Autopilot Achiles Heel?

[xrayvsn]

I am a HUGE autopilot fan and engage autopilot every chance I can get on the highway. There are some limitations however that have made for some uneasy behavior which causes me to take over (even though there is no warning given from the dash):

I find most of these instances occur when there is an 18 wheeler in the next lane slightly ahead of me on the passenger side (right lane for me). I'm not sure the range of the autopilot sensor, but it probably is the length of an 18 wheeler because if the driver sort of drifts near the lane markers, I have noticed that my autopilot markings completely disappear on that side. And rather than use the remaining driver's side lane to guide it, the car all of a sudden acts a bit crazy and actually goes toward the truck as if it is trying to visualize the right lane markings). I try to gently nudge the steering wheel to correct it but on more than a few occasions the vehicle fights me and it feels like we have a sinusoidal course because of it (a bit unnerving for the passenger I had at the time as well).

This has happened on both normal sunny conditions but more prone to happen in the rain. I am curious if that has happened to anyone else.

This also occurs if a truck is ahead of me and then changes to go back into the slow lane (i.e. as the truck crosses the lane markers to get to next lane, sometimes autopilot goes crazy).

I am curious if that situation has happened to others or if this is abnormal and I should get it checked out.

 

[jmsurpri]

Do you see the yellow to red cone shaped sensor visuals on that side when the car starts to drift over? Perhaps the 18 wheeler is high enough that the sensors only read the empty space underneath the trailer and thinks there is nothing there. Since those sensors are mounted so low, I've always wondered if they can pick up high riding vehicles like 18 wheelers.

 

[xrayvsn]

Do you see the yellow to red cone shaped sensor visuals on that side when the car starts to drift over? Perhaps the 18 wheeler is high enough that the sensors only read the empty space underneath the trailer and thinks there is nothing there. Since those sensors are mounted so low, I've always wondered if they can pick up high riding vehicles like 18 wheelers.

Honestly I haven't paid attention to the dash when this happens because when this happens I immediately am trying to actively avoid going into the truck and actually fighting (more like wrestling) with the steering wheel (it actually does a significant counter swerve to my guiding it away from path it appears it wants to go). At this point I do a much harder correction and autopilot comes off.

This actually happens before I get to the side of the truck. In trying to picture it, the truck is actually ahead of me and diagonal to me when it starts happening (i.e I still see the back of the truck and wheels). But I also understand what you are envisioning where maybe the car doesn't think there is anything directly to the side of it if the 18 wheeler is higher than the sensors (I haven't had that issue yet though)

 

[thimel]

I also find myself taking over when I am passing a truck that is near the line. I know AP uses the sonar to move away from things in the next lane, but it is too late as far as I am concerned. I've wondered if they could use the camera to see I'm about to pass a truck that is near the line and move to the opposite side of the lane before I get to the truck.

 

[S4WRXTTCS]

It's been happening to me to such an extent that I simply stopped using lane-keeping until they address it. Quite simply it's an accident waiting to happen.

The problem is you don't know when it will happen. 9 times out of 10 it won't happen, and it will be fine. But, all it takes is one time for the sinusoidal control issue to happen, and suddenly you've knocked into the side of the semi or best case it makes you look like a drunken idiot as you bounce back and forth in your lane.

It is quite simply bizarre because I don't remember having the issue last summer. I don't know if it's a firmware update or something specific about my car.

I do notice a reduction in line detecting when the windwhield has tons of dead bugs on it. That's expected, and I'm okay with that. But, I'm not okay with this quick search thing it does. Where the car immediately tries to find the missing line as soon as it happens. When can't it just hold onto the line it has or ask me to take over?

Last year during the summer I had thousands of miles of uneventful full autopilot miles because I knew what to expect. If I was in the right lane I knew to expect it to dive for the exits. But, I don't recall ever having issues with it being attracted to semi's like a magnet.

This spring it doesn't dive for exits (at least not like it did), but it does immediately slide over as soon as the missing lines things happens just as you're coming up to the truck that's on the right. I can't even go 50 miles without it doing the semi truck magnet thing.

 

[Rockster]

My wife and I refer to this as the bizarre magnetic attraction that our Model S seems to have for trucks. It would almost be comical if it weren't a $100,000+ vehicle acting in a freakishly life threatening manner...

 

[Superloud]

I don't have AP, but I've tried it on a test drive and read countless threads on it. I could probably write a few pages about why I ultimately decided I didn't want AP (until it goes fully autonomous). But this probably isn't the thread for that.

