NHTSA’s full investigation into Tesla’s Autopilot shows 40% crash rate reduction
techcrunch.com
techcrunch.com
It took a combination of problems to cause that crash. The police lieutenant who had informed the railroad of the parade in previous years had retired, and his replacement didn't do it. The police marshalling the parade let it go through red lights. They were unaware that the traffic light near the railroad crossing was tied in to the crossing gates and signals. That's done to clear traffic from the tracks when a train is approaching before the gates go down. So ignoring the traffic signal took away 10 seconds of warning time. The driver thought the police had taken care of safety issues and was looking backwards at the trailer he was pulling, not sideways along the track. People at the parade were using air horns which sounded like a train horn, so the driver didn't notice the real train horn. That's what an NTSB investigation digs up. Those are worth reading to see how to analyze a failure.
[1] https://www.ntsb.gov/investigations/AccidentReports/Pages/HW...
[2] https://www.ntsb.gov/investigations/AccidentReports/Pages/HA...
Even if autopilot was faulty in some way, if more people were living than dying - does it matter?
Similarly, 9/11 would be a footnote compared to sketchy pain killer drug trials. The black lives matter movement would focus more on sentencing disparities and for-profit prison conditions than police shootings, etc, etc.
The reason it matters here is that it's presumably easy for Tesla to fix any bugs that are killing people.
That is, Autopilot makes certain kinds of accidents more likely — those in which the human driver is not monitoring the vehicle carefully enough, Autopilot misses something critical, and an accident results. It is possible, though not necessarily the case, that these accidents tend to be more serious than accidents that Autopilot prevents, such as lane-changing bumps or other less-fatal accidents.
To be clear, I'm not saying that evidence has been presented that the type of accidents prevented are less important than the type facilitated. I'm just pointing out that not all accidents are equal, and that there may be correlation/causation that makes the overall calculus more complicated.
Now, someone makes a robot that has 100% success rate, but every 100 operations, it has a glitch in the sofware, where it stabs and kills the patient instead. It has been calculated that on average, using the robot only kills 1-2 out of 100 people going in for surgery, compared to the old method of 80 people out of 100. The robot is clearly 80 times better than manual procedure, so why not keep using it?
The answer seems to be - because robot killing people is not an unavoidable outcome. It can be fixed. It's a software problem, a mistake that costs someone their lives. Similarily, an auto-driving Tesla may be killing less people on average than manual drivers, but that doesn't mean that any software problems that lead to it don't matter.
Step 2: Correct the bug, so the robot stops killing patients.
The Tesla autopilot may be in a similar situation. Even if there is a number of fatality inducing bugs, disabling it would be even worse. That said, 40% crash reduction is not enough. I want to see 75%, 90%, and more.
Don't even speak of the bug. Just state what matters: this stuff is better than what we had before, and it can (and will) be even better. This may prevent talks about killer robots.
The biggest difference between an autopilot accident and a traditional accident is fault. Most accidents are driver error. Even if it drives accidents, overall, down (which it will), changing the responsibility for safe travel from mostly the individual to mostly the manufacturer has huge legal repercussions. Not to mention many people's discomfort with no longer being in control of their own safety, even if study after study shows that they shouldn't be.
The therac arguably saved way more people than it killed. However... it's one of the most infamous software bugs in history.
It's the same story you came up with above... except it's real, and it got pulled.
They did however improve since then.
Apart from that, more people living than dying could also imply that certain groups of people, maybe kids or bikers, could be at a higher risk of dying while the majority has a much lower risk, resulting in a net positive. So the acceptance of self-driving systems would also depend on the distribution of the risk.
If you let your friend drive you somewhere in your car because you think they're a better driver, is it their fault if they crash?
The reason could be anything - you're drunk and they're not, you're in really congested unfamiliar area near where they live, whatever. If you think they're statistically going to be a better driver it does not absolve them of liability. Someone is always to blame.
Autopilot will reduced the number of fatalities in the same way that going to the hospital when you are sick will reduce the number of fatalities. However some people that use Autopilot will die when they otherwise would not have if they hadn't used Autopilot, just like how some people going to the hospital will die from infection when they otherwise would not have if they hadn't gone to the hospital.
