Autopilot - positioning in lane

[outdoors]

Adjusting for that stuff doesn't seem hard....

AP team has been working on this forever. It is hard. I think Mr. Karpathy would disagree with you. Moving a little that way has consequences, move a little the other way. Different consequences.

Car AP systems don't work on fear. At least I don't want my car to work on fear. I want them to work on facts to the best of their ability. Which I hope is to some degree better than mine.

 

[tzlaxmr]

AP team has been working on this forever. It is hard. I think Mr. Karpathy would disagree with you. Moving a little that way has consequences, move a little the other way. Different consequences.

I get it. You took my quote out of context.

I think general computer vision is hard, and accounting for the issues you need to account for reduces to having to solve that if you lane-position. My point is that lane-centering indicates they are not that close to solving this (admittedly hard) problem in production yet, given the data we as customers have via our experience with EAP (but I don't have v9 yet).

Car AP systems don't work on fear. At least I don't want my car to work on fear. I want them to work on facts to the best of their ability. Which I hope is to some degree better than mine.

It's not 'fear', it's rational margin of safety given the physics of fast moving multi-ton objects. 10yr bet: when autonomous cars are viable, they will not always target dead-center of the lane.

 

[outdoors]

I get it. You took my quote out of context.

I think general computer vision is hard, and accounting for the issues you need to account for reduces to having to solve that if you lane-position. My point is that lane-centering indicates they are not that close to solving this (admittedly hard) problem in production yet, given the data we as customers have via our experience with EAP (but I don't have v9 yet).

It's not 'fear', it's rational margin of safety given the physics of fast moving multi-ton objects. 10yr bet: when autonomous cars are viable, they will not always target dead-center of the lane.

Dude welcome to the forum. I don't bet. I just drive my car, and can't wait for 10 years for perfection. Out of context or in. It is hard, and I am glad you understand that.

Multi ton or not. If a semi wants to rock your world it won't matter if your inches to the left or right for good feelings. I agree at some point it will be as you say, and to not use it or be afraid is a rational feeling. I see it when people are passengers in my car and the car drives next to a semi. Light handed on the wheel. Newbies cringe. Understandable.

AP has saved my bacon at least a handful of times in the last 12 months and 60k. So I will say semi-autonomous cars are very viable to me at this point as long as they stay someplace in between the lines.

Side note. My S is multi ton. So times that by 10 or 20 for a semi.

 

[Dmagyar]

My experiences with auto pilot is, it's quite a nice feature, handled everything OK, but watch for traveling in the right lane on the freeway when encountering exits and the merging lanes that follow, the car will position in the center of the widened lane, moving you further to the right. Got my attention immediately as I thought the car was exiting the freeway. When traveling into heavier traffic, good time to cancel, till traffic opens up. seemingly setting up emergency braking as others entering freeway/changing lanes aren't too discerning about how they accomplish that simple task.

 

[Dana1]

Yeah, I only use autopilot for stretches without anything like that.

In any event, I disagree with keeping the car dead-center. Sure, I'd probably get used to it. But isn't it safer to position a foot or two farther away from traffic if you can? Especially around turns, opposing cars can ride the line, or even cross it, with some frequency.

My guess is that Tesla chose middle because it's simpler, not safer. ie, if they had a rule that moved the car a foot or two away from traffic in other lanes, then they'd have to be that much better at recognizing cyclists, pedestrians, or other obstacles in the shoulder. The fact that they chose to not do that suggests true self drive is far off.

The car would have to have judgment on whether traffic is closer on the right or left. That might be challenge to code right now.

 

[JSergeant]

The car would have to have judgment on whether traffic is closer on the right or left. That might be challenge to code right now.

There are some very easy logical rules which they could implement right now:

- if you are in the left hand lane (on a highway), keep to the left of center
- if you are in the right hand lane (on a highway), keep right of center

 

[Dana1]

There are some very easy logical rules which they could implement right now:

- if you are in the left hand lane (on a highway), keep to the left of center
- if you are in the right hand lane (on a highway), keep right of center

And if there’s a broken down car in the emergency lane? If there’s a concrete barrier rather than lane markings? Center of the lane removes a lot of variables *shrug*. KISS principle at work.

