Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Elon: Autopilot, new constraints on when it can be used

This site may earn commission on affiliate links.
Have you tried using AP on city curves? It has a high pucker factor, and I can tell you with 100% certainty my hands are on the wheel.

Max, here is a very short - 15 seconds video of AP losing the markings because of a combination of a widening road to prepare for an additional lane, happening on a curve.

 
Last edited by a moderator:
Interesting video. At the beginning AP is tracking the painted lines on the sides of the single lane road. As the lane widens, even though the lines are still present AP stopped tacking both the lines even though the lane appeared to be only about a third to a half again wider than at the beginning of the video. Then at the end of the video when the center dashed line appears AP starts tracking it after starting to again track the right side line just a second earlier.
 
Max, here is a very short - 15 seconds video of AP losing the markings because of a combination of a widening road to prepare for an additional lane, happening on a curve.



You didn't grab the wheel? I would have...

I've had that happen to me also, but I stand by my point (and I know we're in agreement), plenty of time to react if you're paying attention. Attention is key, hand placement is secondary.
 
Last edited by a moderator:
You didn't grab the wheel? I would have...

I've had that happen to me also, but I stand by my point (and I know we're in agreement), plenty of time to react if you're paying attention. Attention is key, hand placement is secondary.

@CanadaEV, just to amplify my answer to you above, this is exactly the kind problem case I am scanning for and aware of whenever I have the autopilot on, even with hands a half an inch off the wheel, but not touching it.
 
Max, here is a very short - 15 seconds video of AP losing the markings because of a combination of a widening road to prepare for an additional lane, happening on a curve.



This video timebase seems altered, not realtime is it?

Hmm. trying to guess where on Hwy 1 that video was taken?
 
Last edited by a moderator:
Interesting video. At the beginning AP is tracking the painted lines on the sides of the single lane road. As the lane widens, even though the lines are still present AP stopped tacking both the lines even though the lane appeared to be only about a third to a half again wider than at the beginning of the video. Then at the end of the video when the center dashed line appears AP starts tracking it after starting to again track the right side line just a second earlier.

ecarfan, yes, exactly right. I have been shooting video like this, (side by side) to enable the analysis you make here. It really helps me understand how the Autopilot works, and when it is more or less likely to behave well. It will improve as it learns, for sure. I have a lot of respect for the work that has gone into it.

- - - Updated - - -

This video timebase seems altered, not realtime is it?

Hmm. trying to guess where on Hwy 1 that video was taken?

The time of both video streams is identical, in real time. I match the streams to the same 1/30th of a second of the video frame rate. It was taken north of Victoria BC, south of Duncan.

- - - Updated - - -

You didn't grab the wheel? I would have...

I've had that happen to me also, but I stand by my point (and I know we're in agreement), plenty of time to react if you're paying attention. Attention is key, hand placement is secondary.

Actually, I did react in that situation, almost instantly, as my hands were already on the wheel, so I simply tightened my grip. As you can see, I turned my signal on to indicate I would be choosing the right lane if there was someone behind me. And yes, I believe you would not have had a problem if your hands were close to the wheel and you were watching carefully.

- - - Updated - - -

This video timebase seems altered, not realtime is it?

Hmm. trying to guess where on Hwy 1 that video was taken?

Scott, by 'timebase' you might be referring to the speed of the clip. Yes, it is slowed down so you can see the relationship between the two clips a little more easily.
 
If you look at this in grayscale, at just after the divergence point, you likely get a better determination of why this was the case for the camera
image.jpeg

That central divider change, the leaves obscuring the line and oncoming traffic angle would be my guesstimate why.

again, reinforces why we should continue to reinforce use as a driver aid, not a driver replacement ;-)
 
The autopilot, its behavior and learning fascinate me. When I have it on I sit there scanning the view, focusing on trying to think like the autopilot, or rather I try to anticipate situations of road configuration, lighting (e.g. striped shading from sun through trees), positions of stopped cars, etc., that would give it trouble. If I see potential trouble coming up I have one or both hands hovering a half an inch off the wheel, ready to grab it. Otherwise I have my right elbow on the arm rest and my hand open a few inches away from the wheel ready to grab it. Sometimes, when it gets through a tricky spot without a hitch, I tell it 'good job' (when I'm alone in the car). What I have said applies best when I'm on a good two lane road (federal highway). I don't use it on the narrower, shoulderless country roads where I live, because I know my reflexes aren't quick enough to catch the wheel safely before it sends me into a ditch. Interstates, on the other hand, are clearly within its design parameters, so I tend to have my hands in front of me, but not hovering by the wheel. But I still spend my time trying to think of what I am seeing in terms of how the autopilot is likely to process the scene.

