Truck Lust is a real thing...

[SomeJoe7777]

By focusing on the blue lines you can see clearly that in each instance the car is moved to the far left side of the lane just before Autopilot is turned on. You can see this clearly if you stop the video just after Autopilot is engaged, at 0:17, 0:28 and 0:43.

After AP is engaged when the car is on the far left side of the lane, Autopilot moved the vehicle to the right, toward the center of the lane. Autopilot was then disengaged, presumably because it "felt" like the car was being moved to the right, which it was, because the vehicle was on the left side of the lane, not the center.

I understand your explanation, in that starting to the left of center and then moving towards the right gives the optical illusion that the car is right of center and closer to the truck than it really is.

But analysis of the video does not bear this explanation out. These frame captures are from the 3rd incident:

This is frame 30:01, just after AP was engaged for the 3rd time. The vehicle appears to be perfectly centered in the lane: The vanishing point of the highway lane markers almost directly intersects with a vertical line an exact midpoint between the defroster vents. The only tiny amount of error is due to the BlackVue's lens being on the left-hand side of the camera body, which is directly centered behind my review mirror. AP also indicates that the vehicle is nearly perfectly centered, or perhaps slightly to the left. We can attribute this to the fact that I engaged AP here while the vehicle was slightly left of center, and AP began aligning the vehicle with the lane by moving it to the right. The AP display is lagging behind the actual vehicle position by a small fraction of a second.

I saved these reference markers in a different layer in Photoshop, then loaded frame 32:03 underneath:

This is after the abrupt rightward move by AP, and my subsequent arrest of that movement using the steering wheel, forcing AP disengagement. This was the maximum rightward displacement that the vehicle achieved during the movement.

The width of the yellow 12' lane marker reference line is 536 pixels. The dark blue displacement line is 111 pixels. By ratio, this is 30" (2' 6") of rightward displacement from what should have been a nearly perfectly centered position.

With a 12' (144") lane, and a vehicle width excluding mirrors of 77.3", this should give me 33.5" of margin on both sides if my vehicle is perfectly centered. WIth 30" of rightward displacement, this AP movement placed the edge of my right tires only 3" from the center lane stripes, and that was after I arrested the rightward movement. If you count the vehicle width with mirrors of 86.2", the right tip of the right mirror was 1.1" into the right lane of the highway.

Notice that the AP display still shows that the vehicle is centered (although with very low confidence, as the gray and black lane markers are gone, only the gray road surface is shown). Clearly, the AP system has been fooled by some combination of factors.

This was not an optical illusion on my part, nor a small but normal variation in AP's lane holding. This was a distinct, abrupt, definitive, and significant move almost completely out of the intended lane.

What we don't know is what combination of factors led to this behavior. The truck's existence is a factor, because I did not have this behavior happen either before of after this incident on this same highway, with the same lighting conditions.

I believe the following factors could have contributed in some fashion to this behavior:

• The truck's unusual shape which may be an edge case for the radar.
• Poor lighting conditions with the sun facing the AP camera.
• Heavy shadows from the truck obscuring the center lane markers.
• An odd shadow shape from the truck due to his cargo (note that the right red lane marker in frame 32:03 intersects the exact spot where the sun/shadow form a contrast line that is the same length and direction as a lane marker would be).

Should we call this "truck lust"? I don't know. While the term is amusing, it doesn't fully describe what is going on, since we don't even know what that is. What I can conclude here is that AP began to steer the vehicle out of the lane and place it on a collision course with the truck, requiring my intervention, and I can conclude that the truck's existence was a factor (but perhaps not the only factor) that precipitated AP's error.

 

[Todd Burch]

I can guarantee it was not a calibration or alignment issue that I felt. As I noted in my post it has only happened once and I have passed a large number of trucks with AP engaged. The car pulled towards the trucks trailer axle and I needed to take over when I experienced the phenomenon.

