FSD almost caused a crash..

[Docpaulo]

I was driving today on a slight dow hill, 2 lane divided highway, i was on the inner fast lane.. 55mph

A short 3rd lane appears on my left, for a left turn only through the highway divider

Without warning, car enters the left turn lane for no reason then slams on brake as road disappears in front of me.. i was able to take over and maneuver to avoid the median..

There were no vehicles ahead of me but some behind me. There was no clear reason why it switched lanes as there was no one to pass and my route was straight for 5 miles until i needed to turn left...

I was holding the wheel and paying attention but it still caught me offguard...

Somehow car makes poor lane changing decisions... hope they fix this...


They should allow us to disable all lane changes except to exit or turn...

 

[Tam]

I was driving today on a slight dow hill, 2 lane divided highway, i was on the inner fast lane.. 55mph

A short 3rd lane appears on my left, for a left turn only through the highway divider

Without warning, car enters the left turn lane for no reason then slams on brake as road disappears in front of me.. i was able to take over and maneuver to avoid the median..

There were no vehicles ahead of me but some behind me. There was no clear reason why it switched lanes as there was no one to pass and my route was straight for 5 miles until i needed to turn left...

I was holding the wheel and paying attention but it still caught me offguard...

Somehow car makes poor lane changing decisions... hope they fix this...


They should allow us to disable all lane changes except to exit or turn...

It doesn't sound like an Auto Lane Change with an automatic left signal.

It's a fork-in-the-road decision:

Historically, Tesla automation system might have a hard time to choose which lane to continue when there's another lane added (in your case: 2 lanes become 3, with the left as the left turn only lane).

It's easy for humans just go straight and not deviate into the left but not for an expensive high-tech AI machine.

Tesla should have fixed these kinds of problems a long time ago. Now, it's been 9 years since the introduction of the first Autopilot and I am not sure another 9 years will make any difference.

 

[EVRider-FL]

I assume you’re talking about FSD Beta. Whenever you engage it, you can press the right scroll wheel left or right and select Minimize Lane Changes, which will help reduce (but not completely eliminate) unnecessary lane changes. Moving the scroll wheel will also change the following distance setting (Chill, Average, Aggressive), so you might need to adjust that too.

 

[RTPEV]

They should allow us to disable all lane changes except to exit or turn...

There is such a setting in the latest FSD software (Minimize Lane Changes).

If you had that on, and it still tried to make the lane change, it was apparently confused about which lane was the "through" lane even though it was probably obvious to you. This is a continuing problem with FSD...it even makes these kinds of weird intentions to change lanes in my neighborhood roads with left turn only lanes (it very briefly puts on the blinker like it wants to change lanes, but then figures it out and cancels).

This is an excellent opportunity for you to be able to 'teach' the system by using the feedback capability when you disengaged. One would hope that they would use your snippet as part of their training so they can learn that and similar lane splits.

 

[zoomer0056]

I was holding the wheel and paying attention but it still caught me offguard...

I don't have FSD. Does the car signal before a lane change? This would be a clear time to over ride, before the change occurs.

 

[RTPEV]

I don't have FSD. Does the car signal before a lane change? This would be a clear time to over ride, before the change occurs.

Normally yes, but in this case it sounds more like the car was seriously confused about the lanes. This happens as well where it suddenly and unexpectedly pops into an alternate lane at a lane fork situation.

 

[ewoodrick]

I was driving today on a slight dow hill, 2 lane divided highway, i was on the inner fast lane.. 55mph

A short 3rd lane appears on my left, for a left turn only through the highway divider

Without warning, car enters the left turn lane for no reason then slams on brake as road disappears in front of me.. i was able to take over and maneuver to avoid the median..

Yes, that's why FSD requires driver's attention.
Things like this happen periodically.
When you took over, did you tell the car why when it asked?

 

[SidetrackedSue]

Drive home last night, undivided highway, one lane each direction, speed limit 80kph/50mph. V11.4.2. My incidents 1 & 4 were similar. In all, the drive was pretty good and the fact it was dusk (and extremely cloudy due to storms in the area) didn't affect the confidence of the car. It also didn't flinch at all the lightening happening around us.

1. Car signals right (3 - 4 clicks) when a right turn lane appears, but doesn't attempt to change lanes.

2. Lane splits into a two to create a passing lane (priority use for my direction but can be used in spots by the opposite direction if there are no oncoming cars using it). Car chooses left lane. I signal to change lanes to the right.