I don't want to have my hands on the wheel in part because I don't want to give the system unconscious directional cues that would invalidate my observations of how it's doing. It would also, in my opinion, destroy the fun of using the technology. I would probably just drive the car myself and stop posting reports (in other threads) about specific problem instances, cases of learning, etc., although for all I know, other participants in those threads would rather have me put my hands on the wheel, turn the system off and stop boring them with those posts.:smile:

Having people with your level of interest in the Autopilot as beta testers will contribute a lot to its success. I think there are many others on these forums that are equally avid fans of this technology, and of course most of what Tesla is doing. It's unfortunate that the bell curve has to include people who don't have the respect for these things that they deserve. They make it difficult for everyone. I can completely relate to what you say - I have talked to mine at times too, because you get the sense that it embodies so much intelligence from its designers that you appreciate its accomplishments. And it learns! It's quite amazing.

That's an interesting idea about giving it directional cues. The same idea occurred to me, but I thought it was not possible. I tend to think that it is either on or off, and not susceptible to other real time input. But it may be possible, and even a real benefit. At this time, I speculate that if we correct a wrong AP decision such as trying to exit a freeway when we don't want to, - that set of actions is probably information that is transmitted to Tesla. If Tesla gets repeats of that wrong AP decision through other similar corrections, they then issue a fleet wide instruction to curtail that decision at that location.

Your idea of real time directional cues could be a much more sophisticated and quicker learning method. It's possible, but I don't think they have anything like that in the AP, at least not yet.

- - - Updated - - -

If you look at this in grayscale, at just after the divergence point, you likely get a better determination of why this was the case for the camera

That central divider change, the leaves obscuring the line and oncoming traffic angle would be my guesstimate why.

again, reinforces why we should continue to reinforce use as a driver aid, not a driver replacement ;-)

I've noticed that it is a regular thing for AP to lose one marking when the lane width exceeds a certain amount. It usually tries to find the right side lane marking, which is a lot smarter than trying to find the marking on the oncoming traffic side. After finding the right side marking, it continues driving, and usually either a new left lane marking will appear, as in this case, (but here, after it turned itself off), or the right marking will close with the left marking, and it will pick it up. That is most common in the reverse of this example. When two lanes narrow into one lane, first the center line disappears, but when that happens the road remains wider than one lane. AP finds the right side marking and hugs it. It is then shepherded by that marking to the left as the road narrows to one lane, until it picks up the left boundary marking of the one remaining lane.
 
Last edited:
That's an interesting idea about giving it directional cues. The same idea occurred to me, but I thought it was not possible. I tend to think that it is either on or off, and not susceptible to other real time input. But it may be possible, and even a real benefit.

Your idea of real time directional cues could be a much more sophisticated and quicker learning method. It's possible, but I don't think they have anything like that in the AP, at least not yet.

It's an interesting and unsettled question. There were at least two occasions when I experimentally tugged the wheel at points where it had not yet learned, and it seemed to accept the cue and move into the correct lane (once where a single traffic lane widened to two, once where a right exit lane came off a left hand curve in the road). However, on a couple other occasions in other locations it resisted the cue, so the times it seemed to have taken the cue might have been coincidental independent decisions on its part to go that same way. On the other hand I have seen a couple posts by others saying they could teach it the correct way to go by tugging the wheel. Unfortunately I can't recall which thread.

Relatedly, I had originally thought learning was all either fleet learning or occurred while AP was not engaged, where it compared the driver's steering with it's calculated path and flagged differences. But now I'm not do sure. Please see my post here:

#249

- - - Updated - - -

I've noticed that it is a regular thing for AP to lose one marking when the lane width exceeds a certain amount. It usually tries to find the right side lane marking, which is a lot smarter than trying to find the marking on the oncoming traffic side. After finding the right side marking, it continues driving, and usually either a new left lane marking will appear, as in this case, (but here, after it turned itself off), or the right marking will close with the left marking, and it will pick it up. That is most common in the reverse of this example. When two lanes narrow into one lane, first the center line disappears, but when that happens the road remains wider than one lane. AP finds the right side marking and hugs it. It is then shepherded by that marking to the left as the road narrows to one lane, until it picks up the left boundary marking of the one remaining lane.