 

[JSergeant]

And if there’s a broken down car in the emergency lane? If there’s a concrete barrier rather than lane markings? Center of the lane removes a lot of variables *shrug*. KISS principle at work.

Maybe so, but I personally won't use autopilot on highways with trucks until they fix this. I always disable it when passing trucks. My life is more important than taking the chance of a truck veering into my lane.

 

[jsmay311]

One thing you'll also learn, if you haven't already, is when you see a car about to cross the road in front of you, take it off autopilot. You know the car will be out of your lane in time, but it doesn't so it jams on the brakes. Once you get whiplash once or twice, you learn to take it off autopilot.

Or maybe use it only as intended and not on streets with cross traffic?

Fair enough, but it behaves the same (s***ty) way on TACC too.

Not having any form of cruise control that can be used on 85% of my driving (not limited-access highway) without glaring drawbacks is probably the thing that I hate most about my car.

 

[diamond.g]

Fair enough, but it behaves the same (s***ty) way on TACC too.

Not having any form of cruise control that can be used on 85% of my driving (not limited-access highway) without glaring drawbacks is probably the thing that I hate most about my car.

Does follow distance and FCW sensitivity matter?

 

[Dana1]

Maybe so, but I personally won't use autopilot on highways with trucks until they fix this. I always disable it when passing trucks. My life is more important than taking the chance of a truck veering into my lane.

Fair enough though if a large truck veers into your lane it probably doesn’t matter who is driving...law of gross tonnage always applies

 

[jsmay311]

Does follow distance and FCW sensitivity matter?

I'm not 100% sure on this, but I believe FCW sensitivity does not matter, and following distance either doesn't matter or matters only a little bit. (I'm guessing the latter.)

I almost always have the following distance set to "1" (the minimum), and it still brakes for cars turning left at intersections hundreds of feet away.

 

[shanthenew]

Just started using auto-pilot and navigate on auto-pilot in my new Model 3. Love the car and the technology to death, but several times while on autopilot the car veered off the road on to the shoulder and would have hit the divider if I hadn't intervened.

 

[Phlier]

Just started using auto-pilot and navigate on auto-pilot in my new Model 3. Love the car and the technology to death, but several times while on autopilot the car veered off the road on to the shoulder and would have hit the divider if I hadn't intervened.

MOST of the time, these situations happen when the lane markings are difficult to see. If you get a chance, drive back through that section and look at how well the lane is marked.

Tidbit: Your car's confidence in how it sees the lane markings is shown on your screen. The darker the grey of the lane markings it shows on both sides of your car on the screen, the higher the confidence it has in them. You'll notice that when there's no lane markings at all on the screen, AP won't be available.

 

[shanthenew]

MOST of the time, these situations happen when the lane markings are difficult to see. If you get a chance, drive back through that section and look at how well the lane is marked.

Tidbit: Your car's confidence in how it sees the lane markings is shown on your screen. The darker the grey of the lane markings it shows on both sides of your car on the screen, the higher the confidence it has in them. You'll notice that when there's no lane markings at all on the screen, AP won't be available.

Thanks, Phlier! That's helpful and reinforces my suspicion that deficient lane markings might have been to blame. But it should have informed me in some compelling way (I was hype attentive) that AP was no longer available and I should take back control, instead of me having to monitor the screen to my right while also tracking the lane departure under way outside on my left. I will definitely try it again and given how obsessed I am with Tesla, Elon and his mission, I will take this in my stride. But Tesla do well not to rely too much on the average Tesla buyer to serve as their testers. There's a lot of them on the road right now and not everyone would have the same level of patience. Sketchy lane markings are not unusual in the wild and AP should have mastered it by now. Again, my biggest issue is with lack of feedback to the driver when AP is taking a break.

