Model X Crash on US-101 (Mountain View, CA)

[Peteski]

Classic cruise control has a very simple and reliable behavior. When the driver turns it on, classic cruise control holds the vehicle's speed roughly constant until the driver cancels the activation.

This behavior is easily understood, and relieves the driver from having to modulate the accelerator pedal in order to maintain constant speed (something most drivers aren't very good at, especially on wide open roads). Because classic cruise control doesn't take over the steering at all and is incapable of responding in any way to any traffic, a driver can't rely on it as an excuse to pay less attention to the road.

By contrast, AP has a very complicated behavior. It will keep the car in lane, below a maximum speed, and no closer than a programmed follow distance, except when it doesn't. And there is no surefire way to know when it will or won't work, since its limitations are complicated and its abilities/logic are not clearly published, are complicated (lots of ai code), and change frequently due to thinly disclosed over the air updates. Plus, when it works, AP can handle both steering and acceleration, leaving the driver with little to do-- until the driver suddenly needs to do something.

I agree that AP (as in AS + TACC) gives a driver the ability to become totally inattentive, which can be a problem for those drivers who grossly over-estimate its ability to keep them safe in all situations. I think in practice the vast majority of people who do use AP have enough self-preservation sense to pay full attention to what the system is doing and they learn when it is appropriate to use AP and when to switch it off or at least prepare to over-ride if necessary. I find the system totally reliable when following a simple lane with no complications like gore points or merging lanes with dubious markings, etc. So it's perfect for UK motorway use and most A-road dual carriageways. Very rarely on those roads do you need to over-ride AP.

So is AP safer than manual steering on a motorway? Judging from what I see of other road users, I would say definitely yes! I see cars and trucks wandering all over the place and tailgating dangerously close on a daily basis while distracted by their media lifestyle! Every single day there are crashes, minor and sometimes major on our motorway networks with 99.99% of them as a direct result of some human error while behind the wheel. I reckon many of those crashes could have been avoided using AP, even if the driver was not paying full attention at the time. But obviously we don't hear about any of those scenarios, only the few crashes were a Tesla and possibly AP was involved! It leads to a very skewed picture in the media. Maybe they should analyse ever fatal crash and decide whether or not AP would have prevented that crash. For sure there will be plenty of cases where AP would have probably saved lives.

 

[Peteski]

Shoutout to @Economite

Secondary problem is that the hand-on-wheel recognition technology seems to be extremely limited and thus the data and usability of it is questionable. This results in nags when the wheel is properly and steadily held by two hands, which would be superior for countering such surprises, encouraging instead application of single-hand torque (a common tip on TMC too) - which is not nearly as useful in an emergency counter-move...

Interesting to see how this area develops in the future updates.

I agree entirely with this point. It only uses simple steering torque to register hands on which is not great. It would be better to have 2 touch sensors on either side of the wheel to register both hands on but not necessarily applying any steering force. That coupled with eye recognition would be ideal i.e. checking for hands on and looking out of the front screen.

I think the problem is that the system was initially developed to be hands off and the steering torque sensor was all they had to quickly implement a "nag" solution. Steering touch sensors would require a retro-fit of all steering wheels. But it's something they could implement fairly easily on new car builds with a re-designed wheel.

 

[mongo]

Indeed, there is a difference between:

1) Autopilot failing to prevent an accident that would have happened in any case (expect through intervention).
2) Autopilot taking active steps that result in an accident, when lack of such steps would not have resulted in an accident.

Is there?
1a) Driver fails to guide vehicle safely with heading actively chosen by driver (steering but not paying enough attention)
1b) Driver fails to guide vehicle safely with heading passively chosen by driver (via relying on physics and not paying attention)
2a) Driver fails to guide vehicle safely with heading passively chosen by driver (via relying on AP and not paying attention)
2b) Driver fails to guide vehicle safely with heading actively chosen by driver (via waiting to see what AP does)

I agree with the discussion of ability to maintain attention with reduced workload. However, only in the case of a sudden AP steering change do I see #2 being a separate case, even that requires the driver to not be firmly holding the wheel.

