Annoying FSD/autopilot behavior with on-ramps

[rpo]

I've had a couple of Tesla's now with autopilot and have had the beta FSD from the beginning. This behavior happens with all of them going back years and years, and I don't know why a software fix has not been implemented, but hopefully others have experienced it and not just me. I made some crude Microsoft Paint drawings too. This occurs on highway/freeways with multiple lanes when driving in the far right lane going past an onramp. A human driver can see the onramp and just stays straight down the middle of the lane, and FSD/autopilot should do the same as the camera can see the onramp and the lines merging back together down the road. That logical path is marked with the blue line in the drawing below and results in no odd directional changes:

Inextricably, this is NOT what Tesla's software has the car do. It veers hard right the moment the onramp merges in to center the car and then veers back left (which can cause cars in the lane to the left to think you are drunk, steering into them, etc). It also makes passengers wonder why the car is veering back and forth for no reason. Why, Tesla, why??? The car can clearly see the lane goes back to its original width in 100-200 meters at most. Maintain the distance to the dotted line to the left and continue the course instead of doing this:

 

[Daniel in SD]

This is one of the hardest problems for AI. I expect this will be the last step before robotaxis.

Seriously though people have been complaining about this for years, it used to be even worse. It must be pretty hard to solve.

 

[Dewg]

FSD Beta handles it a bit better than the older AP stack. I don't think Tesla is focusing on any improvement to the AP stack, and hasn't been in some time.

 

[DCGOO]

I've had a couple of Tesla's now with autopilot and have had the beta FSD from the beginning. This behavior happens with all of them going back years and years, and I don't know why a software fix has not been implemented, but hopefully others have experienced it and not just me. I made some crude Microsoft Paint drawings too. This occurs on highway/freeways with multiple lanes when driving in the far right lane going past an onramp. A human driver can see the onramp and just stays straight down the middle of the lane, and FSD/autopilot should do the same as the camera can see the onramp and the lines merging back together down the road. That logical path is marked with the blue line in the drawing below and results in no odd directional changes:

View attachment 867423

Inextricably, this is NOT what Tesla's software has the car do. It veers hard right the moment the onramp merges in to center the car and then veers back left (which can cause cars in the lane to the left to think you are drunk, steering into them, etc). It also makes passengers wonder why the car is veering back and forth for no reason. Why, Tesla, why??? The car can clearly see the lane goes back to its original width in 100-200 meters at most. Maintain the distance to the dotted line to the left and continue the course instead of doing this:

View attachment 867430

It is because the lane markings are wrong. There is supposed to be a dotted/dashed line across the end of the entrance ramp. If that was there, the car would not swerve to re-center itself in the briefly much wider lane.

 

[EVRider-FL]

Not as bad as it used to be (with non-Beta EAP/FSD), but it still happens. As the previous poster said, not an issue when the merge lane is marked with dashed lines.

 

[johnnycnote]

As mentioned previously, the car believes that the road became wider, so it centers itself between the lines. This also happens with FSDb on wider than usual residential streets without line markings other than a center divider. It plops itself straight in the the middle of the road annoying other drivers that want to pass

 

[Individoo]

It is because the lane markings are wrong. There is supposed to be a dotted/dashed line across the end of the entrance ramp. If that was there, the car would not swerve to re-center itself in the briefly much wider lane.

I get the same behavior on all the highways around Vancouver. Other than the yellow 'merge' sign on the side of the road, there's no other indication that the lanes are merging. The moment the middle lane marking is gone, the car comes to the center of the 'lane'. Having the car in front of me driving 'properly' sometimes makes this behavior less pronounced, as Tesla tries to figure out which one to follow. I miss the (often not always) better designed European roads

 

[rpo]

It is because the lane markings are wrong. There is supposed to be a dotted/dashed line across the end of the entrance ramp. If that was there, the car would not swerve to re-center itself in the briefly much wider lane.

Most states don't have dotted/dashed lines there. Some do, and it works fine in those areas.