But it doesn't mean that people dying from faults "don't matter". Every HAI-caused death is a tragedy, and they certainly matter.
The issue is much less black-and-white than the GP's simplistic view.
There's no requirement that driver assist software magically become self-driving software in that situation. Put another way: it's nice if the software can avoid a problem, but not required.
The main remaining concern is that drivers not be confused about what they are required to do; there's discussion about that in this document.
This is probably why Musk pulled radar and vision processing in-house. Combining the data from both sensors is more effective, but more difficult, than combining the results from processing both sensors separately.
Other companies are being even more cautious; Ford isn't putting autonomy on the road until 2021. In the meanwhile Tesla can sell products as Level 2, and perhaps should do so until roads are as safe as commercial aircraft.
So far NHTSA has encouraged such data sharing but they haven't outright mandated it.
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 and after Autopilot installation. 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.
I had hoped to see more information about this specific incident. For instance, any data on whether the driver had his hands on the wheel, what steps the car had taken to prompt his attention, etc. But that doesn't seem to be included.
IIHS research shows that AEB systems meeting the commitment would reduce rear-end crashes[emphasis added] by 40 percent. IIHS estimates that by 2025 – the earliest NHTSA believes it could realistically implement a regulatory requirement for AEB – the commitment will prevent 28,000 crashes and 12,000 injuries.Some excerpts from "Effectiveness of Forward Collision Warning Systems with and without Autonomous Emergency Braking in Reducing Police-Reported Crash Rates", January 2016:
"FCW alone and FCW with AEB reduced rear-end striking crash involvement rates by 23% and 39%, respectively. "
"Among the 15,802 injury crash involvements in these states, the percentage of injury crash involvements that were rear-end striking crashes was larger among vehicles without front crash prevention (15%) than among vehicles with FCW alone (12%) or FCW with AEB (9%). Only 4% of rear- end injury crashes involved fatalities or serious (A-level) injuries."
"Approximately 700,000 U.S. police-reported rear-end crashes in 2013 and 300,000 injuries in such crashes could have been prevented if all vehicles were equipped with FCW with AEB that performs similarly as it did for study vehicles."
[1] http://orfe.princeton.edu/~alaink/SmartDrivingCars/Papers/II...
Did Tesla cars have AEB prior to Autopilot installation? If not, then this suggests the 40% reduction in crashes may simply be due to the installation of an AEB system. What effect Autopilot's other features may have would remain uncertain.
If AEB did lead to a 40% reduction in crash rates for Tesla cars, as it did for other car models, I suspect that moving the dividing line by five months later wouldn't change the figures much: you would still have a lot more crashes prior to AEB and fewer after.
I don't follow the statement about sensor availability. That doesn't seem to change the fact that all the miles driven "after Autosteer" benefited from AEB, while at most a small fraction of the miles driven "before Autosteer" would have had AEB available.
Given that we know AEB systems do reduce frontal collision rates by 40% for all cars, as the NHTSA report stresses, that implies we cannot attribute the reduction in crashes "after Autosteer" to Autosteer alone.
(Which matters because articles such as the one posted are claiming a cause-and-effect relationship between the introduction of Tesla Autopilot and a reduction in crashes, but if a significant portion or even all of the reduction is due to AEB systems that other cars have also started to adopt then we're mistaking correlation for cause.)
Tesla rolled out AEB in March 2015 as a software update for all cars with the appropriate hardware. All Teslas produced starting in mid-September 2014 have the appropriate hardware, so several thousand MY 2014 Teslas have AEB as of early 2015.
I will admit that there have been a couple times it saw the car I was about to merge into before I did. It is not clear if I would have seen the car a moment latter anyway or not.
http://www.businessinsider.com/the-cause-of-the-most-fatal-c...
We can estimate the absolute number from Tesla's press release in December stating that 1.3 billion miles have been driven on Autopilot, while the NHTSA report states the accidents they count occur 0.8 times per million miles.
So in that case it's data from about 1000 crashes in the Autopilot case, and more in the non-Autopilot case. Which tells us that the statistics should be good enough (anything else would be a major scandal).
