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

[Bubslug]

Compare your conclusion (AP induced crash) against all the other instances of the barrier being hit. There are no (lasting) swerve or braking marks from them (else how could we distinguish from this accident). Given the majority (all?) previous collisions were non-Tesla vehicles, placing the blame on a feature of a Tesla seems like a reach.

That's not the way to analyze this. For example, if another vehicle hit the barrier at 11PM, while being driven home from a wedding, the probability the driver was impaired is much higher than this accident. So the reason for another incident may not be reasonably applicable in this situation.

Each situation has it's own probability path to the most likely scenario. You don't compare scenarios from different accidents, you need to compare the different scenarios possible for this incident to conclude which one is the most likely. Certainly you can learn from previous incidents, i.e. what is the probability a vehicle is forced into gore point barriers by another vehicle versus safe transits past the same barrier (no one probably knows this for sure, but you can estimate). In this case what is the likelihood this vehicle was equipped with Autopilot? (100% ?) If so what is the probability it was on? (45% ?) etc. etc. That's the way objective analysis works.

 

[buttershrimp]

I know that when im in autopilot, the automatic lane change is smart enough to know the shoulder vs a lane. Can someone check if auto lane change is available from the HOV lane into that gore zone?

I agree with @verygreen mostly because he is clearly smarter then me. Tesla might know that Autopilot was engaged at some point, but they wouldnt necessarily know at the moment of the accident. Say if the driver disengaged auto pilot to change lanes into the gore zone.. hence my question above. The only way to change lanes would be to disengage auto pilot if auto lane change is not available.

Looking at everything presented here, I could see this scenario:

1) Driver was in the HOV lane and realized he needed to get over.
2) He disengaged Autopilot and moved into the gore zone looking for a window to merge onto the 101, and maybe even accelerating to merge
3) He was not able to find a window and miss judged the distance to the median, maybe in part because he was not the only car trying to get out of the HOV lane. As others have noticed, this is a somewhat confusing lane and even though this person probably knew this stretch of road very well, its possible that general chaos with other drivers around that point in the road could have contribute to him miss judging the distance.

My guess and this is purely speculation. Autopilot was either engaged and he wasnt paying attention and drove directly into the median. I think this is a much lower likely hood given that this seems to be the persons commute so he would have know the issues at this location and would not have made such a huge mistake of getting into the gore zone and enabling Autopilot, assuming the above that auto lane change would not work here.

I find that one of the two scenarios is the most probable. Its very sad, but being an owner and frequent autopilot user, I to have to force myself to be more disciplined. The worst part is that I am often most relaxed where I am most familiar with the roads and this is actually very bad, because stuff happens all the time and just this stuff that Autopilot is not good at. Its nearly perfect when the situation is very stable but any chaos can cause failure and fast.

Totally agreem to me, this is the most likely scenario in some shape or form.

 

[rsum]

Tesla (TSLA) stock loses over $4 billion in value in one day over a series of bad news

 

[MorrisonHiker]

My guess is that AP strongly defaults towards "follow the car in front of you" logic whenever it perceives road markings as the least bit ambiguous. Thus, I think that AP's behavior when there is a car in front of it (even at some distance) is not particularly indicative of how AP will behave in the same location when there is no car to follow.

For the past several months, we've noticed that AP2 would follow the lead car through difficult intersections. When we updated to 2018.10.4 recently, we noticed that our cars no longer follow the lead car through those same difficult intersections.

 

[mongo]

I wonder what speed the was going at the time of the impact?

CHP report said highway speed.

That's not the way to analyze this. For example, if another vehicle hit the barrier at 11PM, while being driven home from a wedding, the probability the driver was impaired is much higher than this accident. So the reason for another incident may not be reasonably applicable in this situation.

Each situation has it's own probability path to the most likely scenario. You don't compare scenarios from different accidents, you need to compare the different scenarios possible for this incident to conclude which one is the most likely. Certainly you can learn from previous incidents, i.e. what is the probability a vehicle is forced into gore point barriers by another vehicle versus safe transits past the same barrier (no one probably knows this for sure, but you can estimate). In this case what is the likelihood this vehicle was equipped with Autopilot? (100% ?) If so what is the probability it was on? (45% ?) etc. etc. That's the way objective analysis works.

People drive into the barrier, a person could drive a Tesla into the barrier just as easily (maybe more so given lack of engine noise as a speed cue and the acceleration)

Correct me if I'm wrong on your train of thought:
Tesla X -> likely has AP -> likely using AP -> likely AP was active and caused crash.

