Teslas running Autopilot have been in 273 crashes in less than a year | Hacker News Reader

Teslas running Autopilot have been in 273 crashes in less than a year

washingtonpost.com

77 pointsboplicity179 comments

The actual, non-editorialized data, along with the two actual NHTSA reports (one for level 2 driver assistance systems, one for self-driving systems in levels 3-5), are here:

* Overview: https://www.nhtsa.gov/press-releases/initial-data-release-ad...

* Data: https://www.nhtsa.gov/node/103486

* Report on level 2 systems: https://www.nhtsa.gov/document/summary-report-standing-gener...

* Report on level 3-5 systems: https://www.nhtsa.gov/document/summary-report-standing-gener...

Going forward, the NHTSA will release data updates monthly. That's exactly what we need to judge progress with these technologies: data. Bravo!

Everyone: Let's at least skim the data and read the reports' summary conclusions before we start sharing our opinions.

The newspaper owned by the billionaire that hates the billionaire that owns Tesla is probably not the most reliable source :-D

Considering the newspaper owned by the billionaire runs stories criticizing said billionaire's company's anti-union efforts [1], mayhap there's editorial independence?

[1] https://www.washingtonpost.com/technology/2022/06/13/amazon-...

Whatever the motives (from whomever), the smear campaign from WaPo against musk is obvious enough that I've seen many people talking about it and even a twitter post detailing the many articles painting musk in a negative light and comparing that to other sources which if the data is taken at face value was very damning of WaPo.

Personally I don't really care enough to investigate, but I thought it might provide some context as there appear to be (what I would assume) neutral parties commenting on the unfair reporting on musk from WaPo.

> obvious enough that I've seen many people talking about it and even a twitter post

If you're accusing an investigative journalism outfit of bias, you'll need better evidence than "people talking about it and even a twitter post."

> If you're accusing

I'm not, I agree that if I was then I would need evidence.

> Personally I don't really care enough to investigate

but you do care enough to comment on the internet with no evidence, thus your context is just disguised misinformation mixed with the hubris that your opinion matters.

Geez, some people really go out of there way to make things hostile. My opinion does matter, just ask my mom!

I simply gave people some context and terms they can google if they're invested in this story. I'm not invested, and really don't much care if wapo is or isn't targeting musk. It doesn't affect my life one bit.

If you think my comment isn't valuable, downvote and move on - that's what the arrows are for.

It's not just that your comment had no value, it had negative value. If it only had no value, I would have done as you said, downvoted and moved on. But since it's existence was a detraction, I felt it was important enough to attempt to mitigate the propagation of potential misinformation.

I do apologize that adding the bit of "... hubris that your opinion matters" came across rather harshly and was unneeded to my point.

> thus your context is just disguised misinformation mixed with the hubris that your opinion matters

This is a really silly thing to say. You're posting your opinion. Is that hubris?

No, because hubris is about having "excessive pride", and it doesn't take much pride to decide that I can parse the English language and determine that the person I am replying to has made a fallacious statement.

I think this is least of the worries. WaPO has extreme left bias which while reporting the truth, omission of important facts as they see fit for their readership.

HN’s left bias matches WaPO in perfect resonance. So, it’s difficult to even discuss it here.

Show me an example of “extreme left bias” on WaPo…

For real, I don't think OP understands the term. Classically liberal is right wing, not left.

Classical liberal is only right wing if you slide left so hard you break the overton window.

Isn't this kind of misunderstanding what the overton window is? The window doesn't pay attention to the content of ideologies per se, just where a culture's sentiments gravitate towards. It doesn't in itself define political beliefs. At some point you need to actually care about principles!

It's right wing in the sense of individualism vs collectivism.

As someone with an actual "extreme left bias" claiming the WaPo is on the "extreme left" is hilarious. Using American political terms, I'd put the NYT as having a centrist Democrat bias, the WaPo as having a centrist Republican position and the WSJ having a Republican bias. All of them are pro-capitalist and different wings of the neoliberal party and really right wing in the way the rest of the world thinks about politics.

I don't even understand this world where people can accuse the WaPo of being far left.

Yeah, they shat all over Trump, but that just reflects how far Trump went completely off the reservation. And it created the surreal situation where every other week I was agreeing with something that Bill Kristol said.

The fact that Trumpers feel the need to scream about how the WaPo is a bunch of communists really shows how intellectually bankrupt the whole thing is.

I'm going to ignore Trump comments because those are irrelevant here.

Quick Google search reveals several sources claiming Left bias. My personal opinion is that it is left of NYT. NYT opinion pieces are extreme left while the news are center-left.

https://adfontesmedia.com/washington-post-bias-and-reliabili...

https://www.allsides.com/news-source/washington-post-media-b...

Here is a the whole chart: https://www.allsides.com/media-bias/media-bias-chart

> I don't even understand this world where people can accuse the WaPo of being far left.