I think you misunderstand me. I didn't say that you didn't feel the car pull toward the truck. I didn't even say that your car didn't pull toward the truck. What I was saying was that your car was pulling toward the truck perhaps due to a calibration or manufacturing assembly tolerance issue.

My point is that if this happens to EVERYONE, then it's an issue in the software. If it only happens to some (as it seems to be the case), then something else is involved that doesn't apply to all of us.

 

[Max*]

I love how there's plenty of people in this thread acting like "well, it didn't happen to me in xx,xxx miles, so it must not be real"


I just did a 500 mile trip this weekend, and it happened once. It's fairly rare (I encountered a LOT of trucks in those 500 miles), but only once did it pull towards one making me turn AP off.

I'm on the latest firmware (something .22 or whatever) that's supposed to fix the truck lust.


And I don't buy it's a calibration thing, because why then does it ONLY do it for trucks? Not cars. It works perfectly otherwise, but once it passes a truck (and not even every time) it pulls towards it.

 

[Dax279]

I think you misunderstand me. I didn't say that you didn't feel the car pull toward the truck. I didn't even say that your car didn't pull toward the truck. What I was saying was that your car was pulling toward the truck perhaps due to a calibration or manufacturing assembly tolerance issue.

My point is that if this happens to EVERYONE, then it's an issue in the software. If it only happens to some (as it seems to be the case), then something else is involved that doesn't apply to all of us.

Sorry I did not misunderstand, I just can't believe you are so quick to rule out the possibility that in a certain scenario this happens.

I think Max sums things up pretty nicely.

 

[BerTX]

I love how there's plenty of people in this thread acting like "well, it didn't happen to me in xx,xxx miles, so it must not be real"


I just did a 500 mile trip this weekend, and it happened once. It's fairly rare (I encountered a LOT of trucks in those 500 miles), but only once did it pull towards one making me turn AP off.

I'm on the latest firmware (something .22 or whatever) that's supposed to fix the truck lust.


And I don't buy it's a calibration thing, because why then does it ONLY do it for trucks? Not cars. It works perfectly otherwise, but once it passes a truck (and not even every time) it pulls towards it.

I am going to respond because I did say it might not be real. But that was because I apparently had a misconception of what was happening to people until I saw the video. Since it has never happened to me, I didn't understand what was being described.

It definitely is real. Apparently, it is not universal. I am not a good metric, though, because I have NOT used AP for a bazillion miles.

 

[RDoc]

I've occasionally noticed Truck Lust, and once it was scary enough that I yanked the wheel back. The car display shows the car drifting to the right as well.

If it's calibration, it's not an uncommon one.

 

[TrafficEng]

I understand your explanation, in that starting to the left of center and then moving towards the right gives the optical illusion that the car is right of center and closer to the truck than it really is.

But analysis of the video does not bear this explanation out. These frame captures are from the 3rd incident:

View attachment 210748

This is frame 30:01, just after AP was engaged for the 3rd time. The vehicle appears to be perfectly centered in the lane: The vanishing point of the highway lane markers almost directly intersects with a vertical line an exact midpoint between the defroster vents. The only tiny amount of error is due to the BlackVue's lens being on the left-hand side of the camera body, which is directly centered behind my review mirror. AP also indicates that the vehicle is nearly perfectly centered, or perhaps slightly to the left. We can attribute this to the fact that I engaged AP here while the vehicle was slightly left of center, and AP began aligning the vehicle with the lane by moving it to the right. The AP display is lagging behind the actual vehicle position by a small fraction of a second.

I saved these reference markers in a different layer in Photoshop, then loaded frame 32:03 underneath:

View attachment 210749

This is after the abrupt rightward move by AP, and my subsequent arrest of that movement using the steering wheel, forcing AP disengagement. This was the maximum rightward displacement that the vehicle achieved during the movement.

The width of the yellow 12' lane marker reference line is 536 pixels. The dark blue displacement line is 111 pixels. By ratio, this is 30" (2' 6") of rightward displacement from what should have been a nearly perfectly centered position.

