our because the wheel doesn’t show the true intentions.

As with most things Tesla, it’s a bug and not by design.

 

[AlanSubie4Life]

As with most things Tesla, it’s a bug and not by design

In this case, I would definitely classify this as a feature. And yes, it is possible it is not by design. I don’t know, of course.

Infact this makes it unlikely people will interrupt a bad turn.

For me it makes it much more likely to disengage, since it automatically disengages as soon as it turns away from the correct direction, so that increases the likelihood of me interrupting a bad turn, by providing the correct steering vector at all times. Just a data point. YMMV. I disengage on the majority of turns, I think - and the few where I don’t disengage actually feel pretty great because they are correct.

 

[EVNow]

I don’t know why the turns can’t be treated the same as the curves in the road, smooth, once FSD is committed and there’s no cross traffic.

There are no lines to guide the car ?

Basically the car has to estimate the path from the current location to where it needs to go and turn the wheel to accomplish that. This part is similar to the twisting road. May be it’s the end point FSD can’t be sure about.

 

[Ramphex]

It has a smooth guide line though, usually

 

[momo3605]

It has a smooth guide line though, usually

I’ve noticed a lot of times, especially for lefts, especially in larger intersections, it doesn’t have a clear view of the lanes at the other end. So as it follows the curve, it keeps making micro-corrections as its confidence level increases as it gets further into the turn

 

[Ramphex]

I’ve noticed a lot of times, especially for lefts, especially in larger intersections, it doesn’t have a clear view of the lanes at the other end. So as it follows the curve, it keeps making micro-corrections as its confidence level increases as it gets further into the turn

Yea, right turns should be pretty smooth though. From my experience it dry turns the wheel too much prematurely and then starts to unwind it as it starts to turn, instead of gradually turning the wheel as it’s making the turn

 

[EVNow]

I disengage on the majority of turns, I think - and the few where I don’t disengage actually feel pretty great because they are correct.

So, you are basically disengaging when the steering wheel jerks ... may be we should all do it and send reports on every turn, till they fix it.

Yea, right turns should be pretty smooth though. From my experience it dry turns the wheel too much prematurely and then starts to unwind it as it starts to turn, instead of gradually turning the wheel as it’s making the turn

Yes - no reason at all for the right turns to be jerky. If they are worried about hitting the curb, the car should drive straight a bit before turning - not jerk to the right and then left and then right and ...

With 10.4, the one right turn I did from stop was especially bad. The car crept in to the road (no obstructions, so no need to creep really) and then jerked quite a bit before making the turn and ending up correctly in the middle of the lane.

 

[TIppy]

As for the rapid steering wheel movements at the beginning of turns, I think it's a matter of FSD trying to get to the center of the path at a given rate. If the car is going fast it takes only a little bit of steering angle to provide the required speed in the lateral direction toward the new path. But going slowly a larger angle from the original path is required to produce the same rate. The path isn't fixed it's constantly shifting to the left and right even if by just a little. You can feel this in the micro adjustments to steering when travelling a normal speeds( e.g. 35 mph).

 

[Mardak]

The path isn't fixed it's constantly shifting to the left and right even if by just a little

I wonder if the full stack (360° video with memory) is already applied to path prediction (like the incoming object persistence), so if not, maybe doing so will improve things before and during a turn. At least right now when predictions aren't confident, the adjustments are definitely necessary especially for odd intersections. There's an angled left turn that FSD Beta 10.x has wanted to turn into oncoming traffic left of the median, but most of the time, it realizes its mistake as it gets closer and adjusts correctly to the right.

 

[AlanSubie4Life]

So, you are basically disengaging when the steering wheel jerks ... may be we should all do it and send reports on every turn, till they fix it.

Yes. Unless I am feeling in an experimental mood AND there is no traffic in the vicinity, this is what I do.

It seems like it should help Tesla to get disengagements whenever the wheel does the wrong thing. They should know that it is not ever supposed to turn the wheel incorrectly!

The only thing that is annoying about this is that I have to be quick to half tap up on the gear stalk or hit the brake to remove TACC immediately. And not put it in reverse by mistake. Good to develop that muscle memory though.

 

[TIppy]

Yes. Unless I am feeling in an experimental mood AND there is no traffic in the vicinity, this is what I do.

It seems like it should help Tesla to get disengagements whenever the wheel does the wrong thing. They should know that it is not ever supposed to turn the wheel incorrectly!

The only thing that is annoying about this is that I have to be quick to half tap up on the gear stalk or hit the brake to remove TACC immediately. And not put it in reverse by mistake. Good to develop that muscle memory though.

I don't think it is turning the wheel incorrectly. It's just that the path has moved sideways and the car needs to steer in the direction of the new center location.

 

[EVNow]

The only thing that is annoying about this is that I have to be quick to half tap up on the gear stalk or hit the brake to remove TACC immediately.

Yes - its annoyingly difficult. I'm not used to TACC at all. Yet, when the turn is wrong, we have disengage by forcing the steering wheel. I'm slowly getting used to just steering correctly and then re-engaging FSD.

 

[AlanSubie4Life]

I don't think it is turning the wheel incorrectly. It's just that the path has moved sideways and the car needs to steer in the direction of the new center location.

If it’s not how I would take the turn, it’s wrong. That’s how I view it. Typically very small errors don’t result in a disengagement. But the violent scrubs and sudden jerks always do.

 

[TIppy]

If it’s not how I would take the turn, it’s wrong. That’s how I view it. Typically very small errors don’t result in a disengagement. But the violent scrubs and sudden jerks always do.

I don't get disengagements if I just let the wheel slide through my hands.

FSD isn't trying to make a turn. It's trying to follow a path. Whenever the car is not on the path, it must turn toward it to minimize the error. The bigger the error the faster it tries to get back to the path. Albeit FSD is generating the path, once it's generated another algorithm doggedly tries to follow it. If FSD generates a new path that is left or right of the current path, there's no choice but to turn toward it. It's just more noticeable when the car is moving slowly because it requires more extreme rotations of the wheel.

The rapid movements are going to lead to a dangerous situation, FSD is just trying to get to the new path.

 

[AlanSubie4Life]

I don't get disengagements if I just let the wheel slide through my hands.

I don’t either but I see no benefit in allowing that. It’s not like the objective is to not have interventions. This is not supposed to make driving easier, at the moment. At least that is my understanding.

I don’t really understand the argument here. If FSD is generating alternate pathing for no good reason (whether stationary or not), that is an error, and a disengagement would seem to be the correct response. Path planning in the immediate vicinity should be rock solid, when there are no changes to the environment. Otherwise it is not good enough.

 

[TIppy]

I don’t either but I see no benefit in allowing that. It’s not like the objective is to not have interventions.

It's an intervention when FSD is performing correctly causing noise in the data sent to Tesla. They have much bigger issues than this one. Once the car can actually drive itself, they can go back and optimize.