All I want to say right now I guess, is that for a while I've been wondering if AP is at least as safe as a human let's say, just throwing a number out there, 99% of the time. But, what if it's not as safe 1% of the time? Then the next obvious question is, well, how much less safe is it during those 1% occurrences? Threads like this make me feel like maybe AP is *way* less safe, a very tiny amount of the time. Which sure, if the driver is really fully ready to take over 100% of the time, it's not a big deal. But, what happens when drivers get complacent, and the 1% of the time you're not paying attention coincides with the 1% of the time AP wants to crash you?

Edit: btw, this isn't meant to bash AP. It's more of a thought exercise. I'm still trying to fully understand the safety trades one makes when they decide to become a daily AP co-pilot.

 

[Boatguy]

AP will never come out of beta with the v1.0 hardware, it simply can't see enough (wide or distance) or process what it sees fast enough. We're a large pool of beta testers, someone else will get the production release. Or we'll all buy new cars!

 

[S4WRXTTCS]

I don't have AP, but I've tried it on a test drive and read countless threads on it. I could probably write a few pages about why I ultimately decided I didn't want AP (until it goes fully autonomous). But this probably isn't the thread for that.

All I want to say right now I guess, is that for a while I've been wondering if AP is at least as safe as a human let's say, just throwing a number out there, 99% of the time. But, what if it's not as safe 1% of the time? Then the next obvious question is, well, how much less safe is it during those 1% occurrences? Threads like this make me feel like maybe AP is *way* less safe, a very tiny amount of the time. Which sure, if the driver is really fully ready to take over 100% of the time, it's not a big deal. But, what happens when drivers get complacent, and the 1% of the time you're not paying attention coincides with the 1% of the time AP wants to crash you?

Edit: btw, this isn't meant to bash AP. It's more of a thought exercise. I'm still trying to fully understand the safety trades one makes when they decide to become a daily AP co-pilot.

I think it's important to shelve any discussion of autonomous driving, or even semiautonomous driving and instead focus on the individual components that comprise of the Version 1.0 AutoPilot system. To think of them instead as driving aids.

TACC when used correctly I believe is safer than a human alone. It keeps the focus of attention on the road ahead, and not on maintaining speed. It causes less aggravation, and stress especially during long drives. There are few edge cases where it will slow down for no particular reason, and this can be startling. The most important part is to error on the side of caution and treat it as if it's just an assistant. The way I see it is TACC is computer+human acting in conjunction vs. one or the other. If I see a cute girl in the car next to me it the TACC will still be looking ahead. Or if the TACC doesn't see a stopped car that the car ahead of me moved over because of then I'll correct for it. Of course there will be cases where one system thought the other system was in control. But, I think overall it's much safer.

Lane Keeping as it's currently implemented is considerably less safe than a human alone. It removes a lot of situational awareness while not really giving much in return. Sure it's nice occasionally to rest your hands, but it's not the level of reduction in fatigue like what TACC is. It also has no way to deal with things that happen on the road. If I was looking at it from a purely safety perspective I wouldn't use it at all. Instead I see it as something that's in fairly early phase of development. Sure Tesla can say it does a better job than a human because it stays more centered in the line, but that does't address the fact that a lot of us drive to one side to avoid something on the other side. Like ruts, huge trucks, or debris.

Summons is fun, but a bit gimmicky. I use it on a daily basis to pull into my garage because I'm too lazy to hook up a tennis ball.

Parking is where Tesla should focus on because apparently Tesla owners are completely incapable of parking. I laugh at how often I see a Tesla Model S parked in a supercharger all crooked.

As to whether someone should get it I believe it's worth it just on TACC alone. But, that assumes they have a AP Capable Tesla and they do a fair amount of freeway driving especially with traffic.

If they don't have the AP hardware then it's probably best to wait for the next generation of hardware. Going to the three camera solution will improve the current system considerably.

 

[pedriscoll]

While I am a real autopilot fan and use it every day, I agree that the TACC is the more mature component. I like using both under good conditions and feel that most of the time the combination is a better driver than I am. However I will disengage lane keeping whenever conditions are not optimal, or in very heavy traffic. In heavy traffic TACC still excels. Overall though, I use both extensively and feel that autopilot reduces my stress level by about 80% during my daily 25 mile commute.

 

[supratachophobia]

My wife and I refer to this as the bizarre magnetic attraction that our Model S seems to have for trucks. It would almost be comical if it weren't a $100,000+ vehicle acting in a freakishly life threatening manner...

It actually already has a name, "truck lust".

Do you see the yellow to red cone shaped sensor visuals on that side when the car starts to drift over? Perhaps the 18 wheeler is high enough that the sensors only read the empty space underneath the trailer and thinks there is nothing there. Since those sensors are mounted so low, I've always wondered if they can pick up high riding vehicles like 18 wheelers.

