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

FSD Beta hugging left centerline

This site may earn commission on affiliate links.
What if there are multiple cars parked on the right ? Would you wave in and out or would you just drive straight towards the center avoiding all the parked cars ?

It should stay toward the left if there are multiple cars. It should move back over when the string of cars is gone. Basically it should be making the call to move back to the right based on how far it can see.

What is there are kids playing on the sidewalk or there are loosely leashed dogs - would you still drive as if the road had no obstacles ?

It should move over to the center if there are VRUs near the drivable boundary. It should move back over to the right when it no longer detects obstacles as far as it can see.
 
  • Like
Reactions: Phlier
It should stay toward the left if there are multiple cars. It should move back over when the string of cars is gone. Basically it should be making the call to move back to the right based on how far it can see.

But how many cars - 5 cars parked 100 ft from each other or 4 cars with 50ft in between ? You can see how this gets tricky ...

It should move over to the center if there are VRUs near the drivable boundary. It should move back over to the right when it no longer detects obstacles as far as it can see.

This is what I meant by what are all the conditions when the car should drive towards the center rather than left. Undoubtedly there are others.

Now the question is - how do you implement all this without hardcoding. After all you want a generic solution where the cost-optimizer chooses the best path.
 
But how many cars - 5 cars parked 100 ft from each other or 4 cars with 50ft in between ? You can see how this gets tricky ...



This is what I meant by what are all the conditions when the car should drive towards the center rather than left. Undoubtedly there are others.

Now the question is - how do you implement all this without hardcoding. After all you want a generic solution where the cost-optimizer chooses the best path.
The thing that comes up again and again is context. The car needs to know the context of the road it is driving on to understand how best to optimize safety, and still get to the destination efficiently.
For myself, the 2 mile road I drive on every day is a good example of needing to basically have it all. It is 1.5 unmarked lanes for half of it, and 2 narrow marked lanes for the rest. There are several small landslides, where the road is either partially missing or the dirt is so far on the road that you must drive over the double yellow or scratch the car/hit a ditch.

The speed limit is 30-35 mph. Most people do less than that, especially where it's very bad. We also have a "temporary" bridge made of 2 train cars. I expect this bridge to be temporary forever.

For the first 0.5 miles the right and left are pretty rough, so staying right in the middle is common. Most everyone slows down here because of the blind corners.
The next 0.5 miles are very windy, so people usually try to position the car where they have best visibility. This can be either the left or right sides depending on location.
The next 0.5 miles start with the bridge crossing, where the car needs to put 2 wheels on each train car, and slow down for the speedbumps. At this point after the bridge, the road has lines, but almost immediately you come to a place where 1/3 of the road is missing and a pair of construction barricades is the only warning that you must cross into the oncoming lane.

The final 0.5 miles has a landslide where 1/3 of the road is missing, and several other parts are slightly too narrow, so many people drive about a foot over the double yellow, to avoid getting the car scratched.

When I look at this road, it really is amazing that the only complaints I currently have, are when it slows too much to 13-15 mph, and it doesn't always know that this corner you really want to be on the inside or outside depending on all the conditions above. Also, sometimes it will cross the double yellow completely on a blind corner.
 
  • Informative
Reactions: EVNow
I just got back from my test drive with 10.5 and it is no better than before. I recalibrated the cameras but that didn't do anything. On a two-lane road, the tires actually touch the double yellow line on straight-aways and cross the lines on curves (especially right curves). C'mon Tesla, with all the AI and technology, is it not possible to stay centered in the lane? Oddly, on multiple-laned roads, it has no problem.
 
  • Like
Reactions: Ixthys
But how many cars - 5 cars parked 100 ft from each other or 4 cars with 50ft in between ? You can see how this gets tricky ...

This is what I meant by what are all the conditions when the car should drive towards the center rather than left. Undoubtedly there are others.

Now the question is - how do you implement all this without hardcoding. After all you want a generic solution where the cost-optimizer chooses the best path.

I think using "as far as the car can see" is a good place to draw the line between staying more toward the center instead of weaving in and out due to gaps. Meaning if there are no obstructions as far as the car can see, it should gently slide back toward the right. If it sees things coming, it shouldn't move over but stay toward center.