The 101 incident involved reaction times much greater than many everyday traffic events such as sudden backup/ stopped traffic, slow cars merging in front of you, or objects falling from trucks.

 

[Jayrod]

I agree that AP (as in AS + TACC) gives a driver the ability to become totally inattentive, which can be a problem for those drivers who grossly over-estimate its ability to keep them safe in all situations. I think in practice the vast majority of people who do use AP have enough self-preservation sense to pay full attention to what the system is doing and they learn when it is appropriate to use AP and when to switch it off or at least prepare to over-ride if necessary. I find the system totally reliable when following a simple lane with no complications like gore points or merging lanes with dubious markings, etc. So it's perfect for UK motorway use and most A-road dual carriageways. Very rarely on those roads do you need to over-ride AP.

So is AP safer than manual steering on a motorway? Judging from what I see of other road users, I would say definitely yes! I see cars and trucks wandering all over the place and tailgating dangerously close on a daily basis while distracted by their media lifestyle! Every single day there are crashes, minor and sometimes major on our motorway networks with 99.99% of them as a direct result of some human error while behind the wheel. I reckon many of those crashes could have been avoided using AP, even if the driver was not paying full attention at the time. But obviously we don't hear about any of those scenarios, only the few crashes were a Tesla and possibly AP was involved! It leads to a very skewed picture in the media. Maybe they should analyse ever fatal crash and decide whether or not AP would have prevented that crash. For sure there will be plenty of cases where AP would have probably saved lives.

That is a great point and if I was Tesla not a single person would leave that delivery until they watched a video that clearly explains what it can and cannot do. They may already do this I have no idea.

 

[Jayrod]

Is there?
1a) Driver fails to guide vehicle safely with heading actively chosen by driver (steering but not paying enough attention)
1b) Driver fails to guide vehicle safely with heading passively chosen by driver (via relying on physics and not paying attention)
2a) Driver fails to guide vehicle safely with heading passively chosen by driver (via relying on AP and not paying attention)
2b) Driver fails to guide vehicle safely with heading actively chosen by driver (via waiting to see what AP does)

I agree with the discussion of ability to maintain attention with reduced workload. However, only in the case of a sudden AP steering change do I see #2 being a separate case, even that requires the driver to not be firmly holding the wheel.

The 101 incident involved reaction times much greater than many everyday traffic events such as sudden backup/ stopped traffic, slow cars merging in front of you, or objects falling from trucks.

Correct, the report said 5 seconds, that's an eternity in the car wreck game. But, like you said, glance at the radio, look at the mirror etc all take that 3-5 seconds to 1 and bam.

Like I said above Tesla need to strongly consider educating people in a Barney level video that says "hey dumbass" this isn't a complete AI like Skynet.

 

[AnxietyRanger]

Indeed, there is a difference between:

1) Autopilot failing to prevent an accident that would have happened in any case (expect through intervention).
2) Autopilot taking active steps that result in an accident, when lack of such steps would not have resulted in an accident.

Is there?
1a) Driver fails to guide vehicle safely with heading actively chosen by driver (steering but not paying enough attention)
1b) Driver fails to guide vehicle safely with heading passively chosen by driver (via relying on physics and not paying attention)
2a) Driver fails to guide vehicle safely with heading passively chosen by driver (via relying on AP and not paying attention)
2b) Driver fails to guide vehicle safely with heading actively chosen by driver (via waiting to see what AP does)

I agree with the discussion of ability to maintain attention with reduced workload. However, only in the case of a sudden AP steering change do I see #2 being a separate case, even that requires the driver to not be firmly holding the wheel.

The 101 incident involved reaction times much greater than many everyday traffic events such as sudden backup/ stopped traffic, slow cars merging in front of you, or objects falling from trucks.

I agree that a sudden steering or braking event is a much worse example of my case type 2), than the Mountain View incident likely was. There is some suggestion that the Greece event may have been a more sudden steering event and obviously many shadow-braking cases are. I agree that Mountain View likely was not a sudden steering event and that simply lack of driver attention or (less likely given age) driver incapacitation were the main reason (alongside with a host of secondary reasons like sun, road condition etc., unrelated to the driver, car or its maker).