However, that is not what they are comparing! As the NHTSA report states in footnote 21, they're not actually comparing Autopilot miles to normal driving. They are comparing crash rates in vehicles before and after Autopilot was installed in those vehicles. That muddies the waters quite a bit in my opinion, especially when they don't tell us how significant this decrease is. I'm not saying it's wrong, but I'm also not very confident that it is a causal effect, as opposed to just a correlation.
Here is a PDF link to the NHTSA report, since TFA only links to a Scribd (thus unusable) version.
https://static.nhtsa.gov/odi/inv/2016/INCLA-PE16007-7876.PDF
It just seems...odd.
It's weird.
A lot of people drive a lot of miles which is why there are a lot of crashes. So you can argue either side of this depending on how you want to twist the numbers.
Of course, whatever algorithm they are using won't work if the car is driving itself, you can't judge the reaction time if all the driver is doing is holding the wheel.
implying that law enforcement won't work with homeland security to subpoena your data at cost under a gag order :)
[Seriously, though. I don't want to downvote everyone in a chain but there's probably a more HN/SNR-friendly way to comment on potential abuses of automaker access to social media feeds.]
immediately pull over and shut down
If your goal is to make things safer I don't think this will help.
And my other suggestion about a TomTomagotchi that you have to regularly drive around to different places to keep happy didn't go over very well, either. Maybe a petroleum company would be more receptive to that idea, though.
Its such an odd patent, perhaps its a high quality signal about their ability to execute on self driving.
[1] https://techcrunch.com/2017/01/18/ford-patents-a-car-camera-...
I think there is a good SaaS business with A/B testing robot/car software OTA. I would imagine some metrics would be: 1. car speed 2. battery consumption per mile. 3. user comfort metrics (if it's possible to measure, maybe like hearthrate, body temperature etc - i can imagine if robots were driving bad, it'd be visible in heart rate or something like that).4
The list probably goes on. It would be interesting to see how much battery consumption changed after OTA, for example.
6.2 was announced in March, 2015.
You mean drivers like the ones behind the wheel of all the BMWs and Audis that cut me off with no turn signals, speed, make aggressive lane changes, tailgate, and generally act like giant douches because a simple traffic ticket is nothing to their pocketbook while their time == money?
Those people are going to be a hell of a lot safer behind the wheel of an autonomous (level 4) vehicle where they can be on the phone and their laptop as the vehicle obeys the speed limits and safe following distance.
Could just be the numbers of those kinds of cars in my area though.
Generally I feel that a douche can drive any type of car, and I try to drive in fear like everyone else on the road is a drunk prison escapee.
(1) provisional licence holders - new drivers (and also people who have lost their licence) have to display "P" plates for a year here (Australia).
I've seen jalopies driven like someone has been up on speed for 4 days straight.
Or the drivers in cheap sensible cars that are terrified of merging onto the freeway faster than 40 mph.
And the distracted drivers who realize they're about to miss their exit so they hammer on the brakes and put on their turn signal and completely fail to match speeds.
Ricers with after market exhaust are another kind of car that usually drives way too fast and aggressive compared to how their lowered suspension actually performs.
The people in road tanks who seem to split the difference between terrified and aggressive and expect you to just get out of their way because they're bigger.
I'm probably missing entire swaths of shitty drivers that I can't think of right now because my blood sugar is a bit low.
Somehow I manage to consistently undertake people who are going slow in the left lane safely. Its not that hard.
Finland has a system where a traffic ticket is a certain percentage of your monthly income. For example, running a red light is typically fined for 25% of your monthly income.
You can buy traction control, ABS, better suspension, steering, better tires, etc. None of that will make the idiot behind the wheel any better, although it may lead to overconfidence.
Thanks for making my point for me though.
"Results Studies 1 and 2. Our first two studies were naturalistic field studies, and examined whether upper-class individuals behave more unethically than lower-class individuals while driving. In study 1, we investigated whether upper-class drivers were more likely to cut off other vehicles at a busy four-way intersection with stop signs on all sides. As vehicles are reliable indicators of a person's social rank and wealth (15), we used observers’ codes of vehicle status (make, age, and appearance) to index drivers’ social class. Observers stood near the intersection, coded the status of approaching vehicles, and recorded whether the driver cut off other vehicles by crossing the intersection before waiting their turn, a behavior that defies the California Vehicle Code. In the present study, 12.4% of drivers cut in front of other vehicles. A binary logistic regression indicated that upper-class drivers were the most likely to cut off other vehicles at the intersection, even when controlling for time of day, driver's perceived sex and age, and amount of traffic, b = 0.36, SE b = 0.18, P < 0.05"
But given that they've shown this 40% crash rate reduction in their current demographic, it's reasonable to expect to see a similar reduction in other demographics (maybe not of the same magnitude, but at least in the same direction).
