The Best Objective Study To Date on Truck Lust

[BerTX]

Lighten up. Or at least understand what statistics can and can not prove. I didn't say my test proves anything. In fact, as I mentioned in my brief statistical discourse with Max, it's impossible to prove any null hypothesis. My interpretation is my own. My results in this small, uncontrolled, limited study, confined to a few hundred square miles of testing range, are my results. Your mileage may vary.

And of course, you have the chance to prove that it does exist. Do a similar test as mine under your driving conditions. I'm not talking about posting one or two videos that show this behavior. Use multiple passes by trucks, and do a two-tailed t-test (or, if you believe there's a better statistical test to use, then go for it) to compare mean distance from the highway divider lines before, during, and after a truck passing event. Report back the p-value. I'm totally open to a better study with more data under a wider variety of speeds and driving conditions.

"The Definitive Word on Truck Lust"
or
"I didn't say my test proves anything"

Sounds a little contradictory, you have to admit.

 

[tel]

"The Definitive Word on Truck Lust"
or
"I didn't say my test proves anything"

Sounds a little contradictory, you have to admit.

Come on. Tongue-in-cheek. I'm good at statistics, but not so good at figuring out how to change titles on a forum thread. You show me how to do that, and I'll update the title to something that makes you feel better.

 

[BerTX]

Come on. Tongue-in-cheek. I'm good at statistics, but not so good at figuring out how to change titles on a forum thread. You show me how to do that, and I'll update the title to something that makes you feel better.

I don't feel bad. I just don't want to get into yet another discussion of truck lust, where people are arguing over two completely different things. One says "the thing I think is truck lust doesn't exist" and the other says "the thing I think is truck lust does exist" and they are talking about different things. Pointless.

And you can't edit titles unless you are a mod.

 

[SomeJoe7777]

I didn't say my test proves anything.

Yet, you use words like "definitive" and "illusion". If you really mean that your test doesn't prove anything, then there is no conclusion you can draw, correct?

Meanwhile, can you explain this:

Analysis here:

Truck Lust is a real thing...


Vehicle movement is definitive, deliberate, and significant.

 

[tel]

Yet, you use words like "definitive" and "illusion". If you really mean that your test doesn't prove anything, then there is no conclusion you can draw, correct?

Meanwhile, can you explain this:

Analysis here:

Truck Lust is a real thing...


Vehicle movement is definitive, deliberate, and significant.

Really? You want to claim proof based on one video? Or five? Or ten? That's not how statistics work. I can post videos of airplanes flying over my car and therefore "prove" that my tesla attracts flying vehicles. That's a classic example of selection bias.

Alternatively, I could do a more thoughtful analysis, and count the number of times that doesn't happen, and the number of times it happens not associated with my car, and show that there's little statistical relation to my tesla and overhead airplanes.

Your vehicle movement may happen from time to time. As I commented earlier in this thread regarding this exact video, how can we be sure that this isn't overcorrection from the initial swerving AWAY from the truck that's evident.

How do we know that this swerving doesn't happen in situations NOT associated with trucks (for example shadows, exits, poor lane markings, poor weather conditions, etc.) How often do you pass trucks and NOT witness this behavior? I've shown you my data taking these variables into account to the degree that I'm able. You show me yours.

And my naming this thread as "definitive" is mostly a joke. I absolutely concede the serious limitations of what I've looked at. If I could rename it "The Most-Objective-Evidence-So-Far-But-Hotly-Contested Data on Truck Lust," I would.

 

[Max*]

That's a classic example of selection bias.

That and it's proving that correlation does not imply causation.

I've shown you my data taking all these variables into account.

To be fair, you didn't actually show any data...

 

[Max*]

Really? You want to claim proof based on one video? Or five? Or ten? That's not how statistics work. I can post videos of airplanes flying over my car and therefore "prove" that my tesla attracts flying vehicles. That's a classic example of selection bias.

Alternatively, I could do a more thoughtful analysis, and count the number of times that doesn't happen, and the number of times it happens not associated with my car, and show that there's little statistical relation to my tesla and overhead airplanes.

Your vehicle movement may happen from time to time. As I commented earlier in this thread regarding this exact video, how can we be sure that this isn't overcorrection from the initial swerving AWAY from the truck that's evident.