With a 12' (144") lane, and a vehicle width excluding mirrors of 77.3", this should give me 33.5" of margin on both sides if my vehicle is perfectly centered. WIth 30" of rightward displacement, this AP movement placed the edge of my right tires only 3" from the center lane stripes, and that was after I arrested the rightward movement. If you count the vehicle width with mirrors of 86.2", the right tip of the right mirror was 1.1" into the right lane of the highway.

Notice that the AP display still shows that the vehicle is centered (although with very low confidence, as the gray and black lane markers are gone, only the gray road surface is shown). Clearly, the AP system has been fooled by some combination of factors.

This was not an optical illusion on my part, nor a small but normal variation in AP's lane holding. This was a distinct, abrupt, definitive, and significant move almost completely out of the intended lane.

What we don't know is what combination of factors led to this behavior. The truck's existence is a factor, because I did not have this behavior happen either before of after this incident on this same highway, with the same lighting conditions.

I believe the following factors could have contributed in some fashion to this behavior:

• The truck's unusual shape which may be an edge case for the radar.
• Poor lighting conditions with the sun facing the AP camera.
• Heavy shadows from the truck obscuring the center lane markers.
• An odd shadow shape from the truck due to his cargo (note that the right red lane marker in frame 32:03 intersects the exact spot where the sun/shadow form a contrast line that is the same length and direction as a lane marker would be).

Should we call this "truck lust"? I don't know. While the term is amusing, it doesn't fully describe what is going on, since we don't even know what that is. What I can conclude here is that AP began to steer the vehicle out of the lane and place it on a collision course with the truck, requiring my intervention, and I can conclude that the truck's existence was a factor (but perhaps not the only factor) that precipitated AP's error.

This is a great analysis! Thanks so much for putting in this effort, and your detailed work on the subject. I can add the following based on my own drive down I 85/I 285 in Atlanta today. I had someone else drive, so I could watch in detail as we passed plenty of trucks on this interstate. I can report that we had one incident where we had to take over steering (the vehicle lost the lines, and drifted towards the truck) and that was also passing a truck on the left hand side, with a strong shadow on the left hand side of the truck. So the hypothesis that the shadow is confusing the lane detection system is still alive. We passed many other trucks, and noticed each time that the shadow was not obscuring the lane line, and AP had no trouble at all with steering.

So can we ask the group to test this hypothesis, that a strong shadow crossing the lane line from a long vehicle, on the side that we are passing on, results sometimes in the lane detection algorithm becoming confused?

Look forward to hearing the results!

 

[S4WRXTTCS]

Maybe this is something that indicates that it is more "perceived" than real? Surely someone would have posted a video if it was possible to get a video. Or, maybe it really isn't demonstrative unless there is a shot of the dash/steering wheel at the same time and the dashcam wouldn't show that.

There is a Youtube video supposedly showing truck lust, but that is not what I have experienced. It shows a dip towards the truck as the car is coming up behind -- mine is right as you are next to the truck.

Mine is right next to the truck as well. More specifically just as your next to the trailer section of it.

Personally I'm convinced it's loses the lines towards the truck so it starts to move closer to the truck. You don't even really need a truck to see how much the car moves over when the line is missing. You just find an area of the road where it's missing the line on the right, and see how much it moves over in the absence of the line.

The truck is a hard variable to control because it entirely depends on whether the ultrasonic sensors can see it. If they can you won't have truck lust with that truck. If they can't you might or might not have truck lust depending on if the lines drop out.

What I find most troubling about truck lust is how infrequent it happens. A good portion of I5 between Seattle, and Portland for all practical purposes is completely owned by semi trucks. I'm not sure why because there is a rail line right next to it. But, it's the way it is. So I'll pass hundreds of hundreds of them when I travel down at night.

In that 3.5 hour drive (not including charging) I might or might not encounter it. If I do it's only going to be a single time.