A Tesla is not a cheap car. I kind of suspect the average Tesla driver is more responsible than the average driver. It may be the case that the average Tesla driver is very well off, and treat the cars as expensive toys though, so it's tough to say without numbers. If i had to bet right now, i'd bet on the more responsible side though.
If you study rich drivers I have no doubt that you would find that they cut people off more, overtake in more dangerous situations, use their indicators less, drive over the speed limit more, proceed through amber signals later, run red lights more often, and generally show less regard for the rules than other drivers.
This behaviour will also correlate to other areas of life where richer people simply care less for the rules. They're more likely to use social contacts or money (e.g.: "lawyer up") to get themselves out of any trouble they find themselves in.
Teslas are not cheap cars. Given that money is (social evidence of) power, the people buying Teslas (and other expensive cars in general) are likely to expect other drivers to respect their status as powerful people.
I would be willing to bet money that the insurance actuaries that actually calculate risk disagree with your extremely prejudiced opinion.
This list is not strongly related to risk /of/ accident, but is correlated with severity (a small change is likely to have a small correlation with severity).
* over the speed limit
* overtake in (more) dangerous situations
* through amber signals later
All of these are very bad things that everyone should avoid doing, and signals should be given starting 5-10 seconds before the action.
* use indicators less
* cut people off
* run red lightsWhich seems like the current driver profile is possibly safer than the overall population since it's unlikely a large percentage of young drivers own a $100k car. It seems like the effect of widespread autopilot could be greater if it included this cohort.
In 2015 2.44 million people were injured in car accidents. That's a much more telling figure of just how bad most drivers are. Apparently our cars are just really safe.
There are some truly consistently horrible drivers. But in general most people do ok most of the time. I think the real challenge is cutting of that really bad end of the distribution, which autopilot seems to be doing.
Another factor is the extent to which bad driving doesn't result in an accident, either because of infrastructure, other drivers reacting, or just a lack of cars. For example, most red light runners don't cause a crash because stoplights usually have an all-red period to compensate, there often aren't other cars to crash into, and other cars might brake to avoid crashing into the culprit.
Many crashes require two bad drivers to misbehave at about the same time. For example, I couldn't tell you how many times I've had somebody change lanes into me, but it hasn't (yet) resulted in a crash because I've seen it and gotten out of the way before they hit.
love you, internet.
For my assumption I assume that due to the higher cost of a Tesla better educated people are opting for them over the alternative because of the environmental benefits.
I mean, who do you think is more likely to buy a Tesla: A 20 year old fresh college graduate, or his 'uneducated' father who has been saving and investing for 30 years and just sold his lifelong home in San Francisco? My bet is on the latter. The boomers in particular are considered the wealthiest generation ever, but not the most educated generation ever.
I'm rather fascinated by how many are questioning the education thing here though. Coming from a farming community, where every (older) farmer I know is multimillions simply by virtue of having purchased farmland when they were young, it's difficult to see how education had any impact on that wealth. Some of them do have educations, some don't. It doesn't seem to make any difference.
OP linked wealth with being better educated, he also inferred that better educated and wealthy meant they would be better drivers. However as others have pointed out there are studies showing that wealthier individuals make worse drivers (the studies don't appear factor in education level and wealth is determined by brand of vehicle).
I disagreed with OPs sentiment that fewer wealthy (and therefore worse educated - in OPs opinion) drivers would result in more crashes.
The imperfect, incomplete, beta, level 2 self driving cars that were supposed to be the "dangerous" area of self driving are ALREADY better than human drivers.
Can we stop the politics and deploy all the real self driving cars to the road immediately, since the government has proven that even the shitty variety is safer than humans?
Tesla's system is level 2, meaning that the driver must (as in, is supposed to) remain engaged and aware of the driving task at all times, although the car will handle common cases on its own.