I'll counter with employee on commute->likely knows route- > likely has AP in same mode as always. Either off because it doesn't work, or on because it does. If it historically didn't work, it would be disabled and not a factor (or closely monitored). If it historically did work, then that supports the case that if it were on, if would pass the section safely. If there is an issue with AP, we should hear about it soon for other commuters of that area.

As @MorrisonHiker just posted, the new version acts differently, so there could be a confounding issue there. However, the driver on their normal commute would likely be aware of the problem spots on the commute in general and pay attention to then. Just like they would likely be aware of which lane to be in. To let the car drive you down a non-lane into a barrier would indicate either a medical emergency or other distraction.

If you are supporting your conclusion on the lack of tire marks, are you comparing that to the level of marks from other incidences? If 90% of crashes do not leave tire marks, does the lack of tire marks indicate anything?
How does this fit into the CHP lost control report (assuming that is accurate)?

 

[buttershrimp]

For the past several months, we've noticed that AP2 would follow the lead car through difficult intersections. When we updated to 2018.10.4 recently, we noticed that our cars no longer follow the lead car through those same difficult intersections.

I’ve been thinking about this scenario, and the more I look at the available data, the more it appears to me to be a driver disengagement of autopilot prior to the impact. In all of my video, I have not been able to trick my car into going straight at a demarcation point.... I have been able to confuse it upon merging with lane change.... see the end of my toll booth willie video.... It shows what happens when you try to signal you are merging and will get confused. But the opposite scenario, let’s say he initiated a lane change prior to the split point. I find this highly unlikely.... when you are lane changing with the auto feature, you are at your most attentive. Lettting it lane change you into a bad situation is the least likely scenario I can think of.... I could imagine someone lane changing into a vehicle in the blind spot based on limitations of ultrasonics, but not into a barricade... .

 

[buttershrimp]

That's not the way to analyze this. For example, if another vehicle hit the barrier at 11PM, while being driven home from a wedding, the probability the driver was impaired is much higher than this accident. So the reason for another incident may not be reasonably applicable in this situation.

Each situation has it's own probability path to the most likely scenario. You don't compare scenarios from different accidents, you need to compare the different scenarios possible for this incident to conclude which one is the most likely. Certainly you can learn from previous incidents, i.e. what is the probability a vehicle is forced into gore point barriers by another vehicle versus safe transits past the same barrier (no one probably knows this for sure, but you can estimate). In this case what is the likelihood this vehicle was equipped with Autopilot? (100% ?) If so what is the probability it was on? (45% ?) etc. etc. That's the way objective analysis works.

Yeah, that’s some hard logic there. We can’t rule out autopilot, but I find the idea that it drove him into the barrier extremely unlikely. I think if the logs are recovered, it will provide useful information on how driver’s interact with autopilot.... and how they behave prior to and upon disengagement. I find that I am probably at most risk of an error when I intentionally disengage autopilot to speed up and get around traffic.... you get used to a certain speed and the autopilot system is intentionally not nimble enough to lane change smoothly around things... in fact, even TACC is likely to brake too much if you are navigating around slower traffic.

 

[Bubslug]

CHP report said highway speed.
.....To let the car drive you down a non-lane into a barrier would indicate either a medical emergency or other distraction....

Fine. The driver was a 38 year old male. Calculate the probability he had a medical emergency while approaching the barrier vs. Autopilot was on and failed to resolve the lane properly. I'm going to say the probability of the former is lower than the latter since the light and lane markings were suboptimal for AP, and I know AP has failed here previously as noted by Alset72. If you told me the 38 year old male had hypertension, I might adjust the probabilities, but come to the same conclusion. If you told me the 38 year old male had just had bypass surgery, I might reverse the conclusion. Probabilities change as more facts emerge.

 

[mongo]

Fine. The driver was a 38 year old male. Calculate the probability he had a medical emergency while approaching the barrier vs. Autopilot was on and failed to resolve the lane properly. I'm going to say the probability of the former is lower than the latter since the light and lane markings were suboptimal for AP, and I know AP has failed here previously as noted by Alset72. If you told me the 38 year old male had hypertension, I might adjust the probabilities, but come to the same conclusion. If you told me the 38 year old male had just had bypass surgery, I might reverse the conclusion. Probabilities change as more facts emerge.

I said

To let the car drive you down a non-lane into a barrier would indicate either a medical emergency or other distraction.

, (don't focus on the M.E. part)

Thank you for the name reference, it let my find this post with the report of trouble there. @Alset72 do you have the newer version of AP SW, and do you travel this way normally?