Is pathing logic currently using hardcoded logic? If so, that has to change at some point. I would hope that AI inference makes the call. It should know the difference between a quiet suburban road and something like this:


There are TONS of suburban and exurban roads like the one I shared where there is hardly any stationary obstacles or VRUs. It doesn't make sense to have the car always toward the center as a long term solution. There's a reason why so many people are griping about this particular issue: the car frequently crosses the centerline. It should never do this unless going around something. Pre-10.5, the car just hung out partially crossed over the centerline. That is not safe, especially on a winding road and curving left. You have little visibility of an oncoming car coming around the bend.
 
  • Like
Reactions: Phlier
Is pathing logic currently using hardcoded logic? If so, that has to change at some point. I would hope that AI inference makes the call. It should know the difference between a quiet suburban road and something like this:


No - they use cost optimization techniques like Monte-Carlo simulation.

There are TONS of suburban and exurban roads like the one I shared where there is hardly any stationary obstacles or VRUs. It doesn't make sense to have the car always toward the center as a long term solution. There's a reason why so many people are griping about this particular issue: the car frequently crosses the centerline. It should never do this unless going around something. Pre-10.5, the car just hung out partially crossed over the centerline. That is not safe, especially on a winding road and curving left. You have little visibility of an oncoming car coming around the bend.
Right - that has been my point.

There are a lot of variations and conditions and the path would depend on those. There is no one "right" answer - everything depends. They need to use a lot of new parameters to figure out the right path in unmarked residential roads.

ps : One more thing. Width of the road - and strangely width of the shoulder. If the road is wider, then it makes more sense to be in the center of the estimated right lane. If it is narrow, may be more towards the center or even at the center ! We have roads here where side lines are marked but not a center line - and some of them have broad shoulders. Makes sense to hug the shoulder (which is what most people do - infact even crossing the shoulder at times). As we dig deeper there will be more and more variations.
 
Last edited:
  • Like
Reactions: nvx1977
I just got back from my test drive with 10.5 and it is no better than before. I recalibrated the cameras but that didn't do anything. On a two-lane road, the tires actually touch the double yellow line on straight-aways and cross the lines on curves (especially right curves). C'mon Tesla, with all the AI and technology, is it not possible to stay centered in the lane? Oddly, on multiple-laned roads, it has no problem.

Completely agree. Why would you ever touch, let alone cross, a yellow center line?

I drive in the country just as much as the city. Plenty of country roads with well-marked lines and no right-sided obstructions and yet the car still hugs the center line or crosses it.
 
  • Like
Reactions: Battouttahe77
On a narrow country road with a centerline, when I can see far and there is no other traffic, I drive on the centerline. Under the conditions that is safest, because all risks come from the sides of the road. What I'm trying to say is, it's not a very simple decision.

It may be illegal though. If I had to stay strictly legal, I would still drive as near as legally possible to the centerline. I have yet to receive a ticket though under such conditions.
 
  • Like
Reactions: EVNow
No better in 2021.36.8.8. (10.5). Still stays left toward center and slows on every approaching vehicle on 2-way roads. Camera calibration in the 2020 MSP??? Can it be done without Tesla service???

M
And I could only wish it was closer to the left lines.. i wish I could just adjust the FSD to try and move 6-10” to the LEFT from where it apparently wants to drive Which is WAY to the right.
Here in lies the issue. As I've been saying figuring out where to drive on unmarked roads is complicated.

Completely agree. Why would you ever touch, let alone cross, a yellow center line?
Because - they found something more undesirable on the other side ?

When there is a curve to the right - the car tends to cross the (imaginary or sometime actual) center line. My guess is this is because they have given somewhat lower cost to crossing the centerline than driver discomfort or hitting / being too close to an obstruction. So, for eg., if the car has to take too tight a turn, it might instead cross over the center line, if safe.

Ofcourse, it could just as well be inaccuracy in figuring out where the center line is.
 
Here in lies the issue. As I've been saying figuring out where to drive on unmarked roads is complicated.


Because - they found something more undesirable on the other side ?

When there is a curve to the right - the car tends to cross the (imaginary or sometime actual) center line. My guess is this is because they have given somewhat lower cost to crossing the centerline than driver discomfort or hitting / being too close to an obstruction. So, for eg., if the car has to take too tight a turn, it might instead cross over the center line, if safe.