I believe the plan for everybody is to go straight from level 2 to level 4, since it looks like level 3 is just not going to work.
In any case, we can't stop the politics and deploy all the real self driving cars to the road immediately, because there aren't any real self driving cars to deploy yet. The technology is advancing extremely quickly and it's getting close, but politics is not the only obstacle.
If and when we have a real, production-worthy self-driving system that's being blocked by politics for no good reason, we can revisit this. I get the sense that, at least in most places, the politics will be pretty easy once the technology gets there. Getting a lot fewer registered voters killed has a way of swaying legislators. Especially in the US, where it's a state-by-state decision, a few states will want to get a jump on it and then the rest will face immense pressure to follow.
The current Tesla autopilot is obviously working well for many people, but for myself I have to wait until level 4. Anything more hands on than that is inherently unsafe because of my habits.
Computer vision and speech recognition could be useful for proving the driver's awake, as well as driving the car itself.
In addition, not having to spend as much energy just keeping the car on the road may help to prevent getting tired. Then again, maybe it won't have any impact.
As for durations at the wheel, I did a Virginia to Wisconsin round-trip over the holidays, and have done other trips of similar lengths in the past. Of course, I'm not spending more than about three hours at the wheel at a time, since the car needs to charge and I have various physiological needs.
Modulating my speed keeps me engaged with driving and I use it as the trigger to cycle through my responsibilities. I check ahead / speed, adjust throttle, check rearview mirror, check left mirror, check right mirror, repeat. When the throttle step is gone I find myself losing siutational focus very quickly. I've always thought it was because it was the only action that involved actual physical engagement, all the rest are just eye movement.
I just bought a new car that alerts when it's below 4 degrees C about the possible icy conditions, but it only does so after the car has been running for several minutes. By the time it goes off I'm already driving and I mistake it for an engine warning or other catastrophic failure. Every time. I live in Canada, of course it's icy in the winter. Toddlers could have told you that. This is an absurdly dangerous thing to have the car do, because when the beep goes off all my attention is on processing the alert, not about the pedestrian I'm about to run over. I have the same basic fear about getting control of a level 2 or 3 car. All my focus will be on why the car is handing me control, not on the actual task of driving.
>The Uniform Vehicle Code states: Upon all roadways any vehicle proceeding at less than the normal speed of traffic at the time and place and under the conditions then existing shall be driven in the right-hand lane then available for traffic ...
>All states allow drivers to use the left lane (when there is more than one in the same direction) to pass. Most states restrict use of the left lane by slow-moving traffic that is not passing. The table below describes the law in effect in each state.
Or you are driving at the speed limit on a 2x2 highway used as a trucking corridor. (For example: most of I-5.)
The right lane is pretty much for trucks going 55.
The left lane is a combination of: * Trucks going 60-65 to pass a truck in the right lane * 'Typical' cars who are going 75ish * Fast cars who are going 80-90
So there is a huge disparity of speed in that lane -- and that disparity seems like it must be a huge factor in many highway accidents.
It always feels really dangerous. I wonder if it is safer than just having the trucks have the same speed limit as the cars.
I thought it allowed the car to require some parameters (such as road type or weather). If my car can handle the hours-long uninterrupted highway sections of a road trip, yet still require manual driving on city streets, that is still a huge value proposition for time & safety.
In my mind, if a level 4 vehicle is unable to reach a given valid destination without assistance, that would be considered a failure.
In reality, macro-scale rerouting around obstacles that the self-driving system does not know how to overcome is not a particular challenge, GPS systems do this easily. The definition doesn't say it has to take the most efficient route. ;)
Oh, that's going to go over really well. Say you're drunk, and want to go home in your self-driving car. However, there's a blind left turn that your car knows it can't safely navigate, so it happily routes you a 15-minute drive through back roads for a trip that should take 5. People are going to love that.
(I've heard that UPS does route planning with a system that discourages left turns, even sometimes doing three rights to avoid some dangerous lefts.)
Many of those scenarios would allow some time for the driver to take over, and would most likely be able to pull over if the driver can't or won't for some reason.
However, Arizona would be a beautiful place to have fully autonomous cars.