 

[brkaus]

You misunderstand my logic. I'm looking for the simplest explanation that matchs all the fact as we know them at this point and invokes the least number of anomalous conditions. So if you consider the most probable case, that the driver was not risk taker, or impaired in any way, and was not interacting with other risk taking or impaired drivers, or traffic movement anomalies of any kind, you're left with an explanation of how a competent driver could run into a fixed barrier, on a road he was familiar with, in normal flowing traffic.

Yes, he could have realized he was in the wrong lane at the last moment, or could have been forced over by another driver's error, but why then apparently no braking, or tire swerve marks? Given we I know so far, I've limited my scenario to the simplest one possible, which assumes only one anomalous condition, poor light conditions and lane marking quality for lane resolving by AP, which I know to be true.

By listing the assumptions required for any scenario for how this could of happen, and assigning probabilities to each assumption, you can mathematically compare any 2 possibilities. Right now I think the highest probability is that he was in lane #2, but if you told me it was only his second day at work at the Apple offices in Sunnyvale, and previously he had worked in Cupertino, I would say it makes more sense he was in the #1 lane, out of habit, since he was used to forking to 85 at that point. In that case another scenario might become more plausible (mathematically).

Drifting while driving is also possible. I see folks drift out of lanes all the time. Reaching for something, spill something, day dreaming, phone call.

A few weeks ago here a pedestrian was killed walking on the shoulder of a road - 8-12 feet from the traffic lane. That particular driver was not in a Tesla. Similar story with a bike two years back (that driver just got out of prison). Heck, a Tesla with AP might have prevented those deaths.

Yes, of course, they could have been on AP. But it sounds like a stretch to say that is "most likely".

 

[jeffro01]

Some of the comments here about AP are not only ridiculously speculative, but grossly irresponsible as well. There is NO WAY AP drove the car into an immovable barrier like that, no freaking way. I don't know what happened, I don't profess to have been there and seen it happen, I don't have the logs, but neither do ANY of you... Yet far too many posters here are absolutely quick to jump on AP when these kinds of accidents happen with all other makes/models of cars... It's just ridiculous...

Why are so many of you so quick to jump on AP? There are so many other more plausible explanations for how something like this could happen... I used to ride a motorcycle and lane split all the time in the Bay Area and I can't even begin to describe to you all the different kinds of distracted driving I saw from every kind of driver in every kind of car.

Jeff

 

[jeffro01]

I see many Tesla owners relying too much on AP - way too much. I think some of the AP users have become oblivious of the commute around them.

It is tragic and I am sorry for the family's loss. That is a horribly marked poorly designed exit.

I have to assume you're camera is good enough to see the dash indicator icons otherwise you really have no way of knowing if anyone has AP engaged or if they are just a horrible driver...

Jeff

 

[Haddock]

I had a thought - do we think this was an AP2 car? The reason I wonder is because the force required to override the wheel in AP2 (at least before 10.4) is a lot higher than with AP1. In my 3 I had a couple of scary disengages at 70mph where I corrected the car with the wheel, AP2 wrestled with me for a bit, then let go, causing me to jerk the wheel. It was scary enough that I stopped using the wheel to disengage because I was worried about losing control at freeway speeds.

There was also this fellow who gave an account of losing control of his AP2 car after using the steering wheel to disengage at 80mph: Scary Experience with AP2.0 Autosteer on the Highway (OP in this thread thought it was a malfunction, but it was probably more likely to be steering overcorrections after the AP disengagement). A couple others in that thread also relayed almost losing control of their cars upon wheel disengagement.

I could see a scenario where AP2 decided to act up right at this bad junction, the X driver disengaged with the wheel, and then due to the high release force, the wheel was jerked at freeway speeds after AP let go. Now the fellow is in a scenario ripe for steering overcorrections and subsequent loss of control of his vehicle.

I think this is very likely and fits with the "driver lost control" description from the police. But my guess is that the AP2 didn't act up, it was just sticking to its lane, and the driver suddenly realized he was leaving the 101, when he wanted to continue. He disengages by turning the wheel, and according to the post linked above it is easy to accidentally steer too suddenly when doing this, especially at highway speeds.

If you read the thread above there are a lot of people saying "this is why I always disengage with the brake, not with the steering wheel". But if the reason for disengaging is that you want a quick lane change it is very natural to disengage by turning the wheel. Everyone should read the post linked above, it perfectly describes a scenario that could have lead to an accident like this one.

