Slashdot Mirror
Tesla Says Autopilot Was Engaged During Fatal Model X Crash (theverge.com)
An anonymous reader quotes a report from The Verge: Tesla says Autopilot was engaged at the time of a deadly Model X crash that occurred March 23rd in Mountain View, California. The company posted a statement online late Friday, after local news reported that the victim had made several complaints to Tesla about the vehicle's Autopilot technology prior to the crash in which he died. After recovering the logs from the crash site, Tesla acknowledged that Autopilot was on, with the adaptive cruise control follow distance set to a minimum. The company also said that the driver, identified as Apple engineer Wei "Walter" Huang, had his hands off the steering wheel and was not responding to warnings to re-take control. Tesla said in a statement: "The driver had received several visual and one audible hands-on warning earlier in the drive and the driver's hands were not detected on the wheel for six seconds prior to the collision. The driver had about five seconds and 150 meters of unobstructed view of the concrete divider with the crushed crash attenuator, but the vehicle logs show that no action was taken."
According to Mercury News, the driver of the car was headed southbound on California's Route 101 when his Model X crashed headfirst into the safety barrier section of a divider that separates the carpool lane from the off-ramp to the left. "The front end of his SUV was ripped apart, the vehicle caught fire, and two other cars crashed into the rear end. [The driver] was removed from the vehicle by rescuers and brought to Stanford Hospital, where he died from injuries sustained in the crash."
According to Mercury News, the driver of the car was headed southbound on California's Route 101 when his Model X crashed headfirst into the safety barrier section of a divider that separates the carpool lane from the off-ramp to the left. "The front end of his SUV was ripped apart, the vehicle caught fire, and two other cars crashed into the rear end. [The driver] was removed from the vehicle by rescuers and brought to Stanford Hospital, where he died from injuries sustained in the crash."
Tesla said in a previous article that autopilot had done this route 85,000 times. I guess repetition doesn't necessarily mean success here. Big surprise.
Laws are rules for the court, but merely a bottom bar to hit for life. Think beyond laws in your actions always.
He apparently had plenty of money; he was driving a Tesla. He was an engineer, so he was educated.
It amazes me that often people don't recognize that driving a car is a potentially extremely dangerous activity. 100% attention is required at all times, particularly since other drivers often do things they shouldn't do.
So you design a car that can safely drive itself in traffic, can track whether the driver is actively using the controls and knows that for six seconds the driver hasn't been using them while driving at speeds the car can't protect them through a crash.
And you didn't design in, "Slow the fuck down because nobody is in control of the vehicle"?
The driver had about five seconds and 150 meters of unobstructed view of the concrete divider with the crushed crash attenuator, but the vehicle logs show that no action was taken.
Having narrowly avoided two separate impending collisions while driving due to insects, one hornet loose in the cab & one bee in the eye through an open window, I have a macabre fascination with the last few seconds in a vehicle before the collision the takes the life of the human witness(es).
Sure, we live in an age of unrivaled electronic distractions, but there have always been ample incentive to pick the wrong five seconds to look away from the road. Outside of law enforcement, we'd never see the video, even if it did exist... but the new tech vehicles are getting makes the 'fly on the wall' view ever more likely.
Happiness in intelligent people is the rarest thing I know.
Ernest Hemingway
Huang reportedly complained that the car’s Autopilot option kept veering the car toward the same barrier on Highway 101, near Mountain View, into which he crashed the car last Friday.
If you've noticed unsafe behaviour and have made complaints about it, why the fuck would you keep using it?
Not surprising that an Apple engineer has no common sense.
And the only common sense thing for Tesla to do is to disable the damn thing. People are too stupid to be trusted with anything.
Those who do not learn from commit history are doomed to regress it.
Actually tesla said the driver ignored the warning earlier in the drive. It could have been an hour before.
Tesla autopilot warns you to take control every minute or so regardless of whether it is "confused" or not. If you don't, after sufficient warnings, it stops the car.
There is no information in the Tesla statement that isn't true of many many tesla autopilot journeys.
It looks to me, as a Tesla owner myself, that autopilot did, indeed, drive him into the barrier and that we have a reminder that every time a driver looks away when on autopilot, they gamble with their life.
If the car had "several visual and one audible hands-on warning" then maybe the autopilot should bring the car to a halt. However I suspect that what happened was more complicated and that we do not know the full story.