How do we know that this swerving doesn't happen in situations NOT associated with trucks (for example shadows, exits, poor lane markings, poor weather conditions, etc.) How often do you pass trucks and NOT witness this behavior? I've shown you my data taking these variables into account to the degree that I'm able. You show me yours.

And my naming this thread as "definitive" is mostly a joke. I absolutely concede the serious limitations of what I've looked at. If I could rename it "The Most-Objective-Evidence-So-Far-But-Hotly-Contested Data on Truck Lust," I would.

I think @SomeJoe7777 has a valid point.

He's shown Tesla veering towards trucks, multiple times.
Anyone who's driven with AP can show that Tesla's don't usually veer to other smaller vehicles.

Truck lust could be something as simple as the car correcting to go away from the truck and then OVERcorrecting to go towards the truck too much/too fast. That's still truck lust though. That would still be valid. Because the end result is that your Tesla is approaching an 18 wheeler at a rate of speed faster than a human would feel comfortable with.

The only way the above point would not be valid, if that correction were to happen in the middle of the road with no cars around. Then you can argue it can happen at any point in time, and it's not specific to trucks. The problem, is that it doesn't happen in other places without explanation. The key words in the last sentence are "without explanation". Poor lane markings and the car swerves, is an explanation. A curve in the road and the car swerves is an explanation. Truck lust has no proven explanation (yet?)

 

[tel]

That and it's proving that correlation does not imply causation.

To be fair, you didn't actually show any data...

You're right. This isn't a peer-reviewed journal, so I omitted the raw data and statistics which I would otherwise have included in a published work. Perhaps I should have said "I have data. Please get some of your own."

 

[Max*]

You're right. This isn't a peer-reviewed journal, so I omitted the raw data and statistics which I would otherwise have included in a published work. Perhaps I should have said "I have data. Please get some of your own."

He has his own data

 

[SomeJoe7777]

You want to claim proof based on one video?

Nowhere in this thread or the other I referenced have I claimed that any of my evidence is proof of anything. That is your inference, not mine.

I clearly state at the bottom of my analysis post in the other thread exactly what I can deduce, which 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.

I don't even claim that what I experienced should be referred to as "truck lust", if that is even a definable term.

What I can tell you from my evidence and what others have presented is that the phenomenon is rare. Various posts have claimed that AP1 has exhibited something of this nature since it was deployed in October of 2015. In 16 months since then, we have what -- maybe 4 or 5 videos that clearly show it? And how many hundreds of thousands or millions of trucks passed? 40 or 50 truck passing tests is meaningless to state anything, especially "it's most likely an illusion". Sorry, that's simply not correct.

Until someone can explain otherwise, I and others have pieces of video that show an unexplained AP behavior that involves trucks. More research is definitely needed.

 

[tel]

More research is definitely needed.

By far the smartest thing said on any of the posts so far!

 

[brucet999]

I think @SomeJoe7777 has a valid point.

He's shown Tesla veering towards trucks, multiple times.
Anyone who's driven with AP can show that Tesla's don't usually veer to other smaller vehicles.

Truck lust could be something as simple as the car correcting to go away from the truck and then OVERcorrecting to go towards the truck too much/too fast. That's still truck lust though. That would still be valid. Because the end result is that your Tesla is approaching an 18 wheeler at a rate of speed faster than a human would feel comfortable with.

My experiences with truck lust did not involve any sort of correction. Straight and steady driving until close to overtaking a truck trailer then swerve sharply toward its back corner.

 

[TSLA Pilot]

I'll add my video to the list of evidence of its existence. If you watch the second time through (after the short black break) I've sped it up to help see the movement in spite of the wide angle lens making everything look rather mild. Also, the last clip shows the car move away from the truck, but departing the lane.. so I included that for fun.

And I should add, that I started pulling these videos down from the camera after about 5-6 others. So it's just 4 of the MANY times that this happened on my road trip from TX to CA.

Have to concur with the 100% existance of Truck Lust.

On our first long distance drive with our P90D (2015), with the March, 2016, software set, AP was a near killer.

