The AP Lane Keeping Bias to the Right is Not Safe, please fix this Elon!

[RDoc]

I was out on a highway recently with some slush still on the road but the AP icon came on and the lane markers were solid on the display so I tried AP. What happened was that the car ping ponged from left to right in the lane, never leaving it, but never stabilizing. What really surprised me though was that the dashboard display accurately depicted what the car was doing, moving from side to side within the solid lane markers.

It appears again to me that the system that shows the lane markers and does lane departure warnings is separate from the AP system. This makes zero sense to me.

 

[Andyw2100]

I was out on a highway recently with some slush still on the road but the AP icon came on and the lane markers were solid on the display so I tried AP. What happened was that the car ping ponged from left to right in the lane, never leaving it, but never stabilizing. What really surprised me though was that the dashboard display accurately depicted what the car was doing, moving from side to side within the solid lane markers.

It appears again to me that the system that shows the lane markers and does lane departure warnings is separate from the AP system. This makes zero sense to me.

The following is not directed at RDoc alone. His post just really drove the point home for me, so I felt the need to post.

AP is an abbreviation for Auto Pilot. All of the Driver Assistance features are part of the Auto Pilot functionality.

In the post above, what you are referring to as AP is Autosteer Beta.

I think a lot of confusion about features, what they are supposed to do, what they can't do, etc. would be eliminated if we could all start referring to the features by their correct names.

 

[AWDtsla]

After going for a long drive today, I don't see any corrections to bias, or any evidence of bias. The car drives like a drunk. I thought it was going to hit a concrete barrier today. The ultrasonic next to the barrier was firing, and nothing else, no cars to the side. The sad part is I've never noticed the lane tracking being off, the car knows it's in a weird place in the lane, it knew the same next to the barrier.

I was "nagged" probably 3 times over 2.5 hours, with no real evidence that the car could not track lanes why being nagged.

It actually seemed to do the best on a 30 mph curvy road with a double yellow, it seemed to stay consistently center of lane, less drunk.

 

[Mark Z]

Using AP with Model X, I thought the vehicle wasn't centered, until...

I looked carefully out the left rear view mirror and out the right rear view mirror. (You can take a good look when the vehicle is on AP.)

Perfectly centered.

I have had to slightly redefine my forward windshield view of what perfectly centered is.

UPDATE: Works okay on normal sized lanes. Tried it today on a extra wide lane, and it pulled to the right line, not good.

I ended up driving without AP.

 

[AWDtsla]

Using AP with Model X, I thought the vehicle wasn't centered, until...

I looked carefully out the left rear view mirror and out the right rear view mirror. (You can take a good look when the vehicle is on AP.)

Perfectly centered.

I have had to slightly redefine my forward windshield view of what perfectly centered is.

I've tilted my right mirror down enough to see how far away the line is from the car to confirm where the car actually is, and it hasn't been that far off from where I thought.

Also, the scared drivers pulling way to the right after I pass them with AP on also confirm it.

 

[wk057]

I can confirm that the virtual lane on the instrument cluster indeed comes from the autopilot hardware and is the same hardware making steering decisions. All the weirder that it still lets the car do things like bounce around and even go over the lane marking at times.

 

[Mark Z]

I can confirm that the virtual lane on the instrument cluster indeed comes from the autopilot hardware and is the same hardware making steering decisions. All the weirder that it still lets the car do things like bounce around and even go over the lane marking at times.

Going over the lane markers? I haven't had that problem. I would be concerned as well.

Remember, my limited experience is only with Model X and the 7.0 software with multiple lane freeways.

One more thought. The right rearview mirror is more curved, so IMHO, the distance may appear closer to the line.

If this is repeatable, why not have two people follow the OP in one vehicle. The passenger can video how close the vehicle is to the bridge.

 

[S4WRXTTCS]

What I find odd is before V7.1 I never had any real issue with lane biasing probably because it never really seemed to do lane biasing. It seemed to go for the middle and mostly tried to stay there unless it detected something on one side that was too close.

I used the lane-steering for thousands of miles with V7.0 without much issue except in the usual places (diving for exits, etc).