The snow bunnies make driving there a nightmare.
The truck would know well in advance which exit it would take and could give the driver a heads-up some minutes before they need to take the wheel.
>22 The crash rates are for all miles travelled before and after Autopilot installation and are not limited to actual Autopilot use.
Because of course there were no autopilot miles in the before bucket you would be skewing the analysis if you only looked at autopilot miles in the after bucket.
What we know is after installing the autopilot feature, accidents with airbag deployments reduced 40%. Not just during so-called "safe" autopilot miles, but over all miles driven.
Now it's plausible that autopilot makes you a better driver both while it's on and off, but I'd say it's just exceedingly common to rear-end people due to distracted driving.
the driving mode-specific execution by a driver assistance system of either steering or acceleration/deceleration using information about the driving environment and with the expectation that the human driver perform all remaining aspects of the dynamic driving task
So yes, it is literally human Vs computer driven safety.
You are bringing it up, but I don't think you are doing this for any reason other than to be contrarian.
Also, ALL cars have cruise control right now. So the 40 percent STILL applies to all these cars.
The correct way to frase the study is that "99 percent of cars are level 1 autonomous cars. For this 99 percent, going from level 1 to level 2 would make them 40 % safer"
The name is good for technical people, not so much for lay people who don't know that autopilots are more like cruse controls than something that completely flies the plane (they have been getting better though). The 40% reduction in air-bag activation seems to indicate the name is good enough while its probably not optimal.
Depending on how all the numbers would work out, making other drivers safer could well make you safer too. It might not, but it could.
ODI analyzed data from crashes of Tesla Model S and Model X vehicles involving airbag deployments that occurred while operating in, or within 15 seconds of transitioning from, Autopilot mode. Some crashes involved impacts from other vehicles striking the Tesla from various directions with little to no warning to the Tesla driver.
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.
21 Approximately one-third of the subject vehicles accumulated mileage prior to Autopilot installation.
22 The crash rates are for all miles travelled before and after Autopilot installation and are not limited to actual Autopilot use.
So the actual rates of crashes for Teslas using Autopilot vs Teslas not using Autopilot aren't reported.
This strikes me as suspiciously high: what fraction of airbag-worth accidents are supposed to happen on the highways where one can use Autopilot at all? So maybe there's some other trend going on which is driving this decrease.
Even if the only benefit of Tesla's autosteer is that it makes you more focused in the parts that are actually dangerous, a 40% improvement is still a 40% improvement.
At least they did not compare "all Teslas" vs "Teslas with Autopilot Technology Package checkbox checked", since that would select for drivers who do a lot of miles in boring long distance driving. That kind of driving happens to be the one where any car has an "airbag mileage" far above average, so it would be just natural that they have a lower crash rate, autosteer or not.
But a part of that effect might still be present in the numbers if customers are less likely to pay for a post-factory update than while ordering the car (which I do suspect). If this holds, then the subset of drivers who activated post factory (with miles already on the clock, so that this part of their car usage appears in the "before autosteer" column) will have, on average, a bigger desire for autosteer (and more naturally safe long distance cruising in their driving patterns) than the average of those who got the car with the option factory-activated (and who thus appear only in the "after autosteer" column).
Edit: oh, and here's the really big one: the "before autosteer" number only includes cars that where upgraded. How is this important? All airbag activations that happened in an event that messed up the car beyond repair cannot appear in the "before update" dataset, because nobody spends thousands of dollars on an autopilot update for a car wreck.
[edit 2]: This should actually make the "before" number much better than the "after" number, so i really wonder what's going on here.
The obsession with perfection in self-driving cars is misplaced, they just need to be demonstrably better than humans.
This is obviously the future.
From a safety point of view, I'm more inclined to agree however I'm not entirely convinced that we've passed the novelty stage of it yet where drivers will remain vigilant because it's a new and exciting feature and that mentally primes them to keep focused. It wouldn't suprise me that as it becomes more mainstream, drivers begin to be less vigilant on staying focused as society as a whole begins to trust it more. That's when everything could go bad.
Nevertheless, you're absolutely right that this is the future no matter what so we need to design it as best we can to resolve all of the issues it may give rise to.
