Normalized crash data shows Autopilot is much less safe than Tesla claims
twitter.com
twitter.com
1: https://engrxiv.org/preprint/view/1973/3986
2: https://www.thedailybeast.com/how-tesla-and-elon-musk-exagge...
- Freeways account for 93% of distance travelled with Autopilot
- Freeways account for 28% of the distance travelled by regular drivers.
- "vehicles on non-freeways crashed 2.01 times more often per mile"
a quick back of the envelop calculation suggests merely driving on the freeway lowers accidents per mile by ~40%, which accounts for most of the difference between Autopilot and regular drivers.
I get you a perfect Sample around age, income, sex, nationality, times of day and others. Prove to you 98% of people in this Town love Romantic Comedies.
The only thing I am not telling you, is that I am running the Inquiry at the exit atrium door of a Romantic Comedy movie...
Teslas are on average substantially newer than the average US passenger car fleet and thus have significantly more safety features and driver assistance tech, are better maintained, etc.
If Tesla compared only to vehicles of similar age and class (Audi/BMW/Mercedes etc) their stats would be much less favorable.
Recent statistics released by TüV^1 show that among relatively new cars Tesla competes with Dacia for the highest failure rate among all cars that go through Germanies mandatory road worthiness tests.
^1 https://www.pcwelt.de/news/TUeV-warnt-vor-typischer-Schwachs...
In short: big car brands' cars go through a preliminary check at manufacturers' licensed dealerships, and only after that get sent to the government's checks. It enables manufacturers to fix cars before independent verification so their safety numbers look great. Tesla doesn't have this kind of extensive network for "pre-checks".
Please don't compare it to high end german cars.
This is what matters. If/when it becomes less often, the case will be made (i.e. driverless doesn't have to be perfect to be better).
- Age of driver
- Specific highways used (congestion, road conditions)
- What is classified as a highway
As in one I can hopefully sleep at the wheel on a late night trip between midwestern US cities.
John Henryism can navigate the suburbs, stroads, and drivethrus. If they achieve a backup safety system that monitors the humans, well, ok great.
But right now, I want long distance automation, and I can handle the messy last mile. Highways are much easier to get uniform infrastructure convergence: some sensors, PSA warnings, better signs + markings + construction notices, etc.
I think Uber making as big a splash at it did sidetracked this low hanging fruit. IMO that was where the bootstrap bucks were, because you would have gotten freight/shipping to pay the cost of the next gen software, and would have provided the impetus for the convergent infrastructure.
Alas, they are obsessed with robotaxis.
(I live in Europe which has relatively good railway network, still I prefer my Tesla as transport method.)
Would you rather do this for a business trip to a 400-500 mile away city:
- fixed time to depart (and fixed for leaving your destination!)
- have to book ticket weeks in advance for a reasonable flight
- no flexiblity if plans change
- pack a small bag
- transit to an airport (parking or cab or whatever)
- arrive a sufficient amount of time in advance
- check in
- go through security
- get on plane, stuck in seat, no/crappy food
- get off plane, go to car rental place
- rent car (not yours, sucks)
- cost: 100-200$, and if you want to bring anyone it's a multiple of that.
- no cell phone, restricted to plane wifi if they have it
- door-to-door is what, 4 hours? Man I hope there isn't a delay....
Or, with self-driving:
- pack your car or minivan to the brim with stuff you need
- leave when you want
- sleep / stop to eat / stop to bathroom when you need it
- full cell coverage / hotspot
- take additional people
- EV cost is like... 20-40$ for the trip if you include wear and tear?
- door to door is... 5-6 hours
If the car is driving for you, it is far less stressful, flexible, much cheaper, especially if you want to take your family.
Trains are a ... bit ... better than airplanes, but still inflexible, surprisingly expensive, won't have your car at the other end, have to haul around your luggage rather than chuck it into YOUR CAR.
When good highway self-driving comes, I will probably never fly anything less than 1000 miles. If I get a self-driving RV that's EV self-recharging at charge stations? Hell I might get rid of my apartment.
But until they get built freeway autopilot is nice.
Me too, and my Tesla handles daytime good-weather non-construction highway driving pretty well with just the standard autosteer. I finish a long road trip much less fatigued in a Tesla than I have in any other car. (Part of that is the seats. Tesla really sweats the details of making seats comfortable.)
But FSD is kind of a sad joke. Off-highway navigation only works in particular major cities, the car stops at green lights, tries to turn right from the left lane, etc. And Elon thinks that's worth an extra $15k. Nope. It really isn't, at least for me.
> An estimated 15% of injury crashes and 24% of property damage-only crashes are never reported to police (M. Davis and Company, Inc., 2015), while9% of injury crashes and 24% of property damage-only crashes are reported but not logged (Blincoe et al., 2015). Even with robust data, establishing the statistical significance of automated vehicle safety can be expensive. Kalra and Paddock (2016)demonstrated that establishing that an AV has a fatal crash rate equivalent to the national average with 95% confidence would require driving a fleet of 100 vehicles continuously for 12.5 years.
I presume there's an adjustment for this in the figures you compare to, but I am having trouble finding it, though that may be my own failure so I wanted to ask if you could help me find how that's adjusted for?
It never got accepted for publication because the reviewers didn't think it was significant enough, but I wrote a paper presenting an alternative way to analyse autonomous vehicle reliability data (you'll have to take my word for it, but the reviewers didn't quibble about the results presented, just whether it was important):
https://arxiv.org/pdf/1812.08901.pdf
The major technical contribution was demonstrating that a class of mathematical models called software reliability growth models are a reasonable way to estimate the "anomalous behaviour" rates of an AV system as bugs are found and fixed.
