Slashdot Mirror
There's Growing Evidence Tesla's Autopilot Handles Lane Dividers Poorly (arstechnica.com)
An anonymous reader writes: Within the past week, two Tesla crashes have been reported while Autopilot was engaged, and both involved a Tesla vehicle slamming into a highway divider. One of the crashes resulted in the death of Walter Huang, a Tesla customer with a Model X. The other crash resulted in minor injuries to the driver, thanks largely to a working highway safety barrier in front of the concrete divider. Ars Technica reports on the growing evidence that Tesla's Autopilot handles lane dividers poorly: "The September crash isn't the only evidence that has emerged that Tesla's Autopilot feature doesn't deal well with highway lane dividers. At least two people have uploaded videos to YouTube showing their Tesla vehicles steering toward concrete barriers. One driver grabbed the wheel to prevent a collision, while the other slammed on the brakes. Tesla argues that this issue doesn't necessarily mean that Autopilot is unsafe. 'Autopilot is intended for use only with a fully attentive driver,' a Tesla spokesperson told KGO-TV. Tesla argues that Autopilot can't prevent all accidents but that it makes accidents less likely. There's some data to back this up. A 2017 study by the National Highway Transportation Safety Administration (NHTSA) found that the rate of accidents dropped by 40 percent after the introduction of Autopilot. And Tesla argues that Autopilot-equipped Tesla cars have gone 320 million miles per fatality, much better than the 86 million miles for the average car. These figures don't necessarily settle the debate. That NHTSA figure doesn't break down the severity of crashes -- it's possible that Autopilot prevents relatively minor crashes but is less effective at preventing the most serious crashes. And as some Ars commenters have pointed out, luxury cars generally have fewer fatalities than the average vehicle. So it's possible that Tesla cars' low crash rates have more to do with its wealthy customer base than its Autopilot technology. What we can say, at a minimum, is that there's little evidence that Autopilot makes Tesla drivers less safe. And we can expect Tesla to steadily improve the car's capabilities over time."
You don't want to know how many accidents there were in cars with autopilot, that doesn't matter. What you want to know is miles per accident *with autopilot engaged.* Using the other number is highly misleading.
"First they came for the slanderers and i said nothing."
What is the point to an autopilot if I have to be fully attentive and ready to take over? I would rather just drive then worry about missing something. This and the are still have to much uncertainty to waist my time with it. Also dammit if i'm going to die in a car, I want it to be my fault and freaking awesome. "He almost made it, if it hadn't been for that ...."
The intentionally misnamed "autopilot" may reduce the likelihood of wandering out of your well-marked lane in clear conditions at highways speed but every once in a while it'll drive you right into an obstacle. Reminds me of those "push this button and ten people with terminal cancer get cured but two other random people die from a meteor strike" questions taught in philosophy classes with the intent of humbling people who might otherwise believe they can quantify their way through every obstacle.
It should be noted that the vehicle that, just a week earlier, destroyed the safety barrier that would likely have saved the Tesla Model X driver was not on autopilot. I've had close calls around entrances to HOV lanes myself. They seem to be designed for throughput over safety, often with the lane that continues actually having to turn a bit to not hit the divider instead of having the HOV folks be required to move into a long entry lane separating from the regular lane completely long before the divider.
Unless Travis goes to prison for life he's still doing better than Trump will.
Autopilot really is the wrong name for this technology. It is not what people think of when they hear autopilot.
It seems even worse than regular driving if you have to continuously be on the look out to prevent the steering wheel suddenly sending you into a wall.
Maybe Tesla should focus on automatic braking, parallel parking, and things like that until using their Autopilot is no longer the same as playing the Russian roulette. At some point, these accidents will damage their reputation badly...
'Autopilot is intended for use only with a fully attentive driver'
Your stupid marketing decision is killing people. Seriously.
Again, why is anyone surprised? Fully autonomous tech will likely never be a thing. True AI will never be a thing. You can't break the laws of math, and math is all an algorithm will ever know or be able to process. Data is not the issue, the flawed premise is. How much more money will be flushed down the toilet in pusuit of this hype train?
You're acting stupid again.
what about motorcycles? I know BMW's Traffic Jam Assistant doesn't do well with that since I had a BMW 750 rear-end me at about 10 MPH. The guy that hit me says it usually does a great job of going the correct speed in 0-30 MPH traffic here on I-5 in Seattle. I think it didn't see me, but instead saw the dump truck in front of me and then tried to drive through me.