 

[bro1999]

Some of the comments here about AP are not only ridiculously speculative, but grossly irresponsible as well. There is NO WAY AP drove the car into an immovable barrier like that, no freaking way. I don't know what happened, I don't profess to have been there and seen it happen, I don't have the logs, but neither do ANY of you... Yet far too many posters here are absolutely quick to jump on AP when these kinds of accidents happen with all other makes/models of cars... It's just ridiculous...

Why are so many of you so quick to jump on AP? There are so many other more plausible explanations for how something like this could happen... I used to ride a motorcycle and lane split all the time in the Bay Area and I can't even begin to describe to you all the different kinds of distracted driving I saw from every kind of driver in every kind of car.

Jeff

You jumping to the conclusion AP WASN'T a factor is just as far a leap.

 

[NeverFollow]

Yeah, that’s some hard logic there. We can’t rule out autopilot,
but I find the idea that it drove him into the barrier extremely unlikely.

I think if the logs are recovered, it will provide useful information on how driver’s interact with autopilot....
and how they behave prior to and upon disengagement.

I find that I am probably at most risk of an error when I intentionally disengage autopilot to speed up and get around traffic....
you get used to a certain speed and the autopilot system is intentionally not nimble enough to lane change smoothly around things...

in fact, even TACC is likely to brake too much if you are navigating around slower traffic.

After reading all those posting, I think it would be crucial to determine:
- If the car has been directed by the AP in the grove area by mistake, thus hitting the middle section wall.
- or if the driver decided to use the left lane to pass the congested traffic and then tried merging on the right lane?
But while looking at the right mirror to check the traffic, he didn't realised that he was going to hit the middle section wall.

 

[UtopianThought]

Work near the flyover in question. Own an S without Autopilot and a 3 with.

Heard about the accident shortly after it occurred as co-workers who were delayed by accident mentioned it. Early pictures seem to show more of the car intact, but there is certainly a frightening amount of damage. Post-crash/Post-fire pictures look worse as they cut away more of the car to fight the fire.

https://insideevs.com/wp-content/uploads/2018/03/Capture-1.jpg
Ian Cull on Twitter

The fact that it was significantly worse than the S that hit the fire truck tells you something about the speed.

It sounds like the driver was cleared before the fire, but not able to survive the severe undampened impact. Very sorry to hear the driver died.

Any car hitting that minimized barrier would have been damaged. Why the state hadn't replaced the barrier with barrels of sand/water until it could be properly repaired is questionable at best.

As to cause of the accident, I definitely think we will have to wait for more information. With the fire, it is likely any record keeping device was damaged, perhaps irrecoverably. I too think it unlikely that autopilot is directly at fault, but certainly, we will have to wait and see.

Perhaps Tesla would consider adding 2 forensic devices, 1 to front, 1 to rear to help with situations where 1 half of the car is damaged badly. Minimal cost to maintain safety rep.

As the safest car ever tested, Tesla has a bit of a target on it.

 

[jeffro01]

You jumping to the conclusion AP WASN'T a factor is just as far a leap.

Um... No. Not even close. Nice try though. I say again, there is NO WAY AP drove that Model X into that wall/barrier/whatever. It's simply not how AP behaves/acts/programmed to function.

Jeff

 

[mkspeedr]

I have to assume you're camera is good enough to see the dash indicator icons otherwise you really have no way of knowing if anyone has AP engaged or if they are just a horrible driver...

Jeff

It is tough to tell sometimes between bad and autopilot but it is actually easier than you think.

 

[jeffro01]

It is tough to tell sometimes between bad and autopilot but it is actually easier than you think.

I don't mean to sound as rude as this is going to sound but... I find it awfully hard to believe you can tell a Tesla is on AP without being able to see the dash indicator light... That seems to be an awfully subjective leap you're taking... I haven't lane split in a year or so as I sold my Harley awhile back but the only time I could tell a Tesla was on AP was when you'd come up on one and it wouldn't move over. Perhaps I'm not as perceptive as you are.

Jeff

 

[tropical1]

Even if autopilot is on, wouldn’t it still be incumbent on driver to be paying attention. These aren’t fsd cars. It’s like having a wreck with a car on cruise control right? Also, Any car would,probably burn in this type of impact. I am glad Tesla and NTSB are trying to figure out what happened to prevent in future. But don’t fatal accidents and vehicle fires occur every day with other types of cars. I guess I am curious why Tesla always get singled for every major accident as if major accidents don’t happen to other manufacturers cars all the time. There is also the flip side where folks have walked away unscathed from some major accidents in a Tesla.