Additionally, who's at 'fault' really doesn't matter when in most cases the damages are fully outsourced to insurance providers. In the case of autonomous driving, this insurance burden is shifted up a level to the automaker, but this cost will be passed on to the car's owner in some way, either in the cost of the car or a subscription to autonomous driving.
However if autonomous driving is safer than non-autonomous driving, premiums will go down, probably dramatically. Insurance companies can much more easily monitor the quality of autonomous driving for various automakers / cars either by the testing specification that they meet or the real-world statistics. Are autonomous Tesla's in less accidents than autonomous Volvo's? Cheaper premiums for Tesla.
This is exactly the situation you want to be in. Premiums for individuals are blunt instrument... it's hard to change many individuals behavior, the rewards may be low for the individual (save $3 on your premium!), and the penalties can be pretty random (you sped that one time so you're a high risk driver... or are there just more police in your city?). It's way better to incentivize at a higher level, like the automaker.
The only case where automakers would truly be at risk is something they should control, and that is if they pulled a VW and committed fraud on their autonomous driving tests, or were negligent in overlooking scenarios or providing updates.
All I can find from the article about weather was in section 3.1:
> The manual includes several additional warnings related to system limitations, use near pedestrians and cyclists, and use on winding roads with sharp curves or with slippery surfaces or poor weather conditions. The system does not prevent operation on any road types.
You have to look at what they're using for this, it's cameras with computer vision and radar. Radar has reflection problems like all hell at ground level and computer vision using cameras is generally piss poor (hence why you see videos of the Tesla hitting reflective objects).
It's a miracle that they can do this well using it IMO, even without adverse weather.
In my experience, its the older people that are somehow scared of the concept of a self-driving car and would need to be convinced.
Just a guess: probably.
Still, I know I drive stupidly sometimes, and autopilot might help.
https://www.youtube.com/watch?v=jhUX8qWFGc4
Imagine you are too fucked up to drive. Your car will be able to pick you up. Do you need Pepto Bismol too? Your car will pick it up from a drive through billed through your license plate. I'd give roughly 15~20 years for this to take place.
http://arstechnica.com/cars/2017/01/after-fatal-tesla-crash-...
Strangely enough, giving more people autopilot would probably be better than letting people drive. I think Tesla's picked the right time to enable it, since the cross-over point between autopilots being better than humans in general use cases has been reached.
Call it a beta if you want, but it's a pretty damn promising beta.
[1] http://www.newyorker.com/magazine/2015/05/04/the-engineers-l...
Tesla's have driven >= 3 billion miles
:O
What I found interesting is that figure 10 shows that as you jack up the independence of the machine, the level of driver distraction accordingly increases. Adaptave Cruise Control (ACC) shows a significantly higher percentage of shorter-duration off-road glances than Limited-Ability Autonomous Driving Systems (LAADS). Additionally, countermeasures help to alleviate some but not all of that increase in distraction. Importantly, this is coupled with the point that the duration in which drivers have to react to most impending is under 3 seconds. This may seem obvious but it's a critical set of data to help objectively demonstrate the risks involved with losing or even reducing alertness.
It goes on to say that Tesla has addressed the risks of mode confusion, distraction, etc. and has implemented solutions to address this unreasonable risk, which they in turn define as abuse that is reasonably forseeable. In this, they're talking about the reasonably forseeable risk of eg: the driver not understanding if they're in autopilot or not. It goes on to mention that Tesla has also changed its driver monitoring strategy to promote driver attention, which I take to mean detecting hands on the steering wheel.
Either way, Telsa's main approach to dealing with driver alertness by testing for hands on the steering wheel. My concern is that this doesn't consider the alertness of the driver to their surroundings, particularly other vehicles that may be approaching them or the process of anticipating hazards (approaching an intersecion where there's a blind corner and adjusting focus to pay more attention to what may come from it, for example). I don't see how Tesla's countermeasures address this.
The physical act of manually driving causes drivers to maintain alertness not only to where they're going, but also the situational alertness of what's around their vehicle. Specifically, it's the process of random actions that requires an taking input, making a decision and executing the appropriate action that maintains this alertness. If the driver isn't having to make those random decisions and take action then their alertness drops. Autopilot, even with hands on the wheel, eliminates much of that random decision-making and reacting.