My second point was that using fatal accidents as a way to measure safety is not necessary; you can use more common events such as non-fatal accidents, or in the case of AV systems in development, even the number of occasions the safety driver overrode the automatic control.
> In this study, all disengagement events listed in the company reports were included in the analysis. It is likely that in some cases, disengagements occurred as a precaution by the safety driver, and no adverse event would have resulted if the AV continued in autonomous mode.
Take a look at [1]. It says that:
> Waymo has developed a robust process to collect, analyze and evaluate disengages for this report. We set disengagement thresholds conservatively for our public road testing. The vast majority of disengagements are not related to safety. Our test drivers routinely transition into and out of autonomous mode many times throughout the day, and the self-driving vehicle’s computer hands over control to the driver in many situations that do not involve a failure of the autonomous technology and do not require an immediate takeover of control by the driver.
> To help evaluate the safety significance of disengagements, Waymo employs a powerful simulator program. In Waymo’s simulation, our team can “replay” each incident and predict the behavior of our self-driving car if the driver had not taken control of it, as well as the behavior and positions of other road users in the vicinity (such as pedestrians, cyclists, and other vehicles). Our engineers use this data to refine and improve the software to ensure the self-driving car performs safely.
So effectively, Waymo has secret process, which may change all the time, to decide which disengagements to report and which not to report. There is no way to tell whether the car would have slightly bumped into another car, or killed a child if not disengaged.
Their reports to the DMV do not constitute data, they are PR.
Not trying to single out Waymo, they all do that.
[1] https://web.archive.org/web/20190531071448/https://www.dmv.c...
Drivers driving without a valid license, without insurance, under the influence, wanted by the law, etc would have all kinds of reasons to not report a crash.
I also used to know a woman who thought it was nothing more than a fender-bender, went to the doc a couple of days later because her neck was bothering her. Doc sent her straight to the hospital--incomplete cervical fracture, one wrong move and she could have dropped dead.
In some cities currently even if you want to have police respond to a traffic accident, you'd need to wait a VERY long time - they will tell you that unless you need EMS to try and get a safe spot and call for a tow. Some folks aren't willing to wait hours for police to show up to do what, take a report?
I guess this is the classic scenario where human intuition is defeated by carefully presented statistics.
Adaptive cruise control, tech we've had for what, 20+ years? is more than enough for most highways, it takes out ~80% of the stress/boredom of driving, especially since lane changes are minimal and curves are gentle.
The hard parts of driving (aka not highways) are the real problem, as expected. We solved the easy parts in 2000 and who knows if we'll solve the hard parts by 2050.
You know what would be really nice? Something that detects when it's safe and reasonable to suggest turning on adaptive cruise control so I use it more. Maybe something like detecting I've driven a few miles at a consistent rate of speed above a certain threshold.
The hard parts are definitely the problem, but there are lots of easy wins left on the table too.
I definitely hope your car doesn't put it in some hidden menus on the central screen.
The hard parts are the whole point of autonomy.. the parts so hard that humans still fuck up and kill people. that's the ideal of autonomy imo and short of that its missing the point.
I wish Tesla would celebrate that victory rather than double down on lies.
Is there any evidence that they actually do the easy stuff better than humans? I would want to see it compared with cars of similar age and cost driving on similar roads with similar aged people in control of the vehicle. I'm pretty skeptical that such a comparison would in fact be favourable to Tesla.
I don't know what else to advise people who want evidence. The cars are real things you can experience...
In this case, the effect when the data is separated out is that AP is only marginally safer. When they are combined, it appears to be significantly safer.
I'd say this isn't a case of simpsons paradox at all.
If you're still surprised by this you've not been paying attention to how Telsa operates.
The study seems to extrapolate from this. Tesla has something like 10 billion miles total driven on Autopilot, on a fleet of a million cars.
It covers 5.4 million trips in 3,400 vehicles with digitally collected records and vehicle data and has been independently validated by the Virginia Tech Transportation Institute and Department of Transportation.
I actually wonder if there's another effect: Drivers only eligible for beta FSD if they meet a super high safety score (which isn't a perfect metric, but probably better than nothing). That incentivizes good driving even when FSD not engaged. Of course, that's a recent thing and hasn't shown up in data.
That's because it's marketing, not science.
Every car company does marketing. Few make over-the-top safety claims about their products. This is the company that decided to have its cars not stop at stop signs. Normal rules don't apply to Tesla.
https://www.bbc.com/news/technology-60230072
"In recall documents, the electric vehicle maker notes that if certain conditions are met, the "rolling stop" feature is designed to allow the vehicle to travel through all-way-stop intersections at up to 5.6mph without coming to a complete stop."
Kind of like in many jurisdictions, police will gently talk to you if you drive 60kph on the freeway, or even if you're doing 90kph in the left lane. Yes it's legal, no it isn't safe.
And 5.6mph isn't rolling. That's beyond a walking pace. That's jogging territory. Any cop watching an intersection would ticket this.
But let's make it real, and let's ask a question about the real world:
Does that mean that expectation is all automated cars will go exactly the speed limit on the left lane of North American superhighways?
If so, those cars will inherently be a danger to life and limb. I don't care about any self-righteous driver who indicates houghtily "I drive speed limit on left lane of American super highway", I've spoken to advanced driver safety instructors, highway police officers and city councilors who all agree those people need to bloody well move it along - it's just not safe.
Most cars idle at 5 to 6 mph, so any roll through a stop sign is likely to break the 5 mph mark.
> Any cop watching an intersection would ticket this.