Its every god damn vehicle with any sort of lane assist feature
Recalculate the miles per fatality after you remove all the other cars without the other safety systems the Tesla has.
It would be interesting to see the miles per fatality when the cars counted all have stability control, crumple zones, a dozen or so airbags, side impact beams, etc.
Unless you're comparing 5 star crash rating cars with other 5 star crash rating cars, it's not even a remotely fair comparison of Autopilot.
"Tesla cars have gone 320 million miles per fatality, much better than the 86 million miles for the average car. These figures don't necessarily settle the debate. That NHTSA figure doesn't break down the severity of crashes -- it's possible that Autopilot prevents relatively minor crashes but is less effective at preventing the most serious crashes."
I find the skewed submitter's view of "minor crashes" a bit odd.
The comparison is for fatalities per mile. I'd have a hard time expecting that there would be a substantial percentage of accidents with fatalities being very high.
That's not to say that there aren't issues here....and it might just be semantics.....but I hope human life has enough of a value so as to not consider a fatal accident as minor.
Tesla: Stop calling it "Auto-Pilot". It's "Driver Assist", and until it's "perfected" (when that day comes), then call it "Auto-Pilot". Anything else is very mis-leading and damned dangerous.
Statistics are only as valid as the data they're based on, and the assumptions made about the data that isn't there.
Most transportation statistics are missing a LOT of base data. Things like "miles driven per year" are guesses.
Except in the case of cars like the Tesla, where there is a black box collecting statistics. How does Tesla know that auto pilot was on or off? It's recorded. How many miles are driven with AP on or off is recorded.
Many of the details are tossed out after an interval, but Tesla can collect a whole lot of data that other manufacturers cannot.
Now, the particular problem with dividing lanes is probably tied to trying to stay between the lines, when the lines are spreading out. If you don't stick to one line or the other, you're target is what is in the middle, and it is going to hurt.
If it requires a fully attentive driver, it's NOT safe.
If by "poorly" they mean "catastrophically" then I agree. Slamming into a barrier at cruising speed would be utterly fatal to the poor mobile-device-hypnotized sap behind the wheel. Not that I'd miss those morons - good riddance in fact, but I'd be genuinely concerned for the other drivers in the immediate area who now have to dodge the wreckage being flung all over the place.
Tesla argues that this issue doesn't necessarily mean that Autopilot is unsafe. 'Autopilot is intended for use only with a fully attentive driver,'
Everyone, Tesla included, knows that any trendy moron who wants a Tesla is the same kind of moron who can't tear themselves away from their devices for more than a few minutes at a time. On top of that, given all the talk about fully autonomous cars being "right around the corner", and the colossally stupid choice of "autopilot" as a name for their assisted lane tracking and cruising speed, of course said morons will conflate the two, and of course they will lose interest in watching the fucking road when their hypno-droid beeps and buzzes. Look at the fucking uber-moron who was supposed to be a TESTER who couldn't be bothered to stay focused on the car, and imagine how much more inattentive a random consumer would be who just bought that automotive virtue trophy after being told "IT HAS AUTOPILOT" by a dealer. That's the kind of self-righteous moron who thinks he's invincible - by way of "it'll never happen to me". It is therefore completely unacceptable to claim this shit is safe.
Tesla, and other such manufacturers are ENABLING these morons to be even worse. I consider that a net reduction in safety.
Tesla argues that Autopilot-equipped Tesla cars have gone 320 million miles per fatality, much better than the 86 million miles for the average car.
Show me the stats of JUST the autopiloted hours - specifically, the number of hours that the drivers WEREN'T paying attention - because that's what really fucking counts.
Until they stop calling it autopilot, in my book, every single crash and near crash by that shit system is proof that they are WORSE than a human driver.
For stuffing yourself in a metal can, then propelling it at great speeds just barely passing other metal cans (within inches). Ludacris you folks are...
[($)]
These accidents are evidence that the Tesla autopilot has caused some accidents. We have no real evidence that it prevents any.
Tesla keeps saying autopilot requires a fully attentive driver. But it appears to require a fully attentive driver capable of recogniing the autopilot failures early enough to prevent an accident. It does not appear that either Tesla or department of motor vehicles test motorists for that ability.
Autopilot is misleading, because these aren't airplanes, motorists aren't trained pilots and the "autopilot" is far from automatic. As I understand airplane autopilot, pilots are not waiting to correct its failures. They are watching for conditions where they know it is unreliable.