When you drive, you mentally note the vehicle over your shoulder that is in the lane next to you, and subconsciously consider that they may do something insane. You consider those blind corners as you approach them and that vehicles may spontaneously appear out from them. You see a truck on the road which is approaching a bend and give it a wider berth because its centre throw may cause it to cut the corner into your lane slightly. These are all tasks that you do, that you may not do as well or at all when autopilot is steering, because you are not as engaged with the driving process.
Critically, I don't see how ensuring hands are on the steering wheel causes these alertness tasks to continue as frequently as manual steering. The driver may be in the physical location to quickly take over, but they may not be in the mental location to do so. This is the major issue I have with the rapid autopilot development based on my experience in related areas where maintaining situational alertness proved to be very difficult when the person was engaged with only a limited scope of requirements to prove their conscious presence.
I feel like the report doesn't really drill in to this as much as it needs to. It begins to touch on it around Figure 10 but sort of hand-waves it away saying 'Telsa considered it discharged their responsibility to make sure drivers stay focused by implementing countermeasures', but I believe it's more nuanced than that. It investigates the extent to which Tesla's system is good at ensuring drivers are physically present (that is, their hands aren't on the passenger seat making breakfast) but it doesn't really look at the mental presence that delivers situational alertness.
That mental alertness is the major sticking point for me. I don't really have a solution beyond "drive manually" which isn't reasonable, because this technology is here to stay and will continue to grow, but it's why I've always been bearish about the rapid pace of rollout of these driverless technologies, particuarly when advertised as 'beta'. As I've said before, no amount of disclaimer and 'hey, you should do this' really changes how drivers behave once the equipment is placed in their hands.
Also, still havent seen any autodrive handle off-road driving such as boarding a carferry or navigate a construction zone manned by an inattentive flag person.
It doesn't matter that it is not full autopilot. What matters is that whatever they have right now, as currently deployed in the real world, produced the numbers behind NHTSA's conclusions.
Just sit back and chill. It's going to happen.
You cannot compare every arbitrary thing to how internet has progress and just win the argument. That is just wishful thinking.
In general, I think we are moving way too fast towards these self-driving vehicles because certain factions want to try and replace long and short haul truckers with robotic systems that are cheaper and damn the consequences.
Computers are good at lots of things like playing chess, and still completely incapable at lots of things that involve common sense. Expecting autonomous cars to be better than human drivers eventually is one thing, but it is certainly not obvious that we are there yet.
That's just ridiculous. It is going to be safer in some cases and less safe in others, and obviously it depends on the system and the driving conditions. To say that a computer is necessarily a better driver is just science fiction until we have solid data showing that to be the case.
Over 25 years, actually - CMU's ALVINN (1989) was one of the first full-sized vehicle implementations:
http://repository.cmu.edu/cgi/viewcontent.cgi?article=2874&c...
It was preceded by an E-MAXX based RC car system called DAVE (DARPA Autonomous Vehicle).
CMU also worked on an investigated other systems, going back to 1984 or so:
http://www.cs.cmu.edu/afs/cs/project/alv/www/
They had an interesting system called the Terragator, from at least 1983; I can't find a paper about this system, though I know they exist - there is a video, though:
https://www.youtube.com/watch?v=beGRM9ZmWlI
One might consider the Stanford Cart as an earlier example (I believed it used computer vision techniques - and was really slow at computing the "next move"):
https://web.stanford.edu/~learnest/cart.htm
(side note: As part of SAIL, the Cart was instrumental in paving the way for Thrun's Stanley - which won the DARPA Grand Challenge, and ultimately led to Google's efforts in self-driving cars)
ALVINN was unique in that it used a (simple compared to today's CNNs and deep learning) neural network (and a WARP systolic array):
https://en.wikipedia.org/wiki/WARP_(systolic_array)
...coupled to a camera to digitize and model/learn the road features for driving, while "watching" the human driver (early form of "behavioral cloning" - but it wasn't termed as such in the various papers you can find about ALVINN).
Interestingly, during the ML Class in 2011 (which I participated in), one of the students - after we were shown several times how ALVINN worked - decided to replicate the system using "modern hardware":
http://blog.davidsingleton.org/nnrccar/
We've really come a long way...