Not sure where you live, but this is absolutely not the case where I am. I have only ever once heard of a traffic officer ticketing for this. On the other hand, I see cops watch on as people ignore "no turn on red" signs and let them get away with it pretty much every day. Let alone rolling through an intersection at 5 mph.
(Part of the reason I do it is because as I age I would like to ingrain habits that will make me a safer driver even as my cognitive ability declines.)
Near me there is an intersection where the same cars drive through on a daily basis and where the drivers have habituated themselves to rolling stops. Yes, it's almost always fine. But I have been almost T-boned twice, and was hit once, fortunately with minimal damage. And even though they do not have the right of way, their habit of rolling stops regularly pre-empts the actual rules of the road, and they cut off drivers who have the right of way.
That this is due to the normalization of deviance is abundantly obvious.
For some reason I still remember this story decades later and always come to a full stop at stop signs, that's how I was taught in driving school and that's how I will always do it. I'm a calm driver and I don't care if people are annoyed by me, I will drive the speed limit and I will not rush anywhere, it's not my problem that someone else needs to speed because they woke up late, safety should be the number one priority when you're operating a vehicle.
IDK about the US, but in France the speed limit is a target speed; if conditions are good (weather, visibility, traffic) you're supposed to drive at or near maximum speed. It's illegal to drive too slow, and there's a specific fine for this.
https://www.legifrance.gouv.fr/codes/article_lc/LEGIARTI0000... [in French]
It doesn't say that the speed limit is a target, or that you're supposed to drive at or near it.
You can get ticketed for driving too slow, but I’ve never seen it. I’ve only seen a minimum speed posted on an interstate (45 mph (72 mph)), but conceivably you can get ticked anywhere for impeding the flow of traffic.
- not exceeding posted speed limit
- appropriate for environmental conditions
- your ability to safely control the vehicle
And failure to do so could be fined.
Most retailers make over-the-top claims regarding their products, including safety.
https://www.latimes.com/business/autos/la-fi-honda-fine-2015...
It's a driving profile which must be turned on that implements behavior extremely common among the driving public. My Honda with Traffic Sign Recognition and adaptive cruise control lets me set the speed above the speed limit, too.
Our traffic laws are in many ways ambiguous, such that even the police officers enforcing the laws will readily admit you're allowed to drive above the posted speed limit. When traffic rules are interpreted ambiguously, it's not that strange for a driving profile to do the same.
The trick is to realize when you've jumped the shark and are now the established brand and wean yourself of the "fast and loose" approach.
And because it’s Elon Musk, not some ethical CEO
The "average driver" involved in an accident includes drunk drivers, road ragers, habitual speeders, teenagers, people driving in bad conditions (rain, snow), people driving without sleep and several other risky groups that most people buying Teslas aren't part of. To be worth it, the car has to reduce my personal accident risk under the conditions I usually drive, otherwise I prefer to keep my own hands behind the wheel over handing over control to an algorithm.
>The "average driver" involved in an accident includes drunk drivers, road ragers, habitual speeders, teenagers, people driving in bad conditions (rain, snow), people driving without sleep and several other risky groups that most people buying Teslas aren't part of.
... what? Tesla owners absolutely can be habitual speeders, people who drive without sleep, road ragers, and drunk drivers. Can you please explain your thought process behind this claim? I really don't understand why anyone would assume that Tesla owners are less likely to have those traits than non-Tesla owners.
To use an analogy: the average person has 1.9 legs but a product which results in its user having 1.95 legs is not an improvement for most people.
The average Tesla owner may well be those things. But I know that *I* am not those things.
I know my own driving record well. I know my own accident history well.
So you don't have to convince me that Tesla is better than the average driver. You don't have to convince me that Tesla is better than the average Tesla owner. You have to convince me that Tesla is better than *me*.
Now, apply that same analysis to the 88% of drivers that think they are above the median [1] in terms of driving safety. It doesn't matter if I am _right_, it matters what I think about my relative driving safety.
Moreover it is true in general that most people driving aren't drunk or high, aren't reckless etc., and so most Tesla drivers aren't either. The number of accidents isn't uniformly distributed among the public.
More broadly though, I never said that zero Tesla owners (or any other specific car drivers) do any of these things, but that using autopilot with an above average safety record can still mean that the personal safety of an individual or group of individuals goes down. For the system to be a net benefit it has to be pushed to the below average drivers.
If your driving skill is above the mean (because you're never a road rager, speeder, etc.), then you're worse off using Autopilot, even if Autopilot is as safe as the average driver.
Since most people probably consider themselves above average drivers (and in fact most people may be above the mean if the bad drivers are outliers), this limits the number of people who will believe that Autopilot makes them safer.
I think it was a comparison to Autopilot rather than Tesla owners. Robot drives can't be drunk, don't need sleep etc.
The claim seems to be that Teslas do not drink before driving, not Tesla owners, right?
So at the end of the day, you have spotless records for the autopilot that was driving many many miles on the straight line and humans having accidents at short intersections or construction sites. This translates into very favorable numbers for autopilot when presented as accidents per miles driven.
Case A: Non Tesla Human drives 999 easy miles and 1 hard mile
Case B: Autopilot drives 999 easy miles, and human drives 1 hard mile
If the effect size of the hard mile is so large that its skewing the statistics, you would expect the Telsa human driver to have a horrendous per mile accident rate relative to the non tesla human driver. What's most likely is that the accidents following a human handoff get correctly allocated to the Autopilot and the effect size being described is not actually that significant.
Or that that negative effect is lost in the orders of magnitude of "all human drivers across all miles" versus FSD.