3. Car flawlessly handles the merge back but at the last moment. There was no car attempting to overtake us at the end of the passing lane. This is notable and a success because in the past (AP or prior to V11.3.x) with no other car around, that has been has been a complete fail, either continuing on the shoulder or panic TAKE OVER!

I never let AP/FSD make the choice of when to merge out of the ending lane if there's a car coming up on my left so can't say it would work as flawlessly if there was a car in the left lane. Years ago, on cruise control at the speed limit in an ICE, I was run off the road with no shoulder in front of me, just the start of a bridge railing. As a result, I no longer trust the overtaking car to let me in. In that instance, the emergency braking disabled cruise control until the car was turned off and on again. So I'm a bit gun-shy and proactive at the end of passing lanes. That has nothing to do with FSD but is why I won't test its ability in that scenario.

Part of the problem for FSD is that the line drawings in my province show a dashed line up until the end of the merge. In other provinces, when the zipper merge of traffic should be happening, the dashed line disappears. In that case, AP/FSD safely centres the car between the two solid lines and smoothly merges. This is an example of how jurisdictions hinder/aid AV success.

4. Car signals right and moves into a right turn lane and then back into the straight lane as it ends. In this case, there's no message as to why it made that choice but was the same behaviour as in #1, except in this case the right turn lane was more than a few car lengths so it had time to make the move in and out.

5. There were major electrical storms and I came to a traffic light with power restored but not working so it was flashing yellow for me, red for the cross road. The car slowed dramatically but the message said "stopping for traffic light" so it obviously interpreted it as an amber. It accepted being nudged through with the go-pedal. This is an extreme edge case so I don't fault FSDb at all in this case.

A non-FSDb observation: Once we were into the storm, all driver assist was cancelled due to the heavy rain. I got to experience "highway carwash mode" once again (where the deep, pooled, water splashes up over the windscreen to completely blind the driver.) This is something we've never had in any other car, including the Prius. The water appears to be coming from the tires (no other vehicles were near), so best guess is the aerodynamic design and bigger wheels pull the water over the car when it is moving fast enough. Pooled water once we were on city streets at 30 mph did the normal wave action I'd expect. My first time experiencing this, I was only going 40 mph on a highway, so maybe it is just depth and not speed (that was even worse torrential rain than last night.)

 

[Supcom]

With FSDb 11.4.4, I find my car occasionally moving into a left turn lane in the middle of a block. Usually, it swerves right back into the correct lane - presumably when it sees that the lane does not continue on. This is an old defect that has reappeared in 11.4 with the lane selection "improvements". Perhaps a neural network replaced some old procedural code and needs to watch a few more training videos.

I expect Tesla is getting plenty of "safety critical" disengagement videos on this.

 

[ewoodrick]

Drive home last night, undivided highway, one lane each direction, speed limit 80kph/50mph. V11.4.2. My incidents 1 & 4 were similar. In all, the drive was pretty good and the fact it was dusk (and extremely cloudy due to storms in the area) didn't affect the confidence of the car. It also didn't flinch at all the lightening happening around us.



Part of the problem for FSD is that the line drawings in my province show a dashed line up until the end of the merge. In other provinces, when the zipper merge of traffic should be happening, the dashed line disappears. In that case, AP/FSD safely centres the car between the two solid lines and smoothly merges. This is an example of how jurisdictions hinder/aid AV success.



5. There were major electrical storms and I came to a traffic light with power restored but not working so it was flashing yellow for me, red for the cross road. The car slowed dramatically but the message said "stopping for traffic light" so it obviously interpreted it as an amber. It accepted being nudged through with the go-pedal. This is an extreme edge case so I don't fault FSDb at all in this case.

A non-FSDb observation: Once we were into the storm, all driver assist was cancelled due to the heavy rain. I got to experience "highway carwash mode" once again (where the deep, pooled, water splashes up over the windscreen to completely blind the driver.) This is something we've never had in any other car, including the Prius. The water appears to be coming from the tires (no other vehicles were near), so best guess is the aerodynamic design and bigger wheels pull the water over the car when it is moving fast enough. Pooled water once we were on city streets at 30 mph did the normal wave action I'd expect. My first time experiencing this, I was only going 40 mph on a highway, so maybe it is just depth and not speed (that was even worse torrential rain than last night.)

The car does not know how to correctly handle lane endings, not does it know how to correctly use temporary passing lanes. And it definitely doesn't read the lane ending signs.

