Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

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

This site may earn commission on affiliate links.
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.

Does it actually cross the limit line in those left hand turns on undivided roads? Maybe it's on purpose to stay far to the right - the logic being it's safer to stay as far away from oncoming traffic as possible.
 
Sorry if dual posting, but my experience with cars on right in 7.1 has been influenced by a couple of rightward drifts as cars come up from behind on the right. Particularly big cars/trucks (white seems more sensitive). It makes me wonder if the proximity sensors know only one value, as fed thru the car's AP processing, and don't account for the value's angle. When a car is behind, the distance seen by the sensor is greater than the cement wall to my left. In that case, I think it might try to equalize by moving toward the activated white sensor, cutting off the driver behind. Also, it is as if lane markings got "second billing" in 7.1's AP processing.

With no proximity alert on my right, I would say 7.1 does stay closer to the wall (which is desired). It's when right side proximity is active, that I start having problems on my right.

I'm FAR from turning AP completely off, and find it wonderful in stop & go. But everybody will have this tech soon, and I sorta wish they'd focus more on the car.
 
Sorry, I meant does it cross the limit line on the right side of the lane (since you're saying it biases too much to the right) - or just hug the line but not cross it?
I'm not sure it actually crossed the line, but I had a case where the car went so far over that to the left that it triggered the lane departure warning. I grabbed the wheel because the car was heading for the guard rail.
 
I'm not sure it actually crossed the line, but I had a case where the car went so far over that to the left that it triggered the lane departure warning. I grabbed the wheel because the car was heading for the guard rail.
Hitting the guard rail is totally unacceptable. Once I had the S doing ping-pong (while one hand driving) and a truck approached me from left and behind but no risk of collision, the ultrasonic sensor triggered my car to swerved big time to avoid it.
Overall I think 7.1 is worse than 7.0 for AS.
 
So, as mentioned in the nag thread, I just finished a ~1500 mile road trip. Most of it was at night on mostly empty roads, which was great for autosteer.

However, I can pretty much guarantee that without intervention that autosteer would have plowed into the rear left corner of probably 50 trucks that I passed due to the right-bias issue. Makes no sense. The car know it's on or over the right lane marker based on the visuals. It even sees the truck we're approaching in that lane to the right. Yet it maintains its position which at best leaves a few inches between my mirror and the truck, and at worst it would have left a few negative inches in between (ie, collision). In cases where the truck was far enough to the right where I let autosteer continue without intervention, the ultrasonics would eventually see the truck very close and start nudging to the left and would be maybe center of the lane by the time I passed the truck. Still not good.

Why doesn't the car just stay centered in the first place on a major highway? I don't get it.

This is a little different than the OP where there is something to the left for the ultrasonics to ping constantly. In my case it was a large grass median the whole way, no guard rails or anything to ping off of. Yet still, riding in the far left lane, the car would bias to basically on top of the line to the right. Sometimes it would drift to the center, then back again to the right, sometimes over the right line. Required much more babysitting than 7.0, which is definitely a step backwards.

I'll be very surprised if there aren't a bunch of accidents because of this issue in 7.1. Not that the people shouldn't be paying attention or anything and taking over (like I do basically every time I approach a vehicle to pass) or to pass blame to autopilot, but I could see a momentary lapse in attention much more easily causing an accident with 7.1 vs 7.0.

As good as a human my ***.
 
What is there about individual cars that could make two different Model S adopt a different lane bias? I honestly don't see lane bias under 7.1 being much if any different from that in 7.0. The earlier version always made me nervous when passing big trucks on divided highways, by allowing the car to get (and remain) too close, side-to-side. That hasn't changed: it still makes me nervous under 7.1. But I don't see it as being obviously worse. I put in 2500 miles or more in AS under 7.0 and am rapidly approaching that many miles under 7.1.

I have experienced the lane departure warning in AS under 7.1; it seems to happen mostly on modestly curved, undivided roads. I can't really pin down the behavior, though: it seems to vary from trip to trip.
 
The solution is actually very simple and should have been done from the start ! Allow the driver to adjust it position within the lane. So let the drive center the car while autopilot is active and use that position ! There is no position fit all situation and fit all driver style. So this would resolve pretty much all complaint.

The only issue with that is the AP hardware is actually pretty weak and I'm not sure any software upgrade will fix that. From my limited Tesla AP experience (loaner), the Mobileye system don't see really far and have issues extrapolating potential line position or even following a car in front of you.

For example, at street light where there is no line because of the intersection, it will try to follow the car in front, but it will zigzag around when the while line are not present, EVEN if the white line are present and visible just 10 meter ahead, and the car in front and yourself are going in a straight line.

In that instance, it actually easy to fix, you extrapolate the while line continuation for X number of meters and confirm that with the car in front.

But those are hack to a lack of good sensors, and there is limit to what you can do with those.

For example, I've made a taxi dispatch software that predict where you will want to be pickup by analyzing previous calls. What great is you can use actual pass data to verify your accuracy. You can take a call from 2 day ago, and run the engine with data from before that 2 day ago and you can compare the actual client address versus what you predicted allowing you to quickly optimize your engine.

So the car could do the same thing as a security system, and if deviation is too high, then disable that system.

You can also verify your actual position by comparing 3 image. You know your speed and heading, so you can predict how the image will look like before you actually get that second image. Then you can compare what you predicted and the actual image and find deviation of your car.

Also, the car also need a bunch of other camera to get an actual 360o view of the car. And not just that, but position them at useful position. For example, central camera require it effective range of of the line being visible. Having 3 forward facing camera would be best (1 central and 1 on each side mirror).