Tesla tends to say that autopilot crashes occur 1/2 as often as non autopilot crashes. That's likely not normalized to road conditions. But if you assume that Tesla is secretly just putting all the hard miles on the humans, then that would imply humans are driving many more hard miles and should have higher accident rates. The autopilots meanwhile must be performing worse on the easy miles and racking up additional accidents that wouldn't have otherwise happened.
If you combine those two, the overall rate of accidents should be higher than average, but it's actually lower by a fair margin. Again, normalization is hard.
Ideally you would be able to compare human drivers of another comparable car brand to the human drivers of Tesla to confirm the Tesla drivers don't seem to be being judged on unreasonably difficult conditions.
Tesla Autopilot is judged only by the miles driven without disengagement, which is quite limited actually. You can watch Youtube videos to see at what kind of situations Tesla autopilot gives up.
There’s no situation where the Autopilot takes over from the human saying “That’s a tricky road, let me handle it”.
The car enables a completely novel driving style which people need getting used to.
This is a fallacy. Ask any demographic you like if they're road ragers, drunk drivers, etc... and they'll deny it. Everyone is sure that rare and terrible outcomes will never happen to them because of their own behavior, including the people to whom it happens!
It's probably true that middle aged premium car owners are safer on average, but you don't get to rule that stuff out by fiat. In fact there have been a few "Autopilot Saves Passed Out Driver" stories over the past few months, where Tesla drivers were clearly impaired.
As for this particular paper: this is just P-hacking folks. You can't take a vague dataset[1] and then "correct" it like this without absolutely huge error bars. Why correct based only these variables? Why did they all push the results in one direction? Why not gender? Why not income? Why not compare vs. like cars?
This isn't good statistics, it's just more statistics. If we want the real answers we should get a better data set (which, I'll agree, would likely involve some regulatory pressure on manufacturers).
[1] And, to be fair: Tesla's safety report is hardly comprehensive and provides no data other than the aggregate numbers.
That ship has already sailed; Tesla makes active safety claims based on that dataset. To hold research to your standard here would be to say that Tesla can make its claims, but nobody can challenge those claims.
> Why did they all push the results in one direction?
If you have two correction steps, then a priori you'd expect a 25% chance that they both act in the same direction. I don't think this is very remarkable.
> Why not gender?
This probably would be a reasonable addition. I doubt it would change the results much, but we have the well-known fact that insurance rates differ for men and women, so it may be relevant.
> Why not income?
First, is this data even generally available? If the data doesn't exist, then we can't control for it.
Second, should we expect crash rates per mile driven to differ greatly by owner/driver income? A priori, I wouldn't think this demographic quality to have a strong impact.
> Why not compare vs. like cars?
I think "personal vehicle" is a reasonable comparative category. Would we expect collision rates to differ greatly between more specific categorizations? For the sake of the overall conclusion, would we expect Tesla-equivalents to be particularly crash prone in the broad dataset?
Any statistical analysis will have its limitations, but when we're talking about life-safety claims from a manufacturer I think we should have wide latitude to look at critical evaluations of the original data.
I don't follow that logic. Claiming that Tesla's data set is incomplete is fine.
That doesn't mean that any junk science that refutes it must be correct. Bad statistics is bad statistics. This kind of "correlate dissimilar data sets by using more dissimilar data" is a terrible way to do science. Quite frankly it's almost guaranteed to be wrong. There are regulatory bodies out there with access to real data. Get them to figure it out.
OTOH: Tesla could still claim that if you replace all the distracted/drunk drivers in beat-up cars with new Teslas running Autopilot, the roads would be safer.
If we're not gonna switch to sane urban planning, may as well bring on the robo-taxis that can go anywhere
E.g. I would not be at all surprised if drivers of high-end BMWs have lower deaths per mile due to older population, more affluent, etc.
I'm not going to boast about how good of a driver I am (I am!!), but lets say I am a median driver. I would want a FSD to be better than a median driver. Not the average crash/mile stats that include all kinds of long tail idiots doing the most damage.
* ... alcohol-impaired driving crashes, accounting for 28% of all traffic-related deaths in the United States.
* Drugs other than alcohol (legal and illegal) are involved in about 16% of motor vehicle crashes.So 44% intoxicated to some degree (unless the stats overlap). That number is not even including reckless drivers not driving under influence, teenagers that just got the license, bad health elderly etc.
* I have even stared at this guy struggling to get out of his fancy Jeep who opened his door into my car. I know this anecdotal, but all it takes is a few counter examples to show that car price etc does not have a super high correlation to how careful & considerate you are.
Driver and buyer/owner are different things. Few teenagers ever buy a Tesla, but they certainly drive them. A car cannot only be safe in the hands of the rich initial buyer who lives in a nice climate. It must be safe in all the other people who that owner may let drive. It must also be safe for subsequent second and third owners, people who might not be wealthy enough to stay home when it rains or when the autopilot thinks conditions are too rough. (It was -35c on my drive to work this morning. Dark. Ice fog then blowing snow.)
Until someone pointed out that vehicle registration... disagreed. By nearly 30%.
Then "Oops. We made a mistake and counted 3,000+ deliveries which haven't actually been, well, delivered."
I hope they get spanked hard for that. Australia has very onerous "truth in advertising" laws.
Plus in your hypothetical the autopilot can drive at least as well as me, which is also a big assumption and not something I believe with what I have seen so far.
The damage done in a car distribution is heavily skewed towards people that get in fatal crashes or total a car. I'd randomly guess like 20-30% of drivers are below mean. Most people won't total a car, vast majority will never kill anyone.