If the light was flashing amber (yellow?) then the car didn't really act correctly. Flashing yellow is a caution indicator (the other direction should have been flashing red). The car should have slowed, watched for traffic and continued.

Evidently you don't get that many storms. It's just as easy for the car next to you to make a blinding splash all over your windshield. The only reaction is basically no reaction.

 

[RTPEV]

Part of the problem for FSD is that the line drawings in my province show a dashed line up until the end of the merge. In other provinces, when the zipper merge of traffic should be happening, the dashed line disappears. In that case, AP/FSD safely centres the car between the two solid lines and smoothly merges. This is an example of how jurisdictions hinder/aid AV success.

I can't tell exactly what your thoughts are, but it seems like you prefer the unmarked merge lanes versus how it's done in your province.

We have the same difference here: in North Carolina the merge lanes are dashed right up to the end, but in Virginia they are not, creating really wide merge "lanes".

I suppose when actually entering the highway, the unmarked option is better, but when already on the highway (which is by far the more common situation when traveling more than a few exits), when traveling in the right lane, the car will move over to the right to center itself in the "wide" lane., and then come back to the travel lane. Not only is this possibly confusing to other vehicles around you, but if the on-ramp happens to be on a right hand curve, jumping back into the travel lane can actually be quite rapid and very concerning to other traffic around you.

That said, the latest version of FSD (shared highway and surface street stack) has actually fixed that behavior.

 

[zoomer0056]

Normally yes, but in this case it sounds more like the car was seriously confused about the lanes. This happens as well where it suddenly and unexpectedly pops into an alternate lane at a lane fork situation.

Are you saying the car did not signal? I think the OP should weigh in on this question.

 

[kabin]

There is such a setting in the latest FSD software (Minimize Lane Changes).

If you had that on, and it still tried to make the lane change, it was apparently confused about which lane was the "through" lane even though it was probably obvious to you. This is a continuing problem with FSD...it even makes these kinds of weird intentions to change lanes in my neighborhood roads with left turn only lanes (it very briefly puts on the blinker like it wants to change lanes, but then figures it out and cancels).

This is an excellent opportunity for you to be able to 'teach' the system by using the feedback capability when you disengaged. One would hope that they would use your snippet as part of their training so they can learn that and similar lane splits.

From my experience Minimize Lane Changes gives little relief even in Chill mode. As many have said, straight clear roadway with no traffic and an immediate upcoming turn needed but still FSD does a needless willy nilly lane change out of the proper lane. It feels like a bug possibly related to the random errant turn signal bug.

 

[RTPEV]

Are you saying the car did not signal? I think the OP should weigh in on this question.

The OP would have to address the question about whether it signaled in their case.

I was responding to your question of whether or not FSD normally signals before a lane change:

"Does the car signal before a lane change?"

The answer to that question is "normally, yes" (with the exceptions given in my post)

 

[cbdream99]

I start seeing this behavior since 11.4.2 for FSDb, on roads that have a turn lane opening, it may signals and then turn into it for no reason, often too late to cancel the turn, sometimes it quickly swerve back to the original lane, sometimes it doesn't, either way it shouldn't do that, no matter what mode or minimize turn is enabled. This reminds me couple of years ago NOA before FSDb - it always mistaken a turn lane as the route at certain location ... Just be careful with this behavior, if it turns into a very short turn lane at 55mph, you only have couple of seconds to correct.

 

[zoomer0056]

The OP would have to address the question about whether it signaled in their case.

I was responding to your question of whether or not FSD normally signals before a lane change:



The answer to that question is "normally, yes" (with the exceptions given in my post)

Yes, hopefully the OP chimes in on this question.

 

[Docpaulo]

I don't have FSD. Does the car signal before a lane change? This would be a clear time to over ride, before the change occurs.

It does, but sometimes seconds before it switches lanes

 

[Docpaulo]

Normally yes, but in this case it sounds more like the car was seriously confused about the lanes. This happens as well where it suddenly and unexpectedly pops into an alternate lane at a lane fork situation.

Correct.. it thought the 3rd lane was the fast lanei

Yes, hopefully the OP chimes in on this question.

It signaled a lane change a second before moving... i didn't realize it was signaling to the left, i thought it was going right as i wasn't looking at the screen and just heard the signal and the right was the only legit travel lane

There is such a setting in the latest FSD software (Minimize Lane Changes).