It happens really fast, and the recovery isn't subtle. Where you could potentially overcorrect. In fact at one point I switched steering sensitivity off of sport just because that makes it less abrupt. Now days don't even use AP when there are tons of trucks that I'll be passing. It's just not worth the anxiety.

I'm sure part of why it's so scary is because it's a truck. Not only would I normally move to the left of the lane to give more room, but I wouldn't move towards the truck. It's likely a subtle movement that's amplified by a humans aversion to big scary things.

I'm not sure a dash cam would pick it up. If I was really dead set on documenting it I would use lidar-lite sensor modules to document the distance to an object on the right.

I'm pretty convinced that Tesla did as much as they can with AP1 to try to mitigate it. Like it no longer does the dive for exits like it used to do. It also seems less "searchy" than it used to.

I don't expect an AP2 car to have truck lust (long term) because it has side cameras and it has better ultrasonic sensor. It has multiple ways to see it.

What I'm most curious about is how AP1 cars in Australia handle truck lust when next to the MASSIVELY long train trucks they have down there. So I'll probably post a question for them in their area of this site. They might have the side curtain things that would give the ultrasonic sensors something to bounce off of.

 

[Electric700]

I understand your explanation, in that starting to the left of center and then moving towards the right gives the optical illusion that the car is right of center and closer to the truck than it really is.

But analysis of the video does not bear this explanation out. These frame captures are from the 3rd incident:

View attachment 210748

This is frame 30:01, just after AP was engaged for the 3rd time. The vehicle appears to be perfectly centered in the lane: The vanishing point of the highway lane markers almost directly intersects with a vertical line an exact midpoint between the defroster vents. The only tiny amount of error is due to the BlackVue's lens being on the left-hand side of the camera body, which is directly centered behind my review mirror. AP also indicates that the vehicle is nearly perfectly centered, or perhaps slightly to the left. We can attribute this to the fact that I engaged AP here while the vehicle was slightly left of center, and AP began aligning the vehicle with the lane by moving it to the right. The AP display is lagging behind the actual vehicle position by a small fraction of a second.

I saved these reference markers in a different layer in Photoshop, then loaded frame 32:03 underneath:

View attachment 210749

This is after the abrupt rightward move by AP, and my subsequent arrest of that movement using the steering wheel, forcing AP disengagement. This was the maximum rightward displacement that the vehicle achieved during the movement.

The width of the yellow 12' lane marker reference line is 536 pixels. The dark blue displacement line is 111 pixels. By ratio, this is 30" (2' 6") of rightward displacement from what should have been a nearly perfectly centered position.

With a 12' (144") lane, and a vehicle width excluding mirrors of 77.3", this should give me 33.5" of margin on both sides if my vehicle is perfectly centered. WIth 30" of rightward displacement, this AP movement placed the edge of my right tires only 3" from the center lane stripes, and that was after I arrested the rightward movement. If you count the vehicle width with mirrors of 86.2", the right tip of the right mirror was 1.1" into the right lane of the highway.

Notice that the AP display still shows that the vehicle is centered (although with very low confidence, as the gray and black lane markers are gone, only the gray road surface is shown). Clearly, the AP system has been fooled by some combination of factors.

This was not an optical illusion on my part, nor a small but normal variation in AP's lane holding. This was a distinct, abrupt, definitive, and significant move almost completely out of the intended lane.

What we don't know is what combination of factors led to this behavior. The truck's existence is a factor, because I did not have this behavior happen either before of after this incident on this same highway, with the same lighting conditions.

I believe the following factors could have contributed in some fashion to this behavior:

• The truck's unusual shape which may be an edge case for the radar.
• Poor lighting conditions with the sun facing the AP camera.
• Heavy shadows from the truck obscuring the center lane markers.
• An odd shadow shape from the truck due to his cargo (note that the right red lane marker in frame 32:03 intersects the exact spot where the sun/shadow form a contrast line that is the same length and direction as a lane marker would be).