If you get median people into an autopilot car that's got a mean safety record we end up pulling the average down, the road becomes less safe. And luxury sedan buyers tend to be one of the safest demos on the road, which makes the problem even worse.
This has come up before. When the new FSD beta started, people started claiming that safe driving was a function of vehicle price, and therefore Tesla drivers, especially those who had paid for FSD, were more likely to be safer. When I noted that my current vehicle costs more than a Model S, based on the logic, Tesla should be recruiting me to beta test FSD, well it was hard to find a refutation.
That's quite an interesting observation, and it shows the importance of proper statistics. It also raises questions about companies producing their own stats, it's often only when someone takes the time to dig into it that we discover this kind of thing.
I am not holding my breath waiting for the common consumer to understand statistical critical thinking though. I sometimes catch myself forgetting the due diligence even after many years of forcing myself to think about stats.
Perhaps some sort of ombudsman could do this, pointing out when stats are lying to people.
Companies saying "trust us, we have the data to prove it" while keeping that data secret should be ignored. Especially if their CEO has a long history of blatant dishonesty! I'm most annoyed that people gave any weight to Tesla's safety statistics claims in the first place.
TSLA stock consistently made gains for a lot of people who otherwise wouldn't be paying attention let alone putting their weight in anything Tesla related.
While true, I do not think this is a great heuristic. How would the average person know whether the "data" is available? Most people don't even know what "data" really is. They have never heard of CSVs or SQL. We need a better signal for consumers to know what they can and can't trust.
I think it's a good default to assume any self-beneficial corporate message is false, absent evidence to the contrary.
What truly makes a difference is the kind of road they're driving on. The accompanying paper provides a figure showing Autopilot is essentially only used on freeways which are inherently safer. Adjusting for that seems to explain most of the difference.
Though personally I'd prefer to have just compared the accident rate on freeways for both types rather than this weird weighting method they try to use.
A number of these near-crashes seem to be pretty odd errors (a couple of times it just decided to drive on the wrong side of the street?) -- albeit ones that one _could_ see drivers unfamiliar with the area making.
If a human driver started driving on the wrong side of a street, etc -- at least in Boston -- they would be greeted by a chorus of horns and rude gestures from those around, which would hopefully indicate to them that they were in error. Brings up an interesting question: does anyone know if Teslas are monitoring for surrounding horns/etc?
Another common problem in Boston is being blocked by geese crossing the road. Are Teslas able to detect such low, slow-moving obstacles?
It seems like it's not super feasible to rely on self-driving vehicles in a dense city like Boston, but maybe such features could be geofenced to areas like highways?
https://twitter.com/ButchSurewood/status/1488602615407681541
Isn't the whole idea that you're not doing anything?
To give you an idea of the incredible features in some cities, I have recently driven in Rennes where they found a good idea to makes roads with one single central lane for cars with wide bicycle path on each side. You are supposed to drive in the center and if a car comes in the other direction each car needs to go on the bicycle lane on their respective side of the road while giving priority to bicycles. I wonder how FSD would handle that...
Same with Summon. Marketing: "Bring your vehicle to you while dealing with a fussy child!" Legal: "Maintain attention and focus on vehicle at all times. Do not use while distracted."
Most regulated products require post market surveillance. You have to actively monitor the market and how consumers behave. If they have a misunderstanding about the intended use of your product, you're overpromising. Which is clearly the case here if Tesla did their post market surveillance correctly.
How do you mandate software quality? I don't think testing alone is enough if you seriously care about safety. Any test can be gamed as tesla has aptly demonstrated so far.
But on the other hand, even the most pessimistic examination of this data doesn't seem to suggest that these cars are below the "can get a drivers license" threshold.
When these regulators get swayed by the car industry into stupid stuff like 25 year import rule, or how they still keep speed limits around which have nothing to do with safety as they are based on outdated MPG savings initiative and end up disproportionally affecting lower income people, they rightfully are criticized.
However, as soon as some manufacturer pops up and does something that the public doesn't like, even for silly reasons like not protecting people from themselves, people immediately jump to wanting the "regulators" to do something about it. Kinda makes it easy to see who has irrational hate for Tesla/Musk versus those who actually care about technology.
The correct thing for the "regulators" to do is to make it mandatory that every car has the same set of cameras and instruments that Tesla cars do, and that data is streamed to a publicly funded data warehouse where its open source and available for anyone to use.
> Kinda makes it easy to see who has irrational hate for Tesla/Musk versus those who actually care about technology.
Another broad brush stroke and false dichotomy. It's entirely possible to care about safety before caring about TSLA's shareholder value, which has nothing to do with "irrational hate" or not "caring about technology".
For example, you won't see me complaining about Waymo's self-driving cars, because their program prioritizes safety instead of moving fast and breaking things/bodies. Does that sound like not caring about technology, or can you only care about technology as long as you're completely uncritical of Tesla's blatant disregard for safety?
> The correct thing for the "regulators" to do is to make it mandatory that every car has the same set of cameras and instruments that Tesla cars do, and that data is streamed to a publicly funded data warehouse where its open source and available for anyone to use.
That's just your opinion, and it's certainly convenient for Tesla. The "correct" and ethical thing for Tesla to have done is to keep their beta testing to closed tracks where all participants agree to be experimented only after giving their informed consent and just compensation for the risks involved. Instead, the entire public is being experimented on without their consent, and they're being experimented on by untrained non-professionals.
The "correct" and ethical thing for regulators to do is ensure that Tesla ceases experimentation that can get the public killed and maimed, that the public never consented to.