1. MS sees truck (or sometimes a large SUV) in right lane, as indicated in display.

2. MS is in, and remains in, the center of its own lane.

3. MS begins to sense truck with ultrasonic sensor displays in amber.

4. MS begins to slide to the right, away from the center of its own lane, AND THE CAR KNOWS THIS as this shift to the far right of our lane is clearly indicated in the IP display.

5. Ultrasonic displays change to RED due to closeness of truck.

6. I urgently grap the wheel to jerk the car away from truck.

Occurred around 1 to 3% of the time I'd pass a truck on perfect days with clear lane markings on wide, Interstate freeways.

Advised SC and they brushed it off without even a test drive.

Left wary of using AP during highway driving with lots of trucks, and also disappointed with Tesla's lack of concern about this safety defect, but it's Tesla--what can you do?

Hoping they'll eventually figure out what the heck is going on, and that no one will get hurt, but I suspect all efforts are now focused on AP2 . . . .

 

[S4WRXTTCS]

I question whether we really should bother proving truck lust since we know all the symptoms that lead to it exists.

We know the blind/side spot monitoring system on HW1 sucks, and misses things that are there.
We know the single camera solution sometimes loses sight of the lines in certain situations.
We know the car has a tendency to search when it loses lines. Now this is a bit dependent on the AP tiles, etc. It's also a bit dependent on firmware version (with later versions it seems to haunt less).

In any case this is how I would test for truck lust where it takes into account the above.

Step 1 -> Go out in the middle of nowhere at night where no Model S/X has gone before or disable the Cell connection preventing the AP tiles from updating.
Step 2 -> Black out about 50-200ft (whatever length of a really long 2 trailer semi) of the line between two lanes. Or find a huge stretch of missing lines. This will simulate a shadow caused by the truck
Step 3 -> Drive on AP through that stretch of road.

Where you're trying to observe if the car slides over. If it does then the car is susceptible to truck lust. It doesn't require a whole lot of movement to freak a person out to the point where they overcorrect.

 

[brucet999]

I question whether we really should bother proving truck lust since we know all the symptoms that lead to it exists.

We know the blind/side spot monitoring system on HW1 sucks, and misses things that are there.
We know the single camera solution sometimes loses sight of the lines in certain situations.
We know the car has a tendency to search when it loses lines. Now this is a bit dependent on the AP tiles, etc. It's also a bit dependent on firmware version (with later versions it seems to haunt less).

In any case this is how I would test for truck lust where it takes into account the above.

Step 1 -> Go out in the middle of nowhere at night where no Model S/X has gone before or disable the Cell connection preventing the AP tiles from updating.
Step 2 -> Black out about 50-200ft (whatever length of a really long 2 trailer semi) of the line between two lanes. Or find a huge stretch of missing lines. This will simulate a shadow caused by the truck
Step 3 -> Drive on AP through that stretch of road.

Where you're trying to observe if the car slides over. If it does then the car is susceptible to truck lust. It doesn't require a whole lot of movement to freak a person out to the point where they overcorrect.

Bad news. Your shadow hypothesis doesn't work when the problem occurs westbound at 10am in Central California, where truck trailer shadows would be well outside of Tesla's lane line, and also eastbound at 2pm where shadows would be within Tesla's lane line.

 

[S4WRXTTCS]

Bad news. Your shadow hypothesis doesn't work when the problem occurs westbound at 10am in Central California, where truck trailer shadows would be well outside of Tesla's lane line, and also eastbound at 2pm where shadows would be within Tesla's lane line.

We have to think in terms of a computer vision problem. Now obviously shadows do come into play, and can hamper the ability for the computer to see the scene. But, there are things in the scene that can hamper it too.

Lets say you have two objects that you want the computer to see. Like a tennis ball and a baseball bat. Where you spend tons of your time training it on images of tennis balls, and images of baseball bats.

It does pretty well, and your thinking this is awesome.

But, then what happens if you combine the two together? Then the system doesn't see either.

So my hypothesis is the truck being so close to the line is obscuring the ability for the computer to see it as a separate entity.

The problem with truck lust is no one ever documents both the IC (to see a loss of the line), and the road.

I've had a handful of truck lust incidents and most of them have been at night. Where if there was a lighting issue I didn't see it. I came to the conclusion that it happened more at 80MPH than 70MPH. So I largely stopped used it if I intended on going 80mph.