Time to fix an autopilot command defect beyond the FAA ATP flight standard on an airplane with a professional experience ATP trained crew. About 25-35 seconds, typically 1-3 miles and 400 feet from assigned altitude at 250kts. The navigation aids are external to the flight director and self-testing and validating and basically straight or performance envelope limited commands in 3-dimensional space. Basically aircraft autopilot standard keeps pilots in a 400x400x16000 foot box with any amount of consistency over the period of an hour at cruise. The primary navigation information is GPS accurate to 1/100th of that box.
Time to fix a Telsa autopilot command defect is on the order of 1-3 seconds and the deviation from lane center need be no more than 4 feet (quarter of a lane) at 50kts. The navigation cues to the vehicle are image and lidar detection of objects under any lighting, any color, any weather conditions with or without 4 foot tumbleweeds. The tesal product keeps the pilot within a 10x30x5 foot box at one-tenth the speed with the odds of a crash because of deviating from that box being high over trip with factors of ten more obstacles. The navaid is a mix of 4+ sensors with blah repeatability.
The Tesla autopilot is immensely more complex than the aircraft autopilot just to position itself in the center of an unmarked lane. The people killed by aircraft autopilots are typically pilots with less than 10 hours of systems training for the type and less than 4 hours of experience with the autopilot in actual conditions. Tesla offers what a youtube video to train on the operation of system navigating within 4 feet of a movable barrier at speeds close to 100 feet per second.
The remarkable thing is that Tesla autopilot has not killed more people.
Autopliot steering directly into a concrete barrier at highway speeds in broad daylight is an enormous bug. Multiple sensors must have detected the barrier under those conditions, yet the onboard AI chose to drive into it. This looks like a great big hole in Tesla's software validation process and badly faulty software. A concrete barrier directly in front of the car is not some one-in-a-zillion anomalous corner case. Autopilot software architecture must be very badly flawed; Even if the lane-detection function misguided the car, there should have been an emergency obstacle-avoidance function which superseded lane-following.
Regardless of Tesla's public efforts to downplay the barrier collision bug, it seems likely that their top priority now inside the company would be to fix it and, more meta, overhaul their software architecture and validation procedures.
Ceci n'est pas une signature.
The 'lane logic' is too stupid to differentiate between highway and 2-lane road. It sees a solid line on the left and tracks it because that's what it is supposed to do on a 2-lane road.
That the other sensors doesn't bother to scan far enough ahead to see a rapidly approaching obstacle and doesn't immediately apply brakes, shake the wheel and override all audio with a ear splitting klaxon is pure malicious incompetence by the software team and the company broadly.
wow, my code word: 'corpse'. No kidding!
Tesla named autopilot as similar in function to the autopilot in an airplane. It just flies in a straight line and has no obstacles to contend with.
I think this makes sense but no-one seems to get this, except maybe pilots. Intentional or not, Tesla touts this as a prelude to full autonomous driving and thus implies that it is in some way similar to it, but really has nothing to do with it. I think this was a big mistake, they should have called it something more obscure like TACC wLC (Traffic Aware Cruise Control with Lane Keeping), because that is all it really does. Then expectations for it would be far lower and there would be no problem here. I also don't think they make it explicit enough that this is a beta feature in the software.
I tend to only use the TACC part, which works really well. The steering really only works in ideal conditions, does not deal well with exit ramps (or similar issues), and can also tend to shimmy side-to-side hunting for markers, or drives too close for comfort to barriers or adjacent cars. I don't trust it when I am possibly a millisecond away from a crash, it just causes me more stress than just driving myself or with just TACC. There are other good reasons to have the feature, such as auto-park, and the accident avoidance is always on regardless of if you are using autopilot or not.
They are technically correct about the statistics, but it is just not working for them as a PR defense, because there is too much focus on Tesla even as people die by the thousands in other types of cars.
- RoS
As a Tesla driver for several years, I can definitely attest to this problem. Itâ(TM)s very unpredictable but I have definitely had my car start to drive over the center line or off of the right lane. You really do have to keep an eye on it.
It should be pretty obvious that anything that sticks up from the surface more than an inch or two should be be considered a hazard to avoid. And we can put hundreds of sensors all around the vehicle that will see everything. But we have problems because everybody is trying to do things on the cheap and cut corners. If it wasn't for that, we would have damn near flawless systems!
You mean Tesla's autopilot turns you into a line divider.
Table-ized A.I.