That said, I still consider an "active" steering decision by Autopilot to go outside of a lane a type 2) event, separate from "passive" events like the Joshua Brown incident, fire trucks - and others where lane simply ends at an obstacle - where mere predictable lane following have resulted in disaster through no fault of Autopilot... that I categorized as type 1) above. The Mountain View event certainly seems to be an active event where Autopilot failed.

Put it simply: Lane keeping, unlike stationary object detection, is an AP2 feature. It seems to have failed here, as AP2 seems to have actively left its lane into a non-lane area.

In many other cases (e.g. "type 1"), there was no AP2 feature to fail - there is no fire truck detection, no perpendicular truck detection, no parked car detection. There is nothing in AP2 to blame in these cases, directly anyway.

In this case - and in other "type 2" steering cases like possibly Greece - there was: AP2 lane keeping failed. That is a significant difference in these scenarios, even if the ultimate legal responsibility is the same. The human failure was preceded by an active failure by AP2 to perform its feature, assuming our information is correct of course.

 

[jimmyjohn]

142 Pages on this thread?

Get a life, people!

 

[Economite]

2: IIHS independent report shows lane assist/ auto steer does improve safety: Lane departure warning cuts crashes

Again, IIHS data disagrees with you.
http://www.iihs.org/frontend/iihs/documents/masterfiledocs.ashx?id=2142

Lane assist/ auto setter helps to prevent head on collisions, side swipes, and ending up in a ditch. It is much more important than your "perfectly centered" comment.

Those IIHS links are about Lane Departure Warning, not AS. Lane Departure Warning is a totally different kind of system. It beeps at a driver when the system believes that the driver is drifting out of his or her lane. The notion of the system is that the beep will get the driver's attention, so the driver can correct their mistake (if any). Clearly that's a good system, as it helps to prevent accidents where a driver is losing attention and drifting out of lane. It can only help. It never will steer a driver into danger and there is no way for a driver to misuse it by become reliant on it. It should stop a driver from drifting into a ditch or off the road just as well as AS, yet It has none of the downsides of AS.

 

[Economite]

So is AP safer than manual steering on a motorway? Judging from what I see of other road users, I would say definitely yes! I see cars and trucks wandering all over the place and tailgating dangerously close on a daily basis while distracted by their media lifestyle!

I agree that TACC alone almost certainly improves safety. My doubt about safety is only when AS is added.

AS does nothing to reduce tailgating behavior. TACC can reduce tailgating in theory, though it sounds like it still follows too close under many circumstances if it is set at follow setting 1 (or some other low follow setting).

I don't think inadvertent "wandering" (poor lane tracking) causes many highway accidents. This behavior is incredibly annoying to other drivers. But it is also generally easy for other drivers to avoid getting into accidents with a drifter since the other drivers can always just slow a bit. I think the main causes of accidents are tailgating and poor lane changes, neither of which are addressed by AS. Besides, "wandering" problems can be mitigated just as well by Lane Departure Warning, which has none of the downsides of AS.

 

[mongo]

Those IIHS links are about Lane Departure Warning, not AS. Lane Departure Warning is a totally different kind of system. It beeps at a driver when the system believes that the driver is drifting out of his or her lane. The notion of the system is that the beep will get the driver's attention, so the driver can correct their mistake (if any). Clearly that's a good system, as it helps to prevent accidents where a driver is losing attention and drifting out of lane. It can only help. It never will steer a driver into danger and there is no way for a driver to misuse it by become reliant on it. It should stop a driver from drifting into a ditch or off the road just as well as AS, yet It has none of the downsides of AS.

Yah, it was mostly about warning only systems (due to lack of direct control systems). Regard the bold part of your post, I would argue that if a sleepy/ing driver gets an alert, there is a non-zero chance they will jerk the steering wheel the wrong way or by an incorrect amount. In the situation where there were heading for the ditch, and now swerve into oncoming traffic, the outcome is worse with the warning. Warnings cannot stop the driver form drifting, only give them a stimulus to take action. AP can stop the drift.

