Why is lane shifting so hard to figure out?

[TjckTock]

I am on the beta release and for the most part am amazed at what it does. However, what I would have thought super easy seems to be kryptonite for the FSD software: know which lane it should be in. This is, no exaggeration, the reason I have to take over 90% of the time. Examples: 1) I am in the right lane coming up on a right turn in 100 feet. Suddenly the Tesla thinks it needs to get in the left lane which would cause it to miss the turn. 2) Similarly, I have an upcoming left turn on a three lane road and the car wants to stay in the right lane until I practically arrive at the intersection and then try to make three lane shifts to get into the left turn lane. 3) Some times I am cruising along on the freeway and it will occasionally turn on the blinker and start a lane change and then change its mind. I could go on but basically more of the same. Thing is it isn't random but very repeatable - at least examples 1 and 2 are and occur every time I take a particular route. Some times, if I initiate a lane change to get in the proper lane it will stay there but some times it tries to go right back to the wrong lane. Seems like it should be pretty simple to tell the computer to get in the correct lane for an upcoming turn *and stay there* barring a disabled vehicle. I searched and have not found anyone else mention this so is it just me or have others observed the same?

With all the amazing decisions the car makes well I am at a loss why this is so troublesome.

 

[ChaoticNeutral]

I definitely experience similar things. It’ll randomly get into the left lane on a particular road to “follow route” even though we need to be in the right lane in five miles. Random “ghost” signaling when going around some curves. More often than not the car doesn’t signal for dual/zipper merges, nor does it usually signal to get into a turn lane.

 

[VanFriscia]

Yes, many have noted similar behavior in other threads about the Beta. I agree with your sentiments - for all of the difficult problems there are to solve with self-driving, this seems like it would be fairly trivial. I mean, the nav knows which lane you need to be in, so why doesn't FSD...?

 

[ChaoticNeutral]

Oh, and the most annoying, my daily struggle of FSD thinking that the right “lane” of a road by my home, which has been converted into parking ONLY, is still a driving lane. It’s been a parking lane for well over a year, and all the spots are clearly marked with parking symbols and white lines. It also thinks that the lined “no parking” area/segment before the street in the picture is a turn lane, that is not a turn lane.

 

[Missile Toad]

It took the car 5 years to get to this point. You guys want it to have common sense too? I'm sure it will be an upgrade to FSD.... perhaps for another upcharge of $2,000? Which you could probably pay for by end of 2022, only to get in 202.... errr... 203_ something.

 

[diplomat33]

I am on the beta release and for the most part am amazed at what it does. However, what I would have thought super easy seems to be kryptonite for the FSD software: know which lane it should be in. This is, no exaggeration, the reason I have to take over 90% of the time. Examples: 1) I am in the right lane coming up on a right turn in 100 feet. Suddenly the Tesla thinks it needs to get in the left lane which would cause it to miss the turn. 2) Similarly, I have an upcoming left turn on a three lane road and the car wants to stay in the right lane until I practically arrive at the intersection and then try to make three lane shifts to get into the left turn lane. 3) Some times I am cruising along on the freeway and it will occasionally turn on the blinker and start a lane change and then change its mind. I could go on but basically more of the same. Thing is it isn't random but very repeatable - at least examples 1 and 2 are and occur every time I take a particular route. Some times, if I initiate a lane change to get in the proper lane it will stay there but some times it tries to go right back to the wrong lane. Seems like it should be pretty simple to tell the computer to get in the correct lane for an upcoming turn *and stay there* barring a disabled vehicle. I searched and have not found anyone else mention this so is it just me or have others observed the same?

With all the amazing decisions the car makes well I am at a loss why this is so troublesome.

This happens to me quite a bit. I think the cause might be bad map data and/or bad localization. The map data is bad, the car thinks it is in the "wrong" lane so it tries to lane change to get into the "right" lane. Or if the localization is off, the car thinks it is in a different lane so it tries to lane change to get into the "right" lane.

 

[EVEnergy]

It was better a couple of years ago (back in 2019-2020) and I didn't have as many issues with it as I do now. This is definitely something that has regressed in the more recent updates.

 

[clydeiii]

Yes, many have noted similar behavior in other threads about the Beta. I agree with your sentiments - for all of the difficult problems there are to solve with self-driving, this seems like it would be fairly trivial. I mean, the nav knows which lane you need to be in, so why doesn't FSD...?

The "nav" knowing which lane you need to be in is very very different from the car being able to determine which lane is which in the real world. Think about it: the car needs to translate 2D mapping data into a 3D real world. That's a massively challenging problem.