Should we call this "truck lust"? I don't know. While the term is amusing, it doesn't fully describe what is going on, since we don't even know what that is. What I can conclude here is that AP began to steer the vehicle out of the lane and place it on a collision course with the truck, requiring my intervention, and I can conclude that the truck's existence was a factor (but perhaps not the only factor) that precipitated AP's error.

Very thorough analysis. Consider contacting [email protected] about this. It really should be fixed (I think this would mean a software update is needed).

 

[Electroman]

Though it has not happened to me (yet) I can see how this can happen, because I have seen this gentle shift for reasons other than trucks, mostly when it loses sight of one lane then it moves in that direction.

 

[Tech_Guy]

I
But analysis of the video does not bear this explanation out. These frame captures are from the 3rd incident:

View attachment 210749

To me looking at the dash in the second shot it no longer shows the lane lines like it can't see them any more.

 

[Jevi_S85D]

I've been involved in computer image analysis programming for 20 years. I have a pretty good idea of how lane markers are detected. The algorithm probably searches for lines that qualify the best by their contrast, angle, length and possibly width.
My theory is that the length of the truck is the reason for truck lust. The algorithm simply "thinks" that the side of the truck represents a lane marker. Obviously this would be a too simplistic approach. The computation should include the radar detection of the truck and reject lines that could be the result of a long truck. To integrate camera and radar information has been one of the challenges for Tesla as two subsuppliers (Mobileye and Bosch) are involved but I'm sure we will see improvements in the future just like v8.0 was a big step forward.

 

[SomeJoe7777]

To me looking at the dash in the second shot it no longer shows the lane lines like it can't see them any more.

Yes, that's correct, it definitely lost the lane lines, and that in turn probably precipitated the abrupt move. As we're now starting to ask, that begs the question of: what is different about trucks that causes difficulties for the lane detection algorithm? Length? Large shadows? Long lines along the truck sides that mimic the contrast of a lane marker?

Jevi above postulates some of these exact questions, and I indeed believe he may be onto something with the image analysis of the side of a truck being confused for a lane stripe. Especially if the truck's color (a lot of times white) contrasts with the area below the truck (dark due to shadow).

 

[TrafficEng]

I've been involved in computer image analysis programming for 20 years. I have a pretty good idea of how lane markers are detected. The algorithm probably searches for lines that qualify the best by their contrast, angle, length and possibly width.
My theory is that the length of the truck is the reason for truck lust. The algorithm simply "thinks" that the side of the truck represents a lane marker. Obviously this would be a too simplistic approach. The computation should include the radar detection of the truck and reject lines that could be the result of a long truck. To integrate camera and radar information has been one of the challenges for Tesla as two subsuppliers (Mobileye and Bosch) are involved but I'm sure we will see improvements in the future just like v8.0 was a big step forward.

it seems likely that the object detection, including lane line detection, is done in the Mobileye processor. Hence improving the performance around trucks may not be under Tesla control, but tighter integration by Tesla with the radar may be possible, to lesson the effect. And in support of the 'white truck theory' our incident yesterday was with a white truck, and my wife commented '... It seems to only happen with white trucks...'

Given that Truck Lust is just a term we are using to describe the event where, when passing a truck on AP, the vehicle suddenly veers towards the truck, the crowd-sourced experience in this thread shows:
1) Truck Lust is a real, but comparatively infrequent event.
2) The direct cause is that the lane line detection processing loses the line next to the truck, and the vehicle then moves towards the lost line, and hence the truck.
3) there are two further details to be tested:
a) it seems to occur with white trucks?
b) it seems to occur when the shadow from the truck falls on the lane line that then gets lost?

We are close to being able to make this a repeatable/predictable bug, which would make it easier for Tesla to fix.

 

[TrafficEng]

What I find most troubling about truck lust is how infrequent it happens. A good portion of I5 between Seattle, and Portland for all practical purposes is completely owned by semi trucks. I'm not sure why because there is a rail line right next to it. But, it's the way it is. So I'll pass hundreds of hundreds of them when I travel down at night.