But, now with V7.1 there are definitely times where in the left lane that it biases way to much to the right. Even when trying to correct it the thing goes back. I don't think its every left lane. Just left lanes with certain types of lane markings on the left. Like on 405 on the left lane next to the HOV lane where it has a double stripe line. In the middle lane it stays nice and centered unless it has to move over because of a car being too close.

I don't have that many miles with it, but the 100 or so that I do have are noticeably worse because of weird lane biasing in the left lane. In some ways it's funny because I assumed I'd have issues with nagging, but the reality is the nagging isn't bad because I think they reduced the amount of steering torque required to make it go away. What I do have issues with is lane biasing.

 

[Cyberax]

It all looks like we're in some kind of a large A/B/C/D... experiment. Perhaps various cars are getting builds with slightly different settings to gather more data for Tesla.

So some of us get lots of nags and some of us get bad biasing behavior. FWIW, my friend's Tesla (P85D, NYC) on 7.1 has no problems with biases but lots of nags. I'm not upgrading mine from 7.0...

 

[David_Cary]

I'm starting to agree with Cyberax points.

Time for a Nag poll cross referenced with a lane bias poll.

For me there is improved lane bias on undivided roads. I don't drive in the left lane very often - or for more than a few seconds - so right lane bias isn't much of an issue for me.

For me there is an unequivocal new rule. If the left sided lane markings are double yellow, bias to the right. That wasn't there before 7.1. Certainly you left lane squatters (jk) who have HOV lanes to the left could be bothered by this new rule.

 

[ecarfan]

It all looks like we're in some kind of a large A/B/C/D... experiment. Perhaps various cars are getting builds with slightly different settings to gather more data for Tesla.

So some of us get lots of nags and some of us get bad biasing behavior. FWIW, my friend's Tesla (P85D, NYC) on 7.1 has no problems with biases but lots of nags. I'm not upgrading mine from 7.0...

I think it more likely that everyone with 7.1 has the same software, but because they are driving on different roads at different times of day (different lighting conditions) with varying degrees of traffic around them they see different behavior. Even people driving the same roads are using different lanes at different times of day with varying amounts of cars around them.
As a scientist, I can tell you that you cannot describe that as an "A/B/C/D experiment" because there are far too many uncontrolled variables!

 

[roblab]

...the car knows it's in a weird place in the lane, it knew the same next to the barrier.

I feel that the car can't see the barrier. It can "feel" it with the side lidar (or sonar or whatever), the the AP camera needs to see painted lines. So it's looking for a line and feels this huge something next to it. We have narrow stone bridges here. The car feels them, but can't see them. The bridges look like a lot of human drivers couldn't tell where the walls were. Lots of work for body shops.

We have lots of dots for lines here, too. The car can't see dots, either.

When I see weak lane markings, dots, cones, barriers, etc., it's time to take over from the car.

 

[JenniferQ]

What I find odd is before V7.1 I never had any real issue with lane biasing probably because it never really seemed to do lane biasing. It seemed to go for the middle and mostly tried to stay there unless it detected something on one side that was too close.

I used the lane-steering for thousands of miles with V7.0 without much issue except in the usual places (diving for exits, etc).

But, now with V7.1 there are definitely times where in the left lane that it biases way to much to the right. Even when trying to correct it the thing goes back. I don't think its every left lane. Just left lanes with certain types of lane markings on the left. Like on 405 on the left lane next to the HOV lane where it has a double stripe line. In the middle lane it stays nice and centered unless it has to move over because of a car being too close.

I don't have that many miles with it, but the 100 or so that I do have are noticeably worse because of weird lane biasing in the left lane. In some ways it's funny because I assumed I'd have issues with nagging, but the reality is the nagging isn't bad because I think they reduced the amount of steering torque required to make it go away. What I do have issues with is lane biasing.

Agree.

 

[garygid]

Driving in freeway light traffic with 2.9.154, at first the lane centering in 7.1 seemed better,
smoother, and more gentle, with noticably less little jerks of the steering wheel.

Later, in very heavy slow traffic, a bus came alongside on the left, and my MS moved
too far to the right in my lane, enough to bother both me and the car traffic in the
lane on my right. The AutoSteering (AS) seemed to be overdoing moving the car
away from the "wall". All 4 lanes were perhaps a bit narrower than usual, due
to construction, or at least it seemed that way.