 

[rpo]

As mentioned previously, the car believes that the road became wider, so it centers itself between the lines. This also happens with FSDb on wider than usual residential streets without line markings other than a center divider. It plops itself straight in the the middle of the road annoying other drivers that want to pass

If it only happened in locations where the car was unable to see that the onramp lane merges in to the point that the lane markings become the same width as prior to the merge, then sure. But it happens even on short merges where the car can see the lane width goes back to the same width as before...in which case it should stick to the same distance to the dotted line on the left until the solid white line rejoins at the regular width. This would be an easy thing to program so it only follows the dotted line when a lane merges in from the right.

 

[RTPEV]

Yes, this is a HUGE problem and I reported it for years (obviously with no success). Eventually I just gave up.

Seriously though people have been complaining about this for years, it used to be even worse. It must be pretty hard to solve.

No, not hard at all. Map data should indicate that there is a merge there and the software can know that it needs to just maintain position against the left lane marking in this area. Or, it can simply see that the lane it is seeing is wider than normal and apply that logic to hug the left side as opposed to center. My wife's VW ID.4 handles this easily, so it is not a hard problem.

FSD Beta handles it a bit better than the older AP stack. I don't think Tesla is focusing on any improvement to the AP stack, and hasn't been in some time.

How do we know that if FSD Beta is not used on highways? Are you talking about similar merge lanes on "city" streets? Given that those merge lanes are likely much shorter than highway onramps, I'm not sure we can make an apples to apples comparison, but maybe there are cases that are similar.

It is because the lane markings are wrong. There is supposed to be a dotted/dashed line across the end of the entrance ramp. If that was there, the car would not swerve to re-center itself in the briefly much wider lane.

Well, wrong may not be the right word. Some states (like NC) do mark all the way to the end of the ramp. Other states (like VA and PA) do not. It's a local variation. The car needs to handle different regions. Sometimes I do think that that's the attitude of the Tesla engineers: Well, here we have lane markings that go all the way to the end of the lane, so that's the "correct" marking and that's all we have to handle. Everyone else is wrong. Well, sorry, but you have to handle regional variances, and like I said above, I really don't think it's that difficult a problem to solve, as my wife's car (and I suspect other brands as well) don't seem to have a problem with it.

And yes, at times it is downright dangerous depending on the road geometry. Particularly when the on ramp is on a right curve. As the car recenters itself, the movement to the left is pretty severe and I imagine scares the crap out of cars passing in the left lane.

 

[Individoo]

If it only happened in locations where the car was unable to see that the onramp lane merges in to the point that the lane markings become the same width as prior to the merge, then sure. But it happens even on short merges where the car can see the lane width goes back to the same width as before...in which case it should stick to the same distance to the dotted line on the left until the solid white line rejoins at the regular width. This would be an easy thing to program so it only follows the dotted line when a lane merges in from the right.

Once the same code will run both on the city streets and highway (one day in the future, right?), I hope it will improve. During AI day 2 they showed the better understanding of road layout and I see that the car might get better at determining the driving direction on these merges.

 

[tivoboy]

As you state, this has been going on since the beginning, not really sure why it can’t factor in the lane divergence and favor the LEFT lane marking as the tracking line for a brief moment.

What I can’t stand is that I SITLL cannot get the car to take the on-ramp when getting on the freeway and ACCELERATE AT ALL to get up to speed and merge into the right slow lane of travel - even when NOBODY else is there. It just seem way too timid at getting ON the highway for solid use.

 

[2101Guy]

It is because the lane markings are wrong. There is supposed to be a dotted/dashed line across the end of the entrance ramp. If that was there, the car would not swerve to re-center itself in the briefly much wider lane.

and this...is why (well, one reason) why I believe that these are true:

The auto industry's $75B bet on autonomy is not paying off

Many self-driving companies raised tens of billions of dollars long before their technology was proven or their businesses came close to being self-sustaining.

europe.autonews.com

Driverless Car Development Sets Ablaze a Bonfire of Billions

Following the money and the footsteps leaving the autonomous-vehicle industry.

www.bloomberg.com

Far too many edge cases combined with situations where things arent "right" or "perfect", where humans can adapt adjust to the "wrongness", but technology cant do that yet...and wont for some years to come.