 

[OncomingStorm]

I am on the beta release and for the most part am amazed at what it does. However, what I would have thought super easy seems to be kryptonite for the FSD software: know which lane it should be in. This is, no exaggeration, the reason I have to take over 90% of the time. Examples: 1) I am in the right lane coming up on a right turn in 100 feet. Suddenly the Tesla thinks it needs to get in the left lane which would cause it to miss the turn. 2) Similarly, I have an upcoming left turn on a three lane road and the car wants to stay in the right lane until I practically arrive at the intersection and then try to make three lane shifts to get into the left turn lane. 3) Some times I am cruising along on the freeway and it will occasionally turn on the blinker and start a lane change and then change its mind. I could go on but basically more of the same. Thing is it isn't random but very repeatable - at least examples 1 and 2 are and occur every time I take a particular route. Some times, if I initiate a lane change to get in the proper lane it will stay there but some times it tries to go right back to the wrong lane. Seems like it should be pretty simple to tell the computer to get in the correct lane for an upcoming turn *and stay there* barring a disabled vehicle. I searched and have not found anyone else mention this so is it just me or have others observed the same?

With all the amazing decisions the car makes well I am at a loss why this is so troublesome.

from my observations my guess, and its just that a guess, is some of this behavior is a result of the computer trying to predict the future. my assumption is its constantly running probability scenarios for where various objects, both seen and unseen, might end up in 5, 10, 20 seconds. given the complexity of the task, the uncertain nature of the real world and, as others have pointed out, possibly having bad/poor map data and likely other factors, determining the "best" current position is likely to be a tough call. and anytime the probabilities are such that it's an equal chance, the car has a hard time deciding, or to put it another way, the fluctuating probabilities are going from 49-51% and back again (or whatever the percentile threshold is for the given task). basically the car is trying to position itself for a future that doesn't end up happening and then tries to correct lol.

given the nature of the beast, so to speak, i would expect this behavior to be fairly persistent until the final stages of development since it likely has a negligible impact on the overall stats of all the cars in the fleet. but like many things it does that technically work but arn't great, it will eventually get much better and we'll likely see minor improvements between now and then.

 

[TjckTock]

The "nav" knowing which lane you need to be in is very very different from the car being able to determine which lane is which in the real world. Think about it: the car needs to translate 2D mapping data into a 3D real world. That's a massively challenging problem.

From what I can tell the mapping is working just fine. The graphical interface is correctly representing all the lanes - the car just decides to move into the wrong one.

 

[TjckTock]

from my observations my guess, and its just that a guess, is some of this behavior is a result of the computer trying to predict the future. my assumption is its constantly running probability scenarios for where various objects, both seen and unseen, might end up in 5, 10, 20 seconds. given the complexity of the task, the uncertain nature of the real world and, as others have pointed out, possibly having bad/poor map data and likely other factors, determining the "best" current position is likely to be a tough call. and anytime the probabilities are such that it's an equal chance, the car has a hard time deciding, or to put it another way, the fluctuating probabilities are going from 49-51% and back again (or whatever the percentile threshold is for the given task). basically the car is trying to position itself for a future that doesn't end up happening and then tries to correct lol.

given the nature of the beast, so to speak, i would expect this behavior to be fairly persistent until the final stages of development since it likely has a negligible impact on the overall stats of all the cars in the fleet. but like many things it does that technically work but arn't great, it will eventually get much better and we'll likely see minor improvements between now and then.

Yeah, the only reason I can think for it to do this is they want a provision to cope with a disabled vehicle in the correct lane and perhaps that is the future scenario it is improperly identifying and switching lanes to avoid. At least that could explain scenario #1. Doesn't really explain #2, though.

 

[totoro722]

This happens to me quite a bit. I think the cause might be bad map data and/or bad localization. The map data is bad, the car thinks it is in the "wrong" lane so it tries to lane change to get into the "right" lane. Or if the localization is off, the car thinks it is in a different lane so it tries to lane change to get into the "right" lane.

That is my best guess too. Problems of this sort I have tend to be in the same predictable places, suggesting the FSD is making decisions based on bad data for those locations.

 

[OncomingStorm]

Yeah, the only reason I can think for it to do this is they want a provision to cope with a disabled vehicle in the correct lane and perhaps that is the future scenario it is improperly identifying and switching lanes to avoid. At least that could explain scenario #1. Doesn't really explain #2, though.

yeah the second one is likely a map issue or some other static data that confuses the AI, since it happens in the same spot every time, i was more referring to the erratic lane changes.... always the problem with stuff like this tho, there is likely dozens of different issues that result in the same symptom

 

[TjckTock]

given the nature of the beast, so to speak, i would expect this behavior to be fairly persistent until the final stages of development since it likely has a negligible impact on the overall stats of all the cars in the fleet. but like many things it does that technically work but arn't great, it will eventually get much better and we'll likely see minor improvements between now and then.