Similarly, the "correct" and ethical thing for Tesla to do is to not try to portray their Level 2 autonomous system as being Level 4 or greater. Regulators, ideally, would handle that issue, too.
Tesla clearly outlines what the autopilot is and isn't, and is not forcing people to use it. If you don't wanna beta test, don't use it. Simple as that. No need for any regulations in the first place. If someone thinks that autopilot means they can go to sleep being the wheel, and they crash, its is 100% their fault, not Teslas, or regulators. You don't blame mountain bike companies because someone decides to take their bike of a jump, go tail up nose down, and smack into the ground, even though their advertising often shows people doing big jumps on said bikes. Be consistent in what you believe.
Also, as far as self driving goes, Waymo isn't as much as pushing for safety as it is burning through VC cash in a giant scam. It started 13 years ago and has zero products. Meanwhile, startup shop Comma AI managed to get self driving highway autonomy system out in 5 years that you can actually buy, on a fraction of the budget.
> However, as soon as some manufacturer pops up and does something that the public doesn't like, even for silly reasons like not protecting people from themselves, people immediately jump to wanting the "regulators" to do something about it.
Weird how people want regulations that benefit the community and not regulations that are detrimental. Absolutely impossible to understand this.
- it’s tricky because the people on the roads expect human convention and drive for that
- as long as there is a mix of humans that drive by convention, it’s probably best to err on the side of following human convention to some degree
Case in point: center lane
When I first started using autopilot on two lane roads the car would stay dead center in the lane. If a car was coming towards my in the opposite lane, it began to notice humans would veer away from center to provide more buffer between themselves and the oncoming car.
Because I didn’t want to piss other drivers off, I would often disengage and drift to the right of my lane while the oncoming car approached me. If I didn’t do that, there was always a last minute extra drift away from me by the other car.. the conventionally expected buffer distance wasn’t enough to make them feel comfortable due to unexpected (lack of) behavior.
It’s not a law / no law issue above. But I have similar experiences navigation into roundabouts with crosswalks in front of them (autopilot stops at empty crosswalk where normal driver would cruise through until the stop line to check for cars in the roundabout — and if none were there might pass through with a rolling stop/check.
Doesn't the Autopilot sometimes just give up in difficult situations? Either explicitly, by stopping behind parked cars or things like that, or worse, by behaving so badly that the driver has to take back control for his own safety.
The worst case would be: autopilot doesn't see an obstacle, driver sees it at the last moment, takes back control, but it is too late, and crash. Is it an autopilot crash or a human crash?
Or: there is some road work, the autopilot cannot maneuver around it and gives up, the driver takes back control and crashes. It is clearly a human error, but still, the autopilot leave the short part where accidents are the most likely to happen the driver.
It things like that happen, it is almost as if the autopilot is to be used only where accidents are the least likely to happen, which can obviously explain why people on autopilot have less accidents.
It also helps statistics both directly: crashes are more likely in the snow and rain, but Teslas don't do self driving in those conditions (that's one of the things this paper addresses).
Tesla has in numerous cases blamed the driver for not taking control of the unsafe "self driving" when crashes occurred, how many near misses have occurred due to drivers intervening? Airlines and aircraft manufacturers are required to report events where a pilot how to take over to prevent an accident but Tesla and co don't
I didn't read the paper fully but grepped for "gender" and "sex" and couldn't find anything.
1: https://xtown.la/2019/12/12/women-drivers-maybe-we-need-more...
Controlling for confounding variables is a complicated and often subtle process that many people don't have an intuitive grasp of, so Elon's BS claims have probably been very convincing for many.
The reason this study has this outcome can be because of the survivorship bias - people are letting autopilot drive only in easiest road conditions where chance for something to go wrong is smaller to begin with.
I would have hard time imagining people are letting autopilot drive regularly in a tight, windy road (at night) or in snow, thus completely preventing those (frequent types of?) accidents from happening under autopilot.
On the one hand I had the same suspicion about the unadjusted data, on the other hand the fact that autopilot isn't worse is pretty promising!
However, average human rates include people driving drunk, tired, elders, teenagers etc.
So at 43 when I am driving sober, I'm probably safer not using autopilot. When I was 19 probably safer to use autopilot.
Presumably drivers take over when it makes mistakes, which tilts the stats a bit.
> When I was 19 probably safer to use autopilot.
As a parent, I wonder about what safety tech I would let my kids use when learning to drive. I want to make sure they fully learn how to drive, which makes me think I shouldn't let them use too much semi-autonomous tech.
At the same time, it seems foolish to tell them not to use blind spot monitors, or to even expect them to "check their blind spot" the old-fashioned way if their cars have monitors.
Interestingly, my father (in his 70s) tells me he'll never trust blind spot monitors, so I'm seeing a generational difference even between him and me. Too bad, since older drivers probably could benefit the most from not having to look back when changing lanes. Older eyes take longer to re-focus, so it's more likely an older driver would miss something happening ahead on account of having looked rearward.
I have never owned a car with blind spot monitoring. Do people really just rely on them completely? Even in dense urban environments with bikes around? Or is it more of a judgement call? Different perspectives on this would be nice.
I have never seen the little light on the mirrors fail to detect me while I am riding but I never trust a driver not to just turn right cutting me off anyways.
I am not trying to pass judgement just understand how these driver aids change the way people drive.
As part of my job, I end up renting vehicles pretty frequently, they're often newer models than my daily and have blind spot monitoring, backup cameras, radar cruise, vision-based lanekeeping assist, etc. It's frightening how in the course of a week you can get used to having to put less effort into centering yourself in the lane, trusting that cruise control will just keep a comfortable distance from the vehicle ahead of you. Driving becomes a lot less stressful. When I get back home and climb into my old manual-everything beater, though, it's quite an adjustment.