Minimize lane changes doesn't really work well.. it still makes a lot of lane changes.. I wish it totally disabled all lane changes except exits... not minimize.. just not do it at all until i signal it to do so.

i know because i have it on most of the time during this road trip.. I'm 2000 miles into it.. and I've been driving in chill mode all the way..


Overall its a godsend, but its definitely not there yet...

Another time it also tried to use a center turn lane as a travel lane (the kind between 2 yellow lines) and i wasn't turning for another mile.. i had to intervene..

The logic really needs to avoid shifting lanes unnecessarily

 

[RTPEV]

Correct.. it thought the 3rd lane was the fast lanei

It signaled a lane change a second before moving... i didn't realize it was signaling to the left, i thought it was going right as i wasn't looking at the screen and just heard the signal and the right was the only legit travel lane

Minimize lane changes doesn't really work well.. it still makes a lot of lane changes.. I wish it totally disabled all lane changes except exits... not minimize.. just not do it at all until i signal it to do so.

Hmmmm....in my experience, Minimize lane changes works exactly how I'd like it. It never changes lanes unless it thinks it needs to to follow the route, at least in cases where the lanes are obvious (i.e. most highways and other roads where lanes are not added or dropped). And while I didn't have a 2000 mile trip to test this one, I did use it on an 800 mile trip, and have used it fairly extensively locally as well. You are aware that each time you start a new drive it resets the setting to off, right? (which is incredibly annoying) (I have to believe that after 2000 miles you've determined this already, but just in case...)

The logic really needs to avoid shifting lanes unnecessarily

I agree. The biggest culprit I see is that it doesn't like the right lane. I understand why it tries to avoid the right lane, but most of the time, at the speeds I travel at and the roads I travel on, the right lane is most appropriate. Yes, I have to anticipate merging traffic (which is the problem because a human would see the car before FSD does and move quickly out of the slow lane, whereas FSD first warns you of the turn, and then takes forever to make the lane change, only to chicken out when it's only halfway through the lane change and the merging car puts a whisker of the tire into your lane--resulting in it slamming on the brakes and tucking in behind the merging car. Very faulty logic. The latest software update I received made some kind of mention of merging traffic-hopefully they've improved that.

 

[Docpaulo]

Hmmmm....in my experience, Minimize lane changes works exactly how I'd like it. It never changes lanes unless it thinks it needs to to follow the route, at least in cases where the lanes are obvious (i.e. most highways and other roads where lanes are not added or dropped). And while I didn't have a 2000 mile trip to test this one, I did use it on an 800 mile trip, and have used it fairly extensively locally as well. You are aware that each time you start a new drive it resets the setting to off, right? (which is incredibly annoying) (I have to believe that after 2000 miles you've determined this already, but just in case...)


I agree. The biggest culprit I see is that it doesn't like the right lane. I understand why it tries to avoid the right lane, but most of the time, at the speeds I travel at and the roads I travel on, the right lane is most appropriate. Yes, I have to anticipate merging traffic (which is the problem because a human would see the car before FSD does and move quickly out of the slow lane, whereas FSD first warns you of the turn, and then takes forever to make the lane change, only to chicken out when it's only halfway through the lane change and the merging car puts a whisker of the tire into your lane--resulting in it slamming on the brakes and tucking in behind the merging car. Very faulty logic. The latest software update I received made some kind of mention of merging traffic-hopefully they've improved that.

Yup, i use the stalk and click minimal most drives into a freeway. This incident I did not have it on as it was a local highway... and i don't bother doing the switch until i enter the ramp..

I've read somewhere that previous versions acted like you say and the 11.4.4 is acting weird.. dunno.. i wish it didn't do lane changes on its own...

During the last drive 30 mins ago, it made another unnecessary lane change, on a freeway with no exits nearby.. went from left lane to right lane where a truck was and it had to slow down from 80 to 70 mph when it came up behind the truck... I'm certain the ai saw the lane was occupied but still decided the right lane was the best lane to go into... Maybe it didn't see it when it initiated the lane change but the brakes kicked in probably 5 seconds after lane change was completed.. 100% minimal lane changes was clicked as i made sure just to test the system again...

Then another time it kept going into the left lane on an empty freeway and i kept moving it back to the right by canceling the lane change...

There's a hidden algorithm that appears faulty... can't make sense of it..

That said, now that I'm more aware of the bugs, i just anticipate it... just hoping they eventually get it right