 

[Max*]

But, then what happens if you combine the two together? Then the system doesn't see either.

Unless I'm not understanding what you're asking, that's not how a good algorithm works.

If you combine them (I assume you mean overlay), if you trained your algorithm well, you'll still detect at least one, and ideally both.

 

[S4WRXTTCS]

Unless I'm not understanding what you're asking, that's not how a good algorithm works.

If you combine them (I assume you mean overlay), if you trained your algorithm well, you'll still detect at least one, and ideally both.

In Deep learning it's highly dependent on the neural network chosen for the tasks. The example I was giving was something that I've seen happen with object detection networks.

This has caused problems for me in my own hobby project, and now I'm trying Image Segmentation to see how well that does with touching/overlapping objects.

Image Segmentation Using DIGITS 5

Now this isn't to say that MobileEye doesn't have a really good solution for detecting the line despite the semi trailer being practically on top of it. I'm simply saying that it's not as easy as a task as it might seem.

 

[Max*]

In Deep learning it's highly dependent on the neural network chosen for the tasks. The example I was giving was something that I've seen happen with object detection networks.

This has caused problems for me in my own hobby project, and now I'm trying Image Segmentation to see how well that does with touching/overlapping objects.

Image Segmentation Using DIGITS 5

Now this isn't to say that MobileEye doesn't have a really good solution for detecting the line despite the semi trailer being practically on top of it. I'm simply saying that it's not as easy as a task as it might seem.

I never said it's an easy task, but I also didn't agree with your statement that when you combine the two the system wouldn't see it.

If the model is properly trained (which is part of the link you posted), I would fully expect it to find one of the objects in an overlay. Possibly both. That's all I was saying.

 

[calisnow]

I wanted to put to rest once and for all the notion of "truck lust;" that is, the feeling that the car will creep closer to a truck in an adjacent lane while autopilot is engaged.
...
Sorry to disappoint many of you, but there was virtually no change in the position within the lane during any of the above encounters. ...

Conclusion:
No measurable truck lust. My interpretation is that the reported truck lust is more likely an illusion, rather than a true vehicle movement.

Sorry to disappoint you @tel but you haven't the foggiest idea what the phenomenon the term "truck lust" refers to actually is. Sorry to be making fun of you but you are hilariously haughty in your tone - so when you show us you have no idea what you are talking about it's just too funny.

Truck lust is *not* a gradual creeping closer to a big rig which is adjacent to you. It is a sudden, very fast dive toward the rear end of a big rig as you pass that truck. You don't understand the phenomenon so you're testing for something which nobody has every reported lol. Also it is nowhere near one in 22 times being next to a truck - it is far less frequent than that. Whether it is entirely gone in new builds is unknown but I experienced it last week on an AP 1 2015 70D - and it was the first time in almost 2,000 miles of driving on that build that it happened. I have experienced it numerous prior times and there are plenty of videos on the net of it happening.

I'm not sure what the point of your "study" was in the first place.

Sorry, that's not exactly how science works. Absence of evidence is not evidence of absence.

It's a good observation, and thank you for doing a controlled experiment. But I can tell you this much, in the time since AP1, and over 35k miles on my car (probably 1/2 on AP), I've had truck lust happen twice. Both times I had to jerk pretty hard in order to avoid hitting the truck. I've probably passed 100s of trucks (thousands?) in that time period in over a dozen states. Could my "truck lust" really be something else? Sure it can. But neither of us has proven that it exists or doesn't exist.

I'm not sure why people are skeptical about this one - so many people have experienced it, there are many videos of it - and there are plausible explanations for it. A glitch in the neural network. What would be surprising would be if there were NO bizarre error behaviors exhibited by autopilot.

Malcom, there's LOTS of video, and all of it is just me driving next to trucks. I can get a few sample clips together and post links to them, if you're that interested in redoing the measurements.

Yep - so that's great - you driving next to trucks - which is not what truck lust is. But A for effort.

...The best we can say is that "there is no evidence to suggest that truck lust exists." .

Not true. The best we can say is you didn't bother to do any reading before you went out to see if a phenomenon nobody has ever reported exists - and came back here believing you took one small step for science lol.