On average there are 700 deaths on US roads EVERY week and two more should not be national news. With safer cars this number has been dropping in the last decade but this is news is actually about computer AI making a choice, or by not making a choice, killing two people. It may not be full AI, but it is still a computer program in control. Two people died because of a computer program. With both accidents the "self-driving" AI program should have saved these people. Both times the person behind the wheel should have been able to avoid or lessen the collision if they were actually driving. We don't hear much about AI driving success in avoiding crashes just like we don't hear about planes that land safely. We only hear about failures. These features will get better with time and debugging (meaning more failures to come). Just as early commercial planes had their problems so does AI self-driving. For now flying is safer than driving no matter who is in control of the car (0 commercial aviation deaths for 2017 in US) and improved technology can only help our chances of making it home safely even if it makes the wrong choice occasionally (well, on average).
Better yet, another story is saying that the man has noticed autopilot having a problem at this particular stretch of road In the past - if you've seen it trying to send you into a k-rail at that bit a few times, what the fuck are you doing letting it drive on that bit, and not paying attention when it tells you to? Did he fall asleep or something?
I sure as shit wouldn't be using it there if I've gone so far as to take it to the service center to have them look at it for trying to drive into exactly that barrier in the past...
Man: Hey Doc, when I shove my head up my ass, I have problems breathing!
Doctor: then pull your head out of your ass, and stop shoving it up there!
Slashdot still doesnâ(TM)t support Unicode after it was added to the HTML standard in 1997.
as the car is driving down route 101, do you honestly expect the computer onboard to STOP the car, right there? in many cases, there is not even a pullover (lay-by) lane.
I'd like to know what you propose, when the computer says 'I need you to do something, I'm not sure what I should do, myself' and the human ignores it for too long. stopping is NOT always the right thing! the correct answer is 'it depends'.
--
"It is now safe to switch off your computer."
... or maybe you're not holding the wheel tight enough.
Well, he did work at Apple. Not holding it right is a distinct possibility...
Reading and re-reading the quote from Tesla, I see I was mislead:
The driver had received several visual and one audible hands-on warning earlier in the drive and the driver's hands were not detected on the wheel for six seconds prior to the collision. The driver had about five seconds and 150 meters of unobstructed view of the concrete divider with the crushed crash attenuator, but the vehicle logs show that no action was taken."
This does not mean that the warning fired during the 6-seconds prior to the collision. It wasn't telling him a collision was imminent. It says that "earlier in the drive" it warned him. So the warning could have been 45 minutes prior. Also, it sounds like the autopilot warning happens any time the user takes their hands off the wheel, not just when it needs help. It might be that autopilot drivers have a tendency to ignore the warning, like a dialog box that comes up so often people just click "OK" to it.
I begin to think that a semi-autopilot is a bad idea. If it is not reliable enough that a person can take their hands off the wheel, and they still must pay full attention to the road in case it makes a mistake, then they might as well drive? It is very hard to pay attention to something you aren't actively involved in. Airline pilots and lifeguards and factory quality inspectors know this. Those industries have specific policies and practices designed to keep people engaged and aware.
"...the victim had made several complaints to Tesla about the vehicle's Autopilot technology prior to the crash...
So the guy who has complained not once, but repeatedly, that his car's autopilot is inadequate engages it and completely ignores what it's doing.
This takes a special kind of stupid.
Somehow I found the strength to ignore the low-hanging fruit: that this potential Darwin Award winner was an Apple engineer.
I've calculated my velocity with such exquisite precision that I have no idea where I am.
How many people are going to die before we are there? 3.22 trillion miles driven and humans have 16,000 accidents a month. They drive over 550K miles without getting in an accident in all weather and road conditions. Let's give Waymo the benefit of the doubt and say they achieve 7000 miles per interaction, that's still only 1.8% the safety of a human *in ideal conditions*; how many injuries and deaths have there been already to get to this point?
People understand the risk of driving and they drive. What people don't understand is how long it will take to make these cars workable.
Laws are rules for the court, but merely a bottom bar to hit for life. Think beyond laws in your actions always.
That story was incorrect. He had taken his car in because of problems with the navigation system, unrelated to autopilot. How that morphed into “autopilot problems at that particular stretch” is a classic example of the telephone game.
If there's a general consensus in the English speaking world that autopilot is synonymous with autonomous, I agree the name was poorly chosen. That's not my impression of the general understanding of the word – we expect a pilot to remain in the cockpit and alert when the plane we're on is on autopilot, after all – but if the data shows otherwise then I'd be the first to argue for the feature to have its name changed.
Uber plowed into a pedestrian at full speed on a well lit road, whereas this driver ignored six seconds of warnings to take control.
Oh god, imagine how much you could freak people out by replacing the standard autopilot-engagement sound files with "KILL ALL HUMANS" ;)
"99 dead duelists of Dios on the wall. 99 dead duelists of Dios! Take one's ring, pass it around..."
Maybe he installed a big notch in the middle of his windshield, and he couldn't see the barrier through it ;)
"99 dead duelists of Dios on the wall. 99 dead duelists of Dios! Take one's ring, pass it around..."