They are called whiskers, not cones.

 

[calisnow]

...rather than use the remaining driver's side lane to guide it, the car all of a sudden acts a bit crazy and actually goes toward the truck...

As of the last few firmware updates my 70D no longer "truck dives" or whatever we call it. Which firmware do you have? I know what you are describing because my car used to do that but it has stopped.

However I can't comment on the rain issue - we don't get enough here in Cali to be able to say. Perhaps in rain the Tesla is still doing the truck dive.

 

[xrayvsn]

As of the last few firmware updates my 70D no longer "truck dives" or whatever we call it. Which firmware do you have? I know what you are describing because my car used to do that but it has stopped.

However I can't comment on the rain issue - we don't get enough here in Cali to be able to say. Perhaps in rain the Tesla is still doing the truck dive.

My version is 7.1 2.20.30. Did not notice this behavior prior to this version (especially the strong fighting back of the car going in opposite direction of where I am trying to nudge it)

 

[calisnow]

My version is 7.1 2.20.30. Did not notice this behavior prior to this version (especially the strong fighting back of the car going in opposite direction of where I am trying to nudge it)

So weird - I have the same firmware. Have driven several hundred miles on autopilot since getting it and passed quite a few 18 wheelers in the lane immediately to my right. No diving for them - but I did get this behavior on earlier firmware versions. I'll keep a look-out and see if it does it at all - if so I'll post back here on this thread.

On the bright side I let it "truck dive" one time without grabbing control back myself and I can report that yes, it got way too close for comfort but no, it did not hit the truck - it swerved away from it before hitting. So - I believe the truck diving would not result in an actual collision but of course we can't go systematically test that theory.

 

[Reeler]

The "truck lust" makes the AP more nerve racking than turning off the feature.

 

[brec]

... It is quite simply bizarre because I don't remember having the issue last summer. ...

Last year during the summer I had thousands of miles of uneventful full autopilot miles ... [emphases added]

Were you a beta tester, or since we now have a beta, maybe I should say alpha tester? Version 7.0's debut was in mid-October 2015.

 

[calisnow]

AP will never come out of beta with the v1.0 hardware, it simply can't see enough (wide or distance) or process what it sees fast enough.

What's your source for deciding the processing speed is not adequate for what AP 1.0 is doing?

 

[calisnow]

I don't have AP, but I've... read countless threads on it. ... I've been wondering if AP is at least as safe as a human let's say, just throwing a number out there, 99% of the time. But, what if it's not as safe 1% of the time? Then the next obvious question is, well, how much less safe is it during those 1% occurrences? Threads like this make me feel like maybe AP is *way* less safe, a very tiny amount of the time. Which sure, if the driver is really fully ready to take over 100% of the time, it's not a big deal. But, what happens when drivers get complacent, and the 1% of the time you're not paying attention coincides with the 1% of the time AP wants to crash you?
I'm still trying to fully understand the safety trades

FWIW I have had autopilot for a few months and about 4,000 miles of use so far. Here is my feedback for you:

1 - You will never understand autopilot until you own it - you get next to no useful understanding from reading threads about it. Your sample here is ridiculously skewed.
2 - It is rock solid in its intended conditions far more than 99% of the time - if it was only safe 99% of the time you'd have heard of it by now.
3 - It has never tried to crash me - not once. Even the few times I have gotten the "take over immediately" buzz it is simply because it is losing the lane marker on an outside curve - but it wasn't headed directly for a wall - it just started to wobble. And that is very, very infrequently. Less than 1% of 1% if distance driven - if even that.

Finally - you don't have access to the statistics and neither do I. We are both largely blowing hot air around in speculation, but in my experience autopilot is extremely safe and getting safer all the time (ie more accurate) when used within its limits. Statistically I believe it is at least as safe as a human when used on freeways - safety being defined as accidents per mile.

Having said all that I will be upgrading to version 2 as soon as Tesla releases an S with it.

 

[davidc18]

Love the autopilot features, use them everyday, everywhere it will engage.

I've also noticed the latest firmware versions have "semi" lust which is a bit of a bummer. Hope it will all get sorted out in newer firmware versions, and I would ask that everyone email these issues to serviceNA@... so that it can get addressed.

 

[David29]

While I am not sure I have observed the "truck lust," I have observed that the sensors do not recognize the spaces below trucks and between the wheels. After reading a few of these posts earlier this week, yesterday I made a point of watching the screen display as I passed trucks. And I was a bit surprised to see how the system recognizes the wheels but completely misses the truck or trailer body between the wheels in many cases, apparently because the sensors are aimed low and the truck bodies are high off the ground. Hopefully my recognition of that will help me to be more cautious when using the autopilot features...