Regarding blind spot checks, yes, the Audi I recently drove had alerts for that, the blinking yellow light in my peripheral vision when cars were passing me was a nice reinforcement, but I'm too conditioned to do head checks to skip those. Likewise, reverse cameras - I've driven many work vans, pickups with headache racks, Jeeps with scratched up plastic, pulled trailers and RVs, etc where the windshield-mounted rearview mirror is useless; lots of pros get used to backing up using the side mirrors only. However, I asked my sister in law (who is extremely competent at most things) to drive my truck for an errand and she asked for help backing it out of the driveway - her car has always had a backup camera, which is honestly lots easier and she was completely uncomfortable using the side mirrors.
It's not hard to imagine that someone who only drives with assistive tools would adapt to become dependent on them; I'd argue it's more unusual to expect that they wouldn't!
Driving for ~25 years. Did quite some kilometers when I was between 19 and 30.
My current car is the first one with a lot of assistance systems. I actually see them as additional safety level, but as a first instance still use my senses. Blind spot monitoring is something that I learned to value as a slice in the Swiss Cheese Model of Security [1].
And I detect myself driving extra carefully when driving a car without them. Way more careful than I was driving before I ever had blind spot monitoring.
[1]: https://en.wikipedia.org/wiki/Swiss_cheese_model?wprov=sfla1
Some Hondas pipe through a camera feed from the side of the vehicle when you turn on your blinker, which wouldn't have the same potential weakness for cyclists that you mention with the binary sensor-based systems.
But I also grew up and had my formative driving learning years without the benefit of such aids.
That’s very dangerous thinking. Autopilot isn’t self driving and will actively try to kill you, from time to time, like all driver assist systems. 19 year old are much more likely to abuse autopilot and drive distracted/drunk. Only reason those systems aren’t crashing left and right is because drivers are paying attention.
That was my first hand experience also, it feels like I’m the nervous dad gritting his teeth while the teenager does something technically legal and safe but nerve-racking.
That might work out for you, but if teens think like that _now_ and avoid years of driving experience, they can't expect to be any better at driving at 43 than they are as teenagers, which kind of kills the argument.
* It doesn't find anything to suggest Autopilot is less safe than claimed. It finds reasons to believe Autopilot's improvement on safety relative to humans is less than claimed
* Under both the road adjusted model; and the road+age adjusted model, Autopilot outperforms the average driver on average. Not by much, but its there.
* Everyone here seems to be assuming that the takeaway is that Autopilot is less safe than humans, but there's no evidence for that. If the numbers here are to be believed, I would take it as a good sign for tesla, and by definition an advantage to people who are otherwise poor drivers?
* Question: Why do crashes per million miles have such large swings in the baseline? It's weird that it varies so heavily by quarter without any obvious seasonal patterns to me.
edit: mildly annoyed at getting downvoted for seeming to be the only person here noting that the blue line is on average below the red line on every chart shown.
Either way, I'm not too impressed. I wouldn't purchase Autopilot personally for similar safety reasons. But let's not twist the stats to suggest they're actually much worse than they are.
Your own hands are literally on the wheel and you remain responsible and capable.
Obviously there are times when driving without it is way more fun. It's not like that option is taken away from you.
These downvotes have got nothing to do with you or with what you pointed out. Just some people here with sour grapes.
"The fact that Tesla consistently refuses to share full data on identical models driving with autopilot turned on, enabled but not turned on, and not enabled is extremely suggestive that the overall safety rates for autopilot on are not better."
Basically accidents involve two people. Said accidents have responsibility divided between both individuals.
A hypothetically perfect driver could still be involved in fatal accidents by virtue of the other driver.
The worse drivers obviously will cause these accidents.
The questions are three fold:
1. Are self driving vehicles safer than the median driver?
2. What’s the ratio between at fault and not at fault for any accidents the median driver gets in?
3. What’s the same ratio for self driving cars?
Normalizes for location, of course.
Not always, no. It's perfectly possible to crash a car with nobody else around.
While the most recent quarter could be an outlier, some other quarters show a similarly dramatic difference.
I'm also wondering about the assumptions the pre-print makes.
https://engrxiv.org/preprint/view/1973/3986
For example, it's assumed that Tesla's data has the same characteristics as the SHRP 2 NDS data, and that the demographic survey data from Hardman et al is accurate WRT the SHRP 2 NDS data.
> The analysis in this paper relies on the assumption that the freeway-to-non-freeway and age group crash ratios found in the SHRP 2 NDS are consistent with the manufacturer’s data, as there are no roadway specific nor age-related factors in the manufacturer safety report.
> Although the ages of drivers in Tesla’s crash rate data are unknown, their ages could be estimated from a 2018 demographic survey of 424 Tesla owners (Hardman et al., 2019).
But... It seems like Hardman et al are using California specific data.
https://escholarship.org/content/qt58t7674n/qt58t7674n_noSpl...
> In this study, we used a cohort survey of Plug-in Electric Vehicles (PEV) owners in California administered by the authors in November 2019. Respondents had been previously surveyed by the UC Davis Plug-in and Hybrid & Electric Vehicle (PH&EV) Research Center between 2015 and 2018 as part of four surveys in the eVMT project when they originally bought their PEV. Respondents for the four phases of the eVMT survey were sampled from the pool of PEV buyers who had applied for the state rebate from the California Vehicle Rebate Program (CVRP). More than 25,000 PEV owners were surveyed between 2015 and 2018. A total of 15,000 of these respondents gave consent to be re-contacted and were invited for the repeat survey in 2019. In all, 4,925 PEV owners responded to the repeat survey.