Ofcourse, it could just as well be inaccuracy in figuring out where the center line is.
But, these are well WELL marked roads.. streets in town, going in and out of town, Menlo Park, Palo Alto, up and down the peninusula, there is no ambiguity of the well marked lines.
 
I rebooted and re-calibrated the cameras again (10.5) and now it stays more centered on the well marked two lane road I travel. It does tend to cut in on left hand turns but not as bad as before (it would go over the double yellow center line at times). It's still not perfect but very good currently.
 
My experience with 10.5 has been that it *finally* drives in the correct position on residential streets with no center line. This has solved most of my disengagements, because now I can finally let the car continue driving on residential streets when an oncoming car approaches. But reading many of the comments on this thread and others, it sounds like some think it's too far left and some think it's too far right, which makes me wonder if (a) it handles different widths of residential streets differently - or maybe there are other street attributes contributing to variations in lane centering, or (b) everyone has a different comfort zone for the best alignment on unmarked streets.
 
It's very weird that some people feel it stays too close to the right and some feel like it stays too close to the left - perhaps its a car-by-car issue? Here's my thoughts and experience:

Generally, an experienced driver is going to find the easiest/fastest line, even if driving under the speed limit - it just becomes second nature when driving. On a well marked, curvy road, this means cheating to the inside of the curve: towards the right-side of the road a little on right-hand curves and towards the center line on left-hand curves IF there is no oncoming traffic. The problem I see with beta 10.5 is that it wants to stay in the center of the lane all the time. However, for "comfort" there has been some hysteresis added into the steering algorithm so that it actually winds up cheating to the outside of the curve on curvy roads: towards the left side of the road on left-hand curves and towards the center line on right-hand curves. This results in the nail-biting occurrence of the car cheating to the outside of the curve - towards or on the center line - in right-hand curves while the "experienced" drivers of oncoming cars are also cheating towards the center line as they approach the curve (their inside). This is what Tesla needs to fix. The hysteresis for comfort should go away if there is oncoming traffic and the car should steer to the center of the lane (or more appropriately cheat to the inside of the curve a little, like a human driver would).
 
I was hoping 10.6.1 would fix this problem, but sadly it has not. I took my car out for a test drive today. We live on a fairly curvy, two-lane road and I engaged FSD. It immediately swerved towards the middle line and stayed there on the straight-away. On the right-hand curves, my tires cross into the oncoming lane and I had to quickly disengage to prevent a collision with an oncoming school bus. I have recalibrated my cameras but that didn't help. Why can't Tesla simply have the car track in the center of the lane? Tesla must address this issue or someone is going to be seriously injured or killed.
 
  • Like
Reactions: Ixthys
I was hoping 10.6.1 would fix this problem, but sadly it has not. I took my car out for a test drive today. We live on a fairly curvy, two-lane road and I engaged FSD. It immediately swerved towards the middle line and stayed there on the straight-away. On the right-hand curves, my tires cross into the oncoming lane and I had to quickly disengage to prevent a collision with an oncoming school bus. I have recalibrated my cameras but that didn't help. Why can't Tesla simply have the car track in the center of the lane? Tesla must address this issue or someone is going to be seriously injured or killed.
That's my experience with 10.6.1 as well. This and the phantom braking/slowing seem like fundamental problems that need to be addressed before moving on to "advanced" maneuvers. As my son said yesterday, the engineers seem to be engaged in a lot of "running before you can walk" type of development. An example of this is the problem with FSD immediately determining that a stopped car is "parked" and that it needs to go around, even on a two-lane road with double yellow line and oncoming traffic in the opposite lane. Yes, it is a safety hazzard that needs to be fixed, but it seems to me they should just make it stop behind the "parked" car and let the driver figure it out. I mean, there are obviously situations that the engineers have decided not to account for yet, like school zones and flashing yellow arrows. Why would they try crossing a double-yellow line into oncoming traffic to go around another car at this stage, whether it is parked or simply waiting to make a turn? Get the phantom braking/slowing and lane keeping nailed, then get it properly staying on correct path through intersections, and ONLY THEN move on to going around "parked" cars, managing school zones, and such.