I don't know. One metric I think they must track is how often the driver has to disengage FSD and take over. If they fixed this, that number for me would drop by an order of magnitude. Unless they do it on purpose to keep the driver attentive... ;-)

 

[S4WRXTTCS]

Because Tesla refuses to use maps in the way they should use maps.

 

[S4WRXTTCS]

The "nav" knowing which lane you need to be in is very very different from the car being able to determine which lane is which in the real world. Think about it: the car needs to translate 2D mapping data into a 3D real world. That's a massively challenging problem.

When it comes to problems like this you have choices.

You can pre-maps the road with high accuracy sensors like Lidar (or the equivalent), and be highly maps based. Or you can try to go a more human style method of visually mapping your position to the lane in the maps.

The problem with Tesla is they don't want to be geo fenced to only pre-mapped roads. So they attempt to do the second one, but both the vision along with the nav maps are really bad.

I can't even follow the nav maps while manually driving without encountering numerous areas where its telling me the wrong thing to do. It's also no secret that FSD Beta testers in California have better results because there maps are better.

At least half the issues I encounter are mapping issues or issues correlating my actual position with a position on the map.

Things like turning on the turn signal at a stop sign, but not actually turning. It turns them on as it thinks its changing lanes as after the stop sign there are two lanes. Human drivers know they can choose either lane, and obviously don't turn on the turn signal.

 

[OncomingStorm]

I don't know. One metric I think they must track is how often the driver has to disengage FSD and take over. If they fixed this, that number for me would drop by an order of magnitude. Unless they do it on purpose to keep the driver attentive... ;-)

honestly as harsh as it might sound i'm guessing they just dont care lol. gotta remember they have their sights set on the end goal, trivial things that dont feel right but still function within parameters would be treated as a success for the purposes of the beta... at this stage they are focused on core functionality, driver comfort, not irritating people around you and general "odd" behavior is going to be ignored throughout the bulk of the project, so long as the end results are sufficient.

its only in the latter stages, once they are happy and confident their model works as intended, that they would then start to tweak it and make it feel more natural and less robotic or w/e, otherwise you risk making tweaks to a model you end up dumping.... the nice part is that last stage will go fast since its tweaks to exsisting functionality, not new functionality, but that is little solace at this point i suppose lol.

 

[EVNow]

From what I can tell the mapping is working just fine. The graphical interface is correctly representing all the lanes - the car just decides to move into the wrong one.

Yep - it is not useful to just blame localization problem.

I'd suggest everyone who sees a problem with the choice of lanes to carefully see whats happening on the visualization (better to record since in real time it is more important to pay attention to the road). Then, look at OSM and other maps to see if there are mapping issues.

I rarely see localization issues - mostly planning errors.

Seems like it should be pretty simple to tell the computer to get in the correct lane for an upcoming turn *and stay there* barring a disabled vehicle.

It would seem so - but its not that simple.

For one thing - the whole planning is not "hard coded". You can't ever hard code and find a decent answer for millions of combinations that can happen. So, they use some kind of cost-optimization (one of the items they will rewrite in Beta 11). So, depending on the costs given, the optimizer can come up with wrong path. This could be because the costs are wrong in some way or because the optimizer found a weird local minima and got "stuck" there.

It is easy to find solutions for specific scenarios - but difficult to find optimal costs that make planning better in this particular scenario without degrading other scenarios.

 

[cinnamon_owl]

This happens to me quite a bit. I think the cause might be bad map data and/or bad localization. The map data is bad, the car thinks it is in the "wrong" lane so it tries to lane change to get into the "right" lane. Or if the localization is off, the car thinks it is in a different lane so it tries to lane change to get into the "right" lane.

Yes! This is my theory as well. There seems to be a LOT of outdated map data causing the problem. Especially on freeways in the Seattle area where it thinks there are actually fewer lanes on the freeway than there are.

 

[Twiglett]

I'm fine with them not relying on map data, it can't be up to date enough, but having it see everything for the first time all the time is problematic, especially when it can't read road signs that are warning what's up the road.
Two examples of crappy/out of date map data are
#1 - the 50mph road with a traffic light after a corner. Lots of road signs about the impending light, but FSDb is still blindly doing 54mph around the corner followed by very hard braking when the light is red - or extremely hard braking if there are other cars waiting.

#2 - It still slows down to 25mph for a removed 4-way stop on a 40mph road near here. Then just after that there's a traffic light that used to be a 4-way stop. Several times its started to move on a red light.
But it would be really nice if it had a local knowledge cache to give it some hysteresis - that and automatically rejecting map data when it sees traffic cones