From the second link:

Because drivers must respond appropriately to warnings for them be effective, another factor that may have limited the size of the system’s benefit is the physical state of the driver. Cicchino and Zuby (2017) examined the physical state of 631 drivers involved in lane drift crashes in the National Motor Vehicle Crash Causation Survey (NMVCCS), in which in-depth crash investigations were performed on a national sample of vehicles involved in crashes where emergency medical services were called and a passenger vehicle was towed that occurred between 6 a.m. and midnight. Thirty-four percent of the drivers were physically incapacitated in some way, with 17 percent medically incapacitated and 17 percent asleep, and an additional 14 percent had a blood alcohol concentration of 0.08 percent or above, non-incapacitating medical issue, or another physical factor that led to the crash . These proportions were higher when considering only crashes involving fatal or serious injuries. Wiacek, Fikenscher, Forkenbrock, Mynatt, and Smith (2017) examined fatal lane-drift crashes in NMVCCS and found that fewer drivers were sleeping or otherwise incapacitated, although the physical states of many were unknown.

Lane departure warnings cannot help medically incapacitated drivers, but they can possibly prevent sleeping drivers from lane drifts. Warnings have been shown to improve lane positioning and reduce lane departures of fatigued drivers in simulators (De Rosario, Soler, Tenas, Dominguis, &Echeverria, 2010; Gaspar et al., 2017; Kozak et al., 2006; Rimini-Doering, Altmueller, Ladstaetter, &Rossmeier, 2005) . However, their ability to wake real-world sleeping drivers in enough time to prevent a crash is unknown.

I don't think inadvertent "wandering" (poor lane tracking) causes many highway accidents. This behavior is incredibly annoying to other drivers. But it is also generally easy for other drivers to avoid getting into accidents with a drifter since the other drivers can always just slow a bit.

For low rate drifts I agree, mostly. However, for sudden drifts or head on cases (non-divided highway) with high speed deltas, I disagree. Especially when considering the potential (non necessarily best) reactions of other drivers.

 

[ArizonaP85]

142 Pages on this thread?
Get a life, people!

The reason there's 142 pages is that someone lost their life.

 

[Peteski]

I agree that TACC alone almost certainly improves safety. My doubt about safety is only when AS is added.

AS does nothing to reduce tailgating behavior. TACC can reduce tailgating in theory, though it sounds like it still follows too close under many circumstances if it is set at follow setting 1 (or some other low follow setting).

I don't think inadvertent "wandering" (poor lane tracking) causes many highway accidents. This behavior is incredibly annoying to other drivers. But it is also generally easy for other drivers to avoid getting into accidents with a drifter since the other drivers can always just slow a bit. I think the main causes of accidents are tailgating and poor lane changes, neither of which are addressed by AS. Besides, "wandering" problems can be mitigated just as well by Lane Departure Warning, which has none of the downsides of AS.

The problem with LDW is that most people just switch it off because of all the annoying vibrations and beeps when it activates. I'd love to see some stats on its actual usage as I've switched it off on every car I've owned with it so far! In theory it is pretty good though.

TACC is great, I use it on follow setting 6 or 7 on the motorway, 4 or 5 at moderate speeds following along in a line of traffic and only use 1 or 2 in crawling stop/start traffic. Tesla could make it more idiot proof by only allowing closer follow settings based on speed e.g. 6 or 7 only available at 75 mph. But it's still much better than many people under manual speed control. Tailgating is of epidemic proportions.

I agree lane wandering isn't the number 1 cause of accidents, but I've had to take action quite a few times to avoid a collision. Not having to do that at all would be a nice bonus! If Tesla can get auto-lane changing to work reliably with full rearward collision detection, that would be a big step forward in road safety. They might be close to that, although certainly not there today.

So yeah, there's plenty of room for improvement as above, but AS + TACC is still a good system when used wisely, as most people actually do otherwise there would be a LOT more reported incidents than there are with the system.

I think training on use of AP should be provided on delivery too, as it's such a new feature to most drivers and not covered at all on any driving test. I got no training whatsoever when I picked my car up. Basically anyone driving the car should have had some basic instruction on AP and it should be disabled for any unknown driver using the car (by driver profile for example).