Can you check next time you experience this
1) Was there a shadow between you and the truck? (one theory is that Truck Lust does not happen at night - it needs a shadow)
2) Was the truck white? (one theory is that it happens with white trucks)

 

[Nietschy]

1) Was there a shadow between you and the truck? (one theory is that Truck Lust does not happen at night - it needs a shadow)
2) Was the truck white? (one theory is that it happens with white trucks)

This is a no.
I drive mostly at night on long trips and it happens.
My trucks were not white...

 

[S4WRXTTCS]

Can you check next time you experience this
1) Was there a shadow between you and the truck? (one theory is that Truck Lust does not happen at night - it needs a shadow)
2) Was the truck white? (one theory is that it happens with white trucks)

In my experiences it tends to happen at night. It could either be that nighttime makes the camera less effective, or it could be the sheer number of semi-trucks on the road (UPS, FedEx, other carriers, etc).

It's been awhile since I last experienced a truck lust incident so I don't have any recent incident to recall. I should make a website counter for "It's been X number of days since your last truck lust incident".

 

[wdolson]

I've been involved in computer image analysis programming for 20 years. I have a pretty good idea of how lane markers are detected. The algorithm probably searches for lines that qualify the best by their contrast, angle, length and possibly width.
My theory is that the length of the truck is the reason for truck lust. The algorithm simply "thinks" that the side of the truck represents a lane marker. Obviously this would be a too simplistic approach. The computation should include the radar detection of the truck and reject lines that could be the result of a long truck. To integrate camera and radar information has been one of the challenges for Tesla as two subsuppliers (Mobileye and Bosch) are involved but I'm sure we will see improvements in the future just like v8.0 was a big step forward.

This is probably why I've seen the car swerve towards bridge railings too. The algorithm that is figuring out where the lane markers are needs to stay focused on information from ground level. Anything above a couple of inches needs to be rejected for the purpose of figuring out the lane markers.

 

[EVie'sDad]

as far as 2.52.22 I believe truck lust has not yet been resolved. I still travel uncomfortably close to big rigs while driving down the freeway.

I think, rather than hypothesize as to the root cause of the issue, that Tesla can take pre-emptive action, that when in and around the vicinity of large trucks, whether they be flatbeds or tall trailers, that tesla should see AND react in a manner to hug the lane opposite, just as it does once the ultrasonic senors finally pick up the trucks wheel-base.

Use the cameras to estimate truck size and compensate accordingly.

This would mitigate the truck lust being experienced, and have the Tesla give larger vehicles a slightly wider berth to compensate for the fact that most of those trucks barely fit in a normal highway lane these days. Lets see if Elon is paying attention, and takes the proper course of action to help us be safer on the highways.

 

[Boatguy]

My theory is that the length of the truck is the reason for truck lust. The algorithm simply "thinks" that the side of the truck represents a lane marker.

I can understand your reasoning, but since the instrument display shows the truck as I'm approaching it, I believe it has correctly identified the truck as a truck and I'm skeptical that the software thinks it's a lane line.

I did a 2,000mi road trip last week and have developed three theories about "truck lust".

1) Steering manually, I tend to move to the left side of the lane when passing a truck. Staying in the center of the lane, as auto-steer tries to do, feels like I'm too close to the truck.

2) I believe the truck creates a low pressure area along its side which tends to suck in the car. This is particularly true of trucks with faring under the bed to improve their aerodynamics.

3) I believe #2 is very pronounced if there is a prevailing wind on the opposite side of the truck. For example, if I was driving north, with a wind from the east (BTW, Teslawinds.com is very cool!), and I am passing on the west side, the prevailing crosswind suddenly stops as I am shielded by the truck, causing the steering bias that was correcting for the crosswind to bring the car closer to the truck.

FWIW, autosteer always seems to bring the car back to the middle of the lane, but there is an anxious moment none the less.