Getting off the freeway, in the leftmost of two exit only lanes, the AS swerved
abruptly to the left, heading into the narrow triangular area that contained lots of
crash barrels (probably filled with water?), and I had to quickly take
control to get back fully into the exit-only lane. My wife gasped, and
had a severe nether-pucker moment, I think, but she was glad that we lived.

Moral: All this is Beta unfinished and unpolished software, and subject to
making unexpected decisions at any time, and acting abruptly.
We all need to be on high alert, not half asleep, when using
any of the AP functions. The fact that the AP worked reliably many
times does not mean that it will the next time, since it is a computer,
it can have a "stroke" at any moment.

Please be extra careful.

 

[dhanson865]

I feel that the car can't see the barrier. It can "feel" it with the side lidar (or sonar or whatever), the the AP camera needs to see painted lines. So it's looking for a line and feels this huge something next to it. We have narrow stone bridges here. The car feels them, but can't see them. The bridges look like a lot of human drivers couldn't tell where the walls were. Lots of work for body shops.

We have lots of dots for lines here, too. The car can't see dots, either.

When I see weak lane markings, dots, cones, barriers, etc., it's time to take over from the car.

dots for lines like Botts Dots AP Guidance Added Under 7.1 ?

You may not have realized it if you were used to 7.0 but you can let the car steer in botts dots lines now.

 

[Todd Burch]

What I find odd is before V7.1 I never had any real issue with lane biasing probably because it never really seemed to do lane biasing. It seemed to go for the middle and mostly tried to stay there unless it detected something on one side that was too close..

Right--in 7.0 the biasing was only when an obstruction was next-door.

I suspect when the video of the moron riding the two-lane, hilly, curvy road on autopilot that almost head-on'd himself into another car and then posted it saying that AUTOPILOT almost killed him (not that he almost killed himself), Tesla decided that whenever a yellow line is seen to the left of the car, it biases to the right side of the lane, obstructions or not.

I'm generally ok with this if we're talking a foot or so of bias. But the bias is way too extreme, and makes me and clearly others pretty uncomfortable to the point that we disengage autopilot.

 

[wk057]

Going over the lane markers? I haven't had that problem. I would be concerned as well.

Remember, my limited experience is only with Model X and the 7.0 software with multiple lane freeways.

One more thought. The right rearview mirror is more curved, so IMHO, the distance may appear closer to the line.

If this is repeatable, why not have two people follow the OP in one vehicle. The passenger can video how close the vehicle is to the bridge.

It definitely goes over the line. It even knows it goes over the line because the lane departure warning noise happens. It's weird. Happens mostly at lower speeds (<50 MPH). I reported it.

 

[brkaus]

Someone that knows more about the internals - would it be possible for Tesla to have a stable and latest AP track selectable by the user?

Is there enough space and/or architected such that it would work? I'm guessing no because of the distributed nature of the code across the hardware pieces.

 

[calisnow]

It definitely goes over the line. It even knows it goes over the line because the lane departure warning noise happens. It's weird. Happens mostly at lower speeds (<50 MPH). I reported it.

Okay so the obvious question is why has Tesla set up 7.1 this way - ie set it up so it overcompensates rather than biasing a "medium" amount? Enough of you are reporting this poor behavior (biasing so far it goes all the way over the lines into the next lane, or at least right up on them) that Tesla must have known what it does before they pushed it out, right?

And if so - what were they thinking? Were they rushed (ie whipped by Elon) to get an update to reign in the Youtube idiots, and thus simply unable to "smooth" the biasing behavior in time for the release - and then made a decision to simply push it out despite this overcompensating behavior that they knew would blow up the forums with complaints? And if so - do you presume they are now working on smoothing out the bias and intend to push it out in another update?

I've never worked in software development so trying to imagine what is going on in the heads of the autopilot team is a real puzzle.

 

[CHG-ON]

I find that mine biases so much to the right that I can't trust it in moderate curves to the left. It gets so close to the limit line that I always take over. It was better in 7.0 because it stuck to the middle of the lane. This is on undivided roads. I am hoping to experience freeway driving this weekend for the first time. But we are expecting torrential rains, and I won't use it then. So it might be a while before I can see how it works in that environment. However, I do find that it works much better in intersections (simple ones) and with left and right turn bump outs. Stays pretty much straight, where it did not with 7.0.