 

[Economite]

TACC is great, I use it on follow setting 6 or 7 on the motorway, 4 or 5 at moderate speeds following along in a line of traffic and only use 1 or 2 in crawling stop/start traffic. Tesla could make it more idiot proof by only allowing closer follow settings based on speed e.g. 6 or 7 only available at 75 mph. But it's still much better than many people under manual speed control. Tailgating is of epidemic proportions.

I agree that TACC is good at preventing sloppy speed control by a driver.

My guess, however, is that drivers who are normally inclined to tailgate when they aren't using TACC/AP tend to be setting the follow distance to 1 or 2 when they do use TACC/AP (even at high speeds). Thus, TACC in its current form might mainly be reducing unintentional tailgating (by folks who mess up at modulating speed/reacting to behaviors of lead cars) rather than intentional tailgating (by folks who are either really impatient, convinced that people will "cut in front of them" if they leave any space, just don't have a good notion of what a safe following distance is). For TACC to dramatically reduce this second type of tailgating (which I suspect is the most common), it would need to have the kind of limitation you suggest in the text I bolded in your quote.

 

[mongo]

I think training on use of AP should be provided on delivery too, as it's such a new feature to most drivers and not covered at all on any driving test. I got no training whatsoever when I picked my car up. Basically anyone driving the car should have had some basic instruction on AP and it should be disabled for any unknown driver using the car (by driver profile for example).

How is this?

Welcome everyone to Telsa TACC and Auto Steer training, known together as AutoPilot.
Who can tell me some things AP does currently?
Uh huh
uh huh
uh huh
OK you're all wrong.

AP is not self-driving, AP does not see stopped objects, AP does not care if the road is icy, AP does not obey any type of traffic signs (including speed limit and stop), AP does not read traffic lights. AP will not pull over for emergency vehicles.

What AP does do is reduce the amount of accelerator, brake, and steering minor adjustments that you, the driver, need to make.

TACC will attempt to stay a selectable distance away from the car in front of you. However it may not, if you feel the car is too close, press the brake. If the car is braking for an unknown reason, press the accelerator.

Auto steer will attempt to keep you in your lane. However it may not, if the car deviates from where you want it to be, steer.

Stay alert, keep your hands on the wheel, and be safe out there

 

[Economite]

So yeah, there's plenty of room for improvement as above, but AS + TACC is still a good system when used wisely, as most people actually do otherwise there would be a LOT more reported incidents than there are with the system.

The problem is that everyone who uses AP (including the people who misuse it) seems to think that they are using it wisely (or at least in a manner that is appropriate). This is partially because Tesla barely defines what appropriate use is, and then seems to wink-wink at some of those statements (as when Elon demonstrates AP with his hands off the wheel, despite Tesla officially defining this as inappropriate use). When crashes occur, they tend to blame themselves for not jumping in fast enough to override AP, rather than questioning whether they have been using AP in an appropriate manner.

And there have been an incredible amount of reports of drivers having to quickly react to correct AP's mistakes. Each of these is an instance that creates an opportunity for a crash. And my guess is that most people who have serious collisions while using AP don't report it on these boards, both because they don't want to be insulted by the folks on the boards and because they don't want to make any admissions or statements about their accident that might cause them trouble down the road. And Tesla's tendency to go publicly nuclear blaming the driver any time there is an AP-involved accident that makes it into the press probably further chills drivers from bringing press attention to such accidents. Thus, we mainly hear about AP accidents that are fatal (and therefore automatically make the local press and then get picked up by other press) and that involve emergency vehicles (since the police and firefighters always publicize the crash themselves).

 

[JohnSnowNW]

The problem is that everyone who uses AP (including the people who misuse it) seems to think that they are using it wisely (or at least in a manner that is appropriate). This is partially because Tesla barely defines what appropriate use is, and then seems to wink-wink at some of those statements (as when Elon demonstrates AP with his hands off the wheel, despite Tesla officially defining this as inappropriate use). When crashes occur, they tend to blame themselves for not jumping in fast enough to override AP, rather than questioning whether they have been using AP in an appropriate manner.