The biggest things for AP system would be to create an camera lens that be clean by an electric signal. Or as a an alternative is to add a bunch of camera with correction software that can do composition from all of them. At some point, you will have to clean sensors, but maybe not as often as you would with just 1 camera.

Anyway, lot of stuff to do on the AP systems :)

I'm personally not disappointed anymore that I don't have the AP system in my car ! I'll probably like the 4th version of it ;)
 
Last edited:
I travel this 55mph road to work each morning and AP's lane biasing, combined with intermittent detection of the low curb on the median side, makes the car act like it's going to cut people off in the right-hand lane. At a minimum, it's unnerving to use AP on this stretch of road. Other 45+ mph roads in this area also have sidewalk-height median barriers and AP isn't really usable in the innermost lane due to the excessive bias. It really ought to back off the bias if the path (as shown on the IC display) is going to put the mirror or the tires over the striped line.

Screen Shot 2016-01-29 at 12.55.18.png
 
Why doesn't the car just stay centered in the first place on a major highway? I don't get it.

This is a little different than the OP where there is something to the left for the ultrasonics to ping constantly. In my case it was a large grass median the whole way, no guard rails or anything to ping off of. Yet still, riding in the far left lane, the car would bias to basically on top of the line to the right.

I'm the OP:

So after watching this for several weeks, I'm convinced that the majority of biasing is triggered by the camera seeing a yellow line on the left side of the car, although curbs/barriers also contribute.

I think Tesla decided that if the car sees a yellow line to its left, to ride on the right side of the lane. I'm fairly certain because I also see strong biasing in a left lane when there's a flat grass division between roads (no obstruction for the ultrasonics to pick up on), and when I ride in the right lane with a white line to the left, the car is perfectly centered.

I suspect this is their safety for the guy who posted the "Autopilot almost killed me!" YouTube video where he let autosteer have its way with him on a hilly, curvy road and almost plowed into an oncoming car. Tesla likely said "to reduce the chances of this, let's shift the car to the right side of the lane to reduce potential for a head-on collision".

I really, really, really hope this is temporary. IMHO it makes using autopilot in the left lane with traffic around a bit worse than uncomfortable, and I've emailed Tesla to let them know how I feel :).
 
The solution is actually very simple and should have been done from the start ! Allow the driver to adjust it position within the lane. So let the drive center the car while autopilot is active and use that position ! There is no position fit all situation and fit all driver style. So this would resolve pretty much all complaint.

The only issue with that is the AP hardware is actually pretty weak and I'm not sure any software upgrade will fix that. From my limited Tesla AP experience (loaner), the Mobileye system don't see really far and have issues extrapolating potential line position or even following a car in front of you.

For example, at street light where there is no line because of the intersection, it will try to follow the car in front, but it will zigzag around when the while line are not present, EVEN if the white line are present and visible just 10 meter ahead, and the car in front and yourself are going in a straight line.

In that instance, it actually easy to fix, you extrapolate the while line continuation for X number of meters and confirm that with the car in front.

But those are hack to a lack of good sensors, and there is limit to what you can do with those.
I specially like the underlined part.
 
The solution is actually very simple and should have been done from the start ! Allow the driver to adjust it position within the lane. So let the drive center the car while autopilot is active and use that position ! There is no position fit all situation and fit all driver style. So this would resolve pretty much all complaint.

It should be 100% suggestible, without AP disengaging if you pull it too far. The whole disengage by steering input thing is just a terrible human interface design.
 
Will the gradual development of high precision maps eventually fix this lane biasing problem for cars on the 1.0 sensor suite which will never get more cameras or any other type of better hardware?

Who knows. I'm not confident that that GPS in the car is accurate enough for that kind of precision, personally.

I think they just need to center the car in the lane and be done with it. I can't think of any cases where being actually centered in the lane would be a bad plan, while there are tons of reasons being biased to either side is a bad plan.

The car knows it's not centered, and it knows where center is. Put it there and be done with it.

- - - Updated - - -

It should be 100% suggestible, without AP disengaging if you pull it too far. The whole disengage by steering input thing is just a terrible human interface design.

I actually tried to nudge autosteer away from the line it was hugging a few times with just enough force to not disengage but enough to move the car to the left a little. Then autosteer comes back for revenge by pulling hard to the right to compensate... and move back towards the other lane that has a vehicle it knows is there. If you actually let go when it goes to do this is will jump all the way over the line and partly into the other lane before correcting.

Long story short, don't try to move the car while autosteer is on. Just take over, and re-engage when you're safe. I don't think the car has the ability to differentiate driver input vs road curvature or other conditions that can cause steering to pull to a side.
 
Not that bad if you consider that the driver might have to do emergency maneuver without fighting the AS too hard, and without having to touch another button to disengage AS.

The steering motor can probably overpower any human. The only reason it doesn't is because the torque sensors, which I assume are exactly the same that are used for electric assist steering along with some code don't let it. Thus you should be able to limit the amount of torque needed to override without having the computer give up completely.

If implemented this should feel like while your hand is on the wheel you're driving but maybe getting gently nudged. If you take your hands of the wheel, computer has full authority.

- - - Updated - - -

I actually tried to nudge autosteer away from the line it was hugging a few times with just enough force to not disengage but enough to move the car to the left a little. Then autosteer comes back for revenge by pulling hard to the right to compensate... and move back towards the other lane that has a vehicle it knows is there. If you actually let go when it goes to do this is will jump all the way over the line and partly into the other lane before correcting.
I call this "my tesla is drunk"

Long story short, don't try to move the car while autosteer is on. Just take over, and re-engage when you're safe. I don't think the car has the ability to differentiate driver input vs road curvature or other conditions that can cause steering to pull to a side.
It must. I don't see another way to design it. Otherwise the road curvature would be electrically boosting the steering wheel force back to your hands.