And... The newer data indicates that age could be positively with more Autopilot use and long-distance travel (also correlated with more Autopilot use), which seems like a confound WRT to the pre-print's conclusions and so on.
> Age is negatively correlated, indicating that the lower the driver’s age the higher the odds of reporting more long-distance travel. This suggests Autopilot induces travel among younger Tesla owners. As with the model with all Level 2 automation, users’ income is negatively correlated.
This is the real question, because a "perfect" autopilot may be impossible because of cases where it must choose between sacrificing something outside the car vs its occupants and there will be disagreement.
Setting the bar at "perfect" is pretty unrealistic given that humans are already pretty bad.
Seriously? you trying to claim your method is more fair using 424 people in 2019, when Tesla has demographics of millions cars across the globe?
My conclusion:
Author is saying it's not fair to compare with some old car. Tesla was one of the first brands to mass adopt this safety features in every Telsa. This is the only brand that include camera, collision warnings in every, even cheapest modification. And they were the first to introduce many of these features. The rest are catching up. Why Tesla can't be proud of that? They give safety to every customer, no matter if it's the cheapest option or $150k+ car. Only volvo invest on par with tesla in testing facilities, in house extended crash tests ets.
Autopilot and similar system is safer compares to any car on the highway with old cruise control because they have features that force you to be focused. They can stop you on a traffic light, they can react faster on animal running across the road etc. And Yes all measures like collision warnings, notifications, makes your driving safe. Yes, you can use it without autopilot, but I as a consumer don't care. If I'm buying a car, fact that this car is much safer compares to 10 years old ford makes me more confident in purchase.
You can debate a lot on how easy is to avoid some of this safety measures. But if you an idiot you can fall asleep in cruise controlled car as well. Tesla will at least try to wake you up. And will stop on a traffic light.
FSD is a good example of an extension of this rational vs emotional thinking. FSD Beta has a very strict focus control. You look not on a road few times you're blocked. You don't hold a wheel few times you blocked. Etc.
So even FSD beta in it's current state is a complete garbage that quite often literally drives you into accident the way they enforce your attention fixes everything.
FSD beta with this limitations has 60k cars on a road. And 0 major incidents that lead to a major injury or death. Not counting few scratched rims and one time driver got off the road because he over reacted on some maneuver.
It's bizarre how much money legacy auto is poring into efforts to make tesla look bad. Before it was infamous data of burning EV's. But if you look at data a single car brand BMW has so 100 times more risk of fire compares to tesla. You don't see paid articles on Times square how dangerous is BMW.
Right now all major brands invest in EV and suddenly all this fire incidents disappear. And they switched to Autopilot and FSD.
Is there a room for improvement yes. Is the car that can stop on a traffic light, after some fixes for a cop's car, notify about potential collision, avoid automatically some collisions safer. Definitely yes. The rest is just hype, speculations and quite often corruption and paid articles.
Google this guy before you buy any of his spin.
Not saying that Tesla shouldn’t be more transparent, but Niedermeyer is famous for how wrong he has been on Tesla on virtually everything. Like many Tesla partisans (not pro and anti) went way beyond rational discourse long long ago.
In case you missed them, there have been 5 tesla stories nuked from the front page today:
'Self driving' Tesla fails miserably on the streets of South Bostonhttps://news.ycombinator.com/item?id=30173566
Tesla Australia admits its sales figures were wronghttps://news.ycombinator.com/item?id=30176635
Tesla drivers report a surge in ‘phantom braking’https://news.ycombinator.com/item?id=30179681
Self-driving Tesla does 'the craziest things you can imagine'
Expressing a view in either direction results in an instant pile-on of trolls and flames.
The reason HN is exceptional is that it has a lot of readers working on or near AI, who will quickly call out the B.S. in Tesla PR. They know this, so all Tesla stories get nuked.
(1) Statistics obtained from Hardmanet al.(2019) (2) Statistics obtained from Blanco et al.(2016) (3) Statistics obtained from Transportation Research Board of the National Academy of Sciences (2013)
I don't understand how such chronologically disparate sets of data could produce reliable statistics. The Telsa demographic data (1) was from a "2018 demographic survey of 424 Tesla owners"
I’ll wait for the peer-reviewed version.
I don't think his request is out of the ordinary here:
https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que...
You'd make the comparison of cruise-control crash rates vs. general driving in all conditions if you were trying to abuse statistics for marketing your gimmick cruise-control feature.
It's unclear to me how you'd even attempt to normalize the data when you don't have national crash rates for cruise-control only miles in non-Tesla vehicles.
I saw mention that one of the TLA's investigating Teslas crashing into emergency vehicles subpoenaed all manufacturers of vehicles sold w/L2 cruise-control features for their safety data. Maybe that will produce some normalized-enough data for making a meaningful comparison.
Autopilot failure statistics are non-ergodic. You don't get to "play again" after a catastrophic Autopilot failure.
So, Autopilot must not be 5x or 10x better than your own driving -- it has to be, like 100x or 1000x better, in order to warrant you risking your life on it.
Trusting an autopilot that is just 10x better than you are?
Insane. It randomly drives under a semi-trailer that you would have easily seen and avoided, and you're dead.
So if humans are 2 times more likely to crash on non-highways than highways, the argument presented here is that the added danger of non-freeways affects autopilot in the same way, degrading its performance. Seems like a far reaching assumption.
If there is no data on autopilot performance on rural roads, then you cannot make a claim either way.