And there have been an incredible amount of reports of drivers having to quickly react to correct AP's mistakes. Each of these is an instance that creates an opportunity for a crash. And my guess is that most people who have serious collisions while using AP don't report it on these boards, both because they don't want to be insulted by the folks on the boards and because they don't want to make any admissions or statements about their accident that might cause them trouble down the road. And Tesla's tendency to go publicly nuclear blaming the driver any time there is an AP-involved accident that makes it into the press probably further chills drivers from bringing press attention to such accidents. Thus, we mainly hear about AP accidents that are fatal (and therefore automatically make the local press and then get picked up by other press) and that involve emergency vehicles (since the police and firefighters always publicize the crash themselves).

Holding onto the steering wheel and looking out the windscreen is hardly a difficult point to get across. Appropriate use of the system is really that simple.

FAR more people are likely to report when AP acts up than comment every time it acts as expected. After all, when things work as they should people tend not to take an interest.

 

[mongo]

The problem is that everyone who uses AP (including the people who misuse it) seems to think that they are using it wisely (or at least in a manner that is appropriate).

I'm with you there. People tend to equate no negative consequences with appropriate action (for many facets of life)

This is partially because Tesla barely defines what appropriate use is,

Disagree, the manual clearly states the limitations and that the driver is in control and needs to remain aware. The next statement regarding official position also speak against not indicating appropriate use.

and then seems to wink-wink at some of those statements (as when Elon demonstrates AP with his hands off the wheel, despite Tesla officially defining this as inappropriate use).

Agree, not setting a great example. However, Elon doesn't drive stock SW and is more aware of what the real limits of the version he has are.

When crashes occur, they tend to blame themselves for not jumping in fast enough to override AP, rather than questioning whether they have been using AP in an appropriate manner.

I think we agree here? If the driver is jumping in, that means there were not engaged with the driving task previously, which is inappropriate usage.

And there have been an incredible amount of reports of drivers having to quickly react to correct AP's mistakes.

Which indicates drivers are not using it appropriately. When I drove an X, I felt it too too little force to disengage (but I'm beyond 95 percentile in size). If the driver applies a disagreeing force to AP, it shuts off. If AP is allowed to make a mistake, the driver already has...

 

[Az_Rael]

I agree with the discussion of ability to maintain attention with reduced workload. However, only in the case of a sudden AP steering change do I see #2 being a separate case, even that requires the driver to not be firmly holding the wheel.

I think the first time someone gets rear ended due to a phantom braking event in AP, that will be an interesting discussion. I know I don't drive with my feet hovering over the pedals when using TACC or AP.

 

[Economite]

Holding onto the steering wheel and looking out the windscreen is hardly a difficult point to get across. Appropriate use of the system is really that simple.

Part of using the system is deciding when to activate AP and when to turn it off, as well as what follow setting to use. It's the guidance on that which is sadly ambiguous (or widely ignored). Folks kind of have to guess (or try to learn via experimentation) what AP will and won't be able to handle.

 

[Peteski]

I agree that TACC is good at preventing sloppy speed control by a driver.

My guess, however, is that drivers who are normally inclined to tailgate when they aren't using TACC/AP tend to be setting the follow distance to 1 or 2 when they do use TACC/AP (even at high speeds). Thus, TACC in its current form might mainly be reducing unintentional tailgating (by folks who mess up at modulating speed/reacting to behaviors of lead cars) rather than intentional tailgating (by folks who are either really impatient, convinced that people will "cut in front of them" if they leave any space, just don't have a good notion of what a safe following distance is). For TACC to dramatically reduce this second type of tailgating (which I suspect is the most common), it would need to have the kind of limitation you suggest in the text I bolded in your quote.

What I've noticed from my own behaviour using TACC is that I tend to be a lot more chilled out when following other cars and don't pressure them to get out of my way like I used to. I find that TACC takes away much of the impatience you get when manually following a slower car. I've heard other people say this too, so not just me. Tailgating is as much a habit for many people rather than a serious attempt to overtake - like when there are 100 other cars ahead all doing the same so there is nowhere to go anyway! Yet they still tailgate as if the car in front is somehow holding them up. TACC does need to sort out its own sensible following distance strategy though as 1 is way too close at high speed. That would be very easy to implement based on speed as I suggested.