 

[Phlier]

Yup, they have *a lot* of work ahead of them, and I completely agree with the majority of your points.

The EU agrees with you to the point that they have severely hobbled Tesla's AP in Europe. In an attempt to make it "safer", they limited the amount of control authority the AP had. This had the very obvious effect of making it less safe; right as you're in the middle of a steep curve, the AP hits the limits of it's authority, and hands control back over to you. /facepalm. Tesla has since had to put in a warning in EU vehicles that tells the driver in advance that it is approaching the limits of its control authority.

And your experience so far is exactly why so many AP 1 owners still think that AP 1 was better. The core functionality of it *did* work better, but unfortunately the technology behind it wasn't scalable enough to build complete FSD off of it.

Elon keeps saying that LIDAR for FSD is a fool's errand. He says that LIDAR has its place, as SpaceX uses LIDAR for the final docking maneuver to dock with the ISS. But it isn't a good tool for FSD. IMO, he's going to learn that it's going to take a combination of his current visual object recognition system enhanced by LIDAR to really get to complete FSD. I truly hope I'm wrong. But that's going a bit off topic.

 

[shanthenew]

Yup, they have *a lot* of work ahead of them, and I completely agree with the majority of your points.

The EU agrees with you to the point that they have severely hobbled Tesla's AP in Europe. In an attempt to make it "safer", they limited the amount of control authority the AP had. This had the very obvious effect of making it less safe; right as you're in the middle of a steep curve, the AP hits the limits of it's authority, and hands control back over to you. /facepalm. Tesla has since had to put in a warning in EU vehicles that tells the driver in advance that it is approaching the limits of its control authority.

And your experience so far is exactly why so many AP 1 owners still think that AP 1 was better. The core functionality of it *did* work better, but unfortunately the technology behind it wasn't scalable enough to build complete FSD off of it.

Elon keeps saying that LIDAR for FSD is a fool's errand. He says that LIDAR has its place, as SpaceX uses LIDAR for the final docking maneuver to dock with the ISS. But it isn't a good tool for FSD. IMO, he's going to learn that it's going to take a combination of his current visual object recognition system enhanced by LIDAR to really get to complete FSD. I truly hope I'm wrong. But that's going a bit off topic.

Not at all .. great conversation .. thank you!

 

[sixela]

Yup, they have *a lot* of work ahead of them, and I completely agree with the majority of your points.

The EU agrees with you to the point that they have severely hobbled Tesla's AP in Europe. In an attempt to make it "safer", they limited the amount of control authority the AP had. This had the very obvious effect of making it less safe; right as you're in the middle of a steep curve, the AP hits the limits of it's authority, and hands control back over to you.

That's no longer true these days. Now the car brakes so that it can take the durve without exceeding the UN standards for lateral acceleration (it's very conservative, but will keep autosteer engaged if you apply a limited amount of accelerator pedal, although if you push it too hard and force it to exceed the lateral acceleration limits it will indeed first warn you autosteer is limited and then if you push it even harder disengage autosteer).

That was a temporary and very stupid hack that lasted only a few months...

Likewise, its behaviour with respect to the five second limit for making lane changes also has changed. At first, the five second would start to tick when you indicated you wanted a lane change, which meant lane changes would be aborted very often just before the car was fully committed to go to the new lane.

Now the clock starts ticking when the car acts on your instruction to change lanes, and they almost never fail (unless it really detects some danger, usually an idiot trying to pass you on the right just when you are trying to move right).

 

[shanthenew]

I don't know if it's just me being a newbie, but I also noticed on my very first run with Navigate using AP on the highway that when AP changed lanes on its own (I had set it to not wait for prompt), it did not appear to use indicator lights. So now I have disabled auto-lane change so that I have to prompt it.

 

[sixela]

It does use indicator lights. You might not hear them when there's a firmware glitch, though, because the sound for the indicator goes over the car entertainment system. BTW, in Europe we're currently forced to indicate we're OK with the lane change.