 

[2101Guy]

This is one of the hardest problems for AI. I expect this will be the last step before robotaxis.

Seriously though people have been complaining about this for years, it used to be even worse. It must be pretty hard to solve.

I get that the focus is on FSD, but there have been MANY AP updates pushed over the "years". One would think this would be easy to solve...

So this isnt a problem AT ALL for FSD Beta? Its been totally solved?

 

[cbdream99]

Yes, this is one of the annoying thing … there is one short section of interstate entrance here, around couple of hundreds feet that goes wide, narrow and then wide again before going normal, people must be thinking I am drunk. However, there is another merge junction where two lanes merged into one for the on-ramp, there are no markings, it used to center early but now it consistently states on the left side until it reaches the end of the merge, which equally annoying and dangerous …. most Texas impatience drivers seeing the opportunity will speed pass me since the Tesla hangs to the far left.

 

[DCGOO]

Well, wrong may not be the right word. Some states (like NC) do mark all the way to the end of the ramp. Other states (like VA and PA) do not. It's a local variation. The car needs to handle different regions. Sometimes I do think that that's the attitude of the Tesla engineers: Well, here we have lane markings that go all the way to the end of the lane, so that's the "correct" marking and that's all we have to handle. Everyone else is wrong. Well, sorry, but you have to handle regional variances, and like I said above, I really don't think it's that difficult a problem to solve, as my wife's car (and I suspect other brands as well) don't seem to have a problem with it.

I think it may be an issue with timing of the construction. In my state, older ramps are often incorrect. New construction or repairs, almost always are marked correctly and compliant with current DOT standards.

 

[sleepydoc]

It’s pretty clear that the algorithm tries to center the car in the lane. When an entrance ramp occurs the ’lane‘ that the car perceives widens. As a human you say ‘no, the extra space on the right is just where the ramp is joining and it disappears in a bit. The lane is still over here,’ but FSD doesn’t do that.

My impression is that part of the problem stems from how far ahead the car is looking. A human tends to see both (the wider area as well as the narrower area further ahead) and aim for the narrower area because that makes sense. FSD focuses more on the perceived lane immediately in front of the car. THe most recent version of FSD seems to try and average the two a bit more but it leads to some other odd behaviors, causing it to sway back and forth a bit.

 

[ThisIsMask]

There's another side effect is the car entering freeway thinks I'm blocking them to pass . However, I'm doing workaround for this by turning off autosteer (keeps smart cruise) at those merges and re-engage afterwards.

 

[enemji]

I've had a couple of Tesla's now with autopilot and have had the beta FSD from the beginning. This behavior happens with all of them going back years and years, and I don't know why a software fix has not been implemented, but hopefully others have experienced it and not just me. I made some crude Microsoft Paint drawings too. This occurs on highway/freeways with multiple lanes when driving in the far right lane going past an onramp. A human driver can see the onramp and just stays straight down the middle of the lane, and FSD/autopilot should do the same as the camera can see the onramp and the lines merging back together down the road. That logical path is marked with the blue line in the drawing below and results in no odd directional changes:

View attachment 867423

Inextricably, this is NOT what Tesla's software has the car do. It veers hard right the moment the onramp merges in to center the car and then veers back left (which can cause cars in the lane to the left to think you are drunk, steering into them, etc). It also makes passengers wonder why the car is veering back and forth for no reason. Why, Tesla, why??? The car can clearly see the lane goes back to its original width in 100-200 meters at most. Maintain the distance to the dotted line to the left and continue the course instead of doing this:

View attachment 867430

Simply because Tesla is designed to stay in the center of two white lines ie the blue line that you drew.

 

[KArnold]

Simply because Tesla is designed to stay in the center of two white lines ie the blue line that you drew.

A.k.a. yet another bug.