Well, I did find an NHTSA report from January of 2017 (after a previous fatality linked to Autopilot use). They found a 40% decrease in crashes among Tesla drivers after Autopilot Autosteer became available. Not super-definitive, but interesting.
5.4 Crash rates
. ODI analyzed mileage and airbag deployment data supplied by Tesla for all MY 2014 through 2016 Model S and 2016 Model X vehicles equipped with the Autopilot Technology Package, either installed in the vehicle when sold or through an OTA update, to calculate crash rates by miles travelled prior to (21) and after Autopilot installation. (22)
Figure 11 shows the rates calculated by ODI for airbag deployment crashes in the subject Tesla vehicles before and after Autosteer installation. The data show that the Tesla vehicles crash rate dropped by almost 40 percent after Autosteer installation.
page 10 on:
https://www.scribd.com/documen...
The driver had complained about trouble with his car to Tesla before the fatal crash:
"Walter Huang's family tells Dan Noyes he took his Tesla to the dealer, complaining that -- on multiple occasions -- the auto-pilot veered toward that same barrier -- the one his Model X hit on Friday when he died."
If his Tesla has a history of doing something reckless, why would he re-enable it? Why would he also not have your hands on the wheel? Why didn't Tesla analyzed the data in his car when he reported this to see what was going on? Seems like it would have been a pretty simple check: Did the car attempt to steer the car towards the barrier or not?
Right. They'll just pop down to their gigafactories after a quick recharge on their supercharging networks and build vehicles on lines and with components that are the result of billions of dollars of investment over years. Easy as pie!
"99 dead duelists of Dios on the wall. 99 dead duelists of Dios! Take one's ring, pass it around..."
I apologize for the length of this post. It is as much about trying to state my thoughts clearly as it is about participating in a conversation.
For all of the following I will assume that the 40% reduction in crashes attributed to Autopilot by NHTSA is real.
I agree that off-highway driving is where you see the most crashes. Let's break all driving into two types: Autopilotable (P), and Manual-required (M). P includes all miles driven that allow the driver to use Autopilot, M includes the rest.
Let's arbitrarily set the ratio of (P miles driven / total miles driven) to 50%, leaving 50% for M. No, these are not the real numbers, but we can see which way adjustment moves the effectiveness of the Autopilot system. Let's further assume that crashes are evenly distributed between P miles and M miles (which neither of us believes). Let's call this (unrealistic) scenario the Baseline.
Now let's grab a sample of the total miles driven that contains exactly 1000 crashes. 500 of them are in P miles and 500 of them are in M miles. If Autopilot can prevent 40% of ALL of those crashes, then it is preventing 400 crashes. It can't prevent crashes in the M miles at all (it's not in use, by definition), so it must prevent 400 of the crashes in the P miles, of which there are only 500. Preventing 400 out of 500 crashes is preventing 80% of P mile crashes.
But we don't believe that crashes are evenly distributed. We know they are skewed toward the bad conditions and the off-highway miles, the M miles. So, let's move the line a little. We'll say that 55% of the accidents occur in the M miles and only 45% in the P miles.
In our 1000 crash sample, we have 450 in the P miles, 550 in the M miles. 40% of all crashes is still 400 crashes. If Autopilot can stop 400 out of the 450 P mile crashes it is stopping 88.9% of accidents in the P miles, up from 80% in the Baseline scenario. The more the crashes concentrate into the M miles, the better Autopilot has to do in the P miles to make that 40% improvement. This is the reasoning behind my second assertion above:
This is especially true if accidents become more common as driving conditions deteriorate, which I take as a given.
But what about my crazy assumption that P miles and M miles were equal? Seems unlikely to me, but I truly don't know. So, what happens if we move it some? Let's say that P miles account for 60% and M miles only account for 40%. What does this do to our Baseline?
1000 evenly distributed crashes gives us 600 in the P miles and 400 in the M miles. Autopilot's 40% reduction is 400 crashes out of 600 in the P miles (again, these are the only miles where drivers are using Autopilot). In this case Autopilot can apparently prevent 66% of crashes in the P miles. Not bad, but less than our (admittedly arbitrary) Baseline. As P miles become more prevalent, it is easier for Autopilot to make that 40% reduction. This is the reason for my third assertion above:
I guess it becomes less true if Teslas are generally driven less in those deteriorating conditions.
Because I absolutely believe that crashes are concentrated in the M miles, I suspect that P miles are way more common than one would intuitively guess. Or alternatively, it could be that the 40% reduction was not caused by the Autopilot rollout at all, but I am taking the NHTSA at their word. I don't usually see them as knuckleheads, but I have to acknowledge the possibility.
If you read all of that, thanks. I appreciate your viewpoint and I am trying to approach this with an open mind.