Yes! But in my experience AP did not initially hug the right line when gaining or losing a lane. When I used it for the first time at locations where the road gains a lane it would either wander in the middle and jerk to the right when it was about to overrun the new dashed line that appears in front of it, or it would just shut off. By the third try it would hug the right side. Maybe the fleet has now learned to hug the right line when the car comes across a new location for the first time. I haven't been able to test this because I tend to drive the same roads, and AP has now learned all the lane pick-ups and drops on those roads. I need to try new roads.
 
The time of both video streams is identical, in real time. I match the streams to the same 1/30th of a second of the video frame rate. It was taken north of Victoria BC, south of Duncan.
I don't know if it's just my computer or what, but that looks much slower than real-time. You're going ~50mph there, and yet trees and cars appear to be passing at walking speed.
 
Those who worry about restrictions, remember that Tesla is of the few manufacturers allowing full time rear view camera and can be trusted to implement restrictions only when necessary. It would be wise to put in restrictions before accidents happen so they will not be forced to overreact.

That's partly true. They allow the rear camera to stay on. They don't have a nag screen. They don't allow video in the browser. They said on their web page that it's for legal reasons. (i.e. it's not legal to do so) Other manufacturers (such as Infiniti) allow video in the front as long as the car is in park and the parking brake is on. Not only did they implement a restriction even when it's not necessary, they outright told people a reason for it that wasn't true.

- - - Updated - - -

Elon is less tolerant of BS then we are and knows well that a lane keeping function that requires your hands on the wheel is near useless.

While some manufacturers want to switch it off if they don't detect hands on the wheel, Tesla should do the opposite. If they detect hands off the wheel, they should turn it on. They should call that a safety feature and tell you not to rely on it any more than you'd rely on active emergency braking to stop the car instead of using the brake pedal.
 
My guesses:

1) Autopilot won't let you go more than 5 mph above speed limit. ...
2) "Hold the wheel" indicators much more often or even continuously when autopilot is used in situations it's not meant to be in only....
3) Seat weight sensor integration.

What happens when the car gets the speed limit wrong? It happens a fair amount. If the car misreads a "55 MPH -- Trucks" and tells me that I need to go 55 mph (60 with margin) then what should the car do when I'm going 70 in a 65 zone?

Item two seems reasonable. I already expect that behavior, and if they tune it, so much the better. Let the warnings go away when the car is sure.

If they integrate the seat weight, then what should they do? If a person is out of the seat and facing the wrong way, obviously disabling the system would make it worse. So designing it so that it can't be used under that circumstance won't work. Perhaps a flashing and beeping "sit your ass down" message while the car gradually slows down would be in order, but it's not something Tesla can simply decide on a whim. No matter what they do, it will have consequences. Since a driver has the option of taking over and using the pedals, it might be justifiable to do things that would be reasonable with a person in the seat who can take over.
 
  • Like
Reactions: davidc18
If they integrate the seat weight, then what should they do? If a person is out of the seat and facing the wrong way, obviously disabling the system would make it worse. So designing it so that it can't be used under that circumstance won't work. Perhaps a flashing and beeping "sit your ass down" message while the car gradually slows down would be in order, but it's not something Tesla can simply decide on a whim. No matter what they do, it will have consequences. Since a driver has the option of taking over and using the pedals, it might be justifiable to do things that would be reasonable with a person in the seat who can take over.

As soon as the car is stopped, the driver is unable to use AP again. Poof. Gone.

That's what I'd do. 'Abuse AP & lose it.' :)
 
I don't know if it's just my computer or what, but that looks much slower than real-time. You're going ~50mph there, and yet trees and cars appear to be passing at walking speed.

For sure it is slow motion, because it all happens a little fast to switch your focus back and forth between the AP view and the driver view. What I meant was that the two video streams were synchronized correctly in real time. After edited into one stream, it was slowed down for convenience.
 
I tend to agree with Douglas Adams -- never underestimate the ingenuity of fools.

I don't think there's any way to make this foolproof. They don't even have proper sensors to tell whether your hands are on the wheel. Detecting whether someone is looking forward is... error-prone, particularly with sunglasses. Et cetera.

I think that YouTube video is the problem. Tesla should probably mail it to her insurance company and see what happens next. That might be the best move. :)