This blithe dismissal of the fact that a machine is making decisions that kill people falls right in line with the need of industry to change the mass perception about machines killing people. And it will get better, right? Death by death. People are expendable apparently.
E Proelio Veritas.
I generally blame Tesla for over selling the Auto Pilot feature. In turn some Tesla owners become to trusting of the technology. All it takes is one time for the system to not predict properly to have a accident. If the human driver is also lulled into a false sense of trust. Then nobody will react in time. The solution is easy, make the system disable when it detects the human driver is not holding the steering wheel. If Tesla really cares about selling cars then it will take seriously these accidents.
'Autopilot is intended for use only with a fully attentive driver,'
Said no logic ever.
"You had one job!"
"showing their Tesla vehicles steering toward concrete barriers. ..... Tesla argues that this issue doesn't necessarily mean that Autopilot is unsafe"
When I think safety, it certainly includes things like purposefully steering into concrete barriers!
I think the auto pilot can help.... but it isn't "auto pilot" of course because you have to pay attention. Which is harder to do when you're the one not driving the car. If my car started aiming itself at divider barriers, I'd probably stop using the feature entirely.
I wanted to watch the video from the near misses, but YouTube is starting with a one minute commercial.
Which are you? Jealous or a Luddite?
Because if you are not fascinated and excited at the idea of a robot driving your car, you might as well turn your geek card in and leave slashdot
Maybe I'm old-fashioned, but how about people just drive while they're driving? Then again, I am one of the seemingly-few people who actually enjoys driving. So perhaps I am being unrealistic.
"What the American public doesn't know is what makes them the American public." -Ray Zalinsky (Tommy Boy)
I have a 2016 ford edge with lane assist, and all I can say is no shit... Lane width and markings, while I am sure there is a standard, are not standard. If you rely on these features to keep you in your lane, or not crash, you are a moron.
How can we expect the onboard processing power of a vehicle to abide by constantly changing non standard conditions, when highly trained and skilled humans cannot...?
Now that several humans have sacrificed their lives to help you realize what you should have assumed all along, the question is -- if you know you must stay completely alert and ready to take the wheel to stay alive, why engage autopilot? Ever?
Compared to driving, it takes more alertness to generally assess road conditions. discern if a correction is necessary, then WAIT to see if the computer is responding appropriately, then ALLOW a sense of urgency to build in your mind over a short time period until finally ACTION is taken by you with a twinge of adrenaline because you have registered this (correctly) as a potential brush with disaster.
As a Tesla autopilot driver. how many times has this happened? And as it happens repeatedly have you permitted another sense of urgency to build, the conviction that your early confidence in the system may have been misplaced and you put yourself and other drivers, and whole families on the road, at risk?
How much are you getting paid extra for this? Is it worth it?
Does the Tesla 'drive better' than you do in some ways, initiate lane changes with greater precision and confidence at times you would hold back until a larger opening appears? If so... now that you're firmly in a mental state of being hyper-aware that you may need to take over... how does that make you feel? Like you are riding a roller coaster with poorly maintained rails? Good! You might get through this.
Now let's talk about those highway skills. Whether or not autopilot navigates the highway as well as you can, what will be the effect on you as days and weeks pass and the mental intuition you used to completely 'just operate' the vehicle has been replaced with a new and different set of instinct and action. As time goes by every move the autopilot makes on your behalf becomes something you once did, a more distant memory, a skill unused. Of course you won't forget but it is possible to lose your edge.
How important is your edge? Taking greater control of your own environment (and destiny) has been the driving force in human advancement for millennia. You are now participating in someone else's lab experiment, to see if you can gain any satisfaction from the idea of losing control of your environment, and destiny.
How much are you getting paid extra for this? Is it worth it?
Turn the damned thing off. And now that it has killed other humans, to avenge them you are honor bound to have Tesla disable the 'feature' and refund some of your money. You still have a cool car.
<blink>down the rabbit hole</blink>
On both videos you can clearly see that as the lanes split up the marking of the "real" lane the car was on disappears.
The car mistakes the right-side lane marking from the left lane as the left-side marking of its own lane and drives towards the concrete block.
Questions:
1. Doesn't Tesla have sensors that detect obstacles (to at least brake)?
2. Shouldn't their algorithm detect that the distance between left and right lane markings grows and panic (i.e. brake)?
3. Is Tesla only able to drive on perfectly marked roads? Most people are fine with faded, or even none-existing lane markers