Recommend exploring more areas of the political spectrum. Not a single source claiming WaPo is center-right. If you can find a better source than your own delusion, I'd love to see!

> Quick Google search reveals several sources claiming Left bias

Why are these sources authoritative? Why is the chart you linked authoritative?

> NYT opinion pieces are extreme left

So when they published an op-ed by Tom Cotton calling for the mobilization of the military against American citizens, that was "extreme left"?

> Not a single source claiming WaPo is center-right

Not a single source /of your sources/.

> Not a single source /of your sources/.

There's no point saying this. Just post better sources if you have them.

You have no idea what "extreme left" means.

And the "allsides" chart with Jacobin lumped in with the NYT and MSNBC is hilariously bad.

Allsides itself was formed by a former Republican which is where its bias lies.

> And the "allsides" chart with Jacobin lumped in with the NYT and MSNBC is hilariously bad.

Yeah I can understand using a left/right scale more relative to American political parties, but that's like meme a tier placing. Left and right are already broad terms and shit like these makes it even more binary.

> off the reservation

Hey, maybe consider not using this phrase.

oh yeah that's a good point.

> Using American political terms, I'd put the NYT as having a centrist Democrat bias,

The New York Times is owned by NewsCorp / Fox News / Rupert Murdoch

You may want to reconsider how well you know this topic

You replied to someone that posted NHTSA links. This only makes you sound more whiny than if you had just posted this as a reply to the article itself.

The reply highlighted an issue that seriously affects interpretation of the data from the NHTSA link. Citing a source isn't enough if either the source or the commenter misrepresents the facts. There was no justification for you to be so insulting.

A similar article was also published in the LA Times, whose billionaire owner is friends with Musk.

https://www.latimes.com/business/story/2022-06-14/data-likel...

The level 3–5 report shows Apple as having a reported incident. Is Apple testing level 3-5 autonomous system?

This should be a headline of its own.

Can we trust Tesla (and other manufacturers) after they did this:

"Tesla eventually agreed to split the cost of the repair, but before they would do that, the owner had to sign away his right to discuss the defect, which could preclude him from reporting it to the National Highway Safety and Transportation Administration (NHTSA)."

Source: https://www.forbes.com/sites/lianeyvkoff/2016/06/09/is-tesla...

No. They are offering bribes. It's anti-competitive behavior in that it withholds information consumers need to make free and informed decisions. And it's fraud. Tesla gains financially from this deception. If offering the bribe is not already illegal, it should be made illegal. If it is illegal it should be prosecuted. By all means start out with a per bribe fine at least equivalent to the profit of a single unit, or else the fine is just a cost of doing business. Otherwise, treat it as a crime.

Here's some actual data I just pulled from the report, showing number of incidents when using a level 3-5 system, sorted in descending order and then alphabetically by reporting entity. Despite having the greatest number of FSD users, Tesla has only a single reported accident when FSD was engaged. Other level 3-5 systems with much smaller installed bases are doing far worse. The data for level 2 is further below.

  # of incidents, levels 3-5:
  Waymo        62
  Transdev     34
  Cruise       23
  GM           16
  Zoox         12
  Argo         10
  Ford          7
  May Mobility  3
  Mercedes Benz 3
  Pony.ai       3
  Easymile      2
  Toyota        2
  WeRide        2
  Apple         1
  Beep          1
  Chrysler      1
  Hyundai       1
  Local Motors  1
  NVYA          1
  Navistar      1
  Porsche       1
  Robert Bosch  1
  Robotic Rsrch 1
  Tesla         1
  TuSimple      1

Here's the level 2 data -- number of incidents, again sorted in descending order and then alphabetically by reporting entity. The biggest caveat with these figures is that the NHTSA does NOT know how the rate of adoption & usage of level 2 technologies varies by reporting entity. It's likely that Tesla and Honda have the greatest rate of adoption & usage of level 2 technologies, but we don't really know:

  # of incidents, level 2:
  Tesla       273  <-- figure quoted by the OP
  Honda        90
  Subaru       10
  Ford          5
  Toyota        4
  BMW           3
  APTIV         1
  Hyundai       1
  Lucid         1
  Porsche       1
  VW            1

Sources: https://www.nhtsa.gov/laws-regulations/standing-general-orde... -- click on "Reporting Entity (ADS)" for the level 3-5 data, and on "Reporting Entity (ADAS)" for the level 2 data.

EDIT: After a quick look at the report and the data, my take is that it's way too early to reach conclusions other than non-Tesla level 3-5 systems appear to have more accidents despite being used in many fewer vehicles. Shame on the Washington Post for publishing such an unjustifiably alarmist piece!

It would be interesting to see per 1000 vehicle owners or normalized in some ways.

Actually, it would be interesting to see the data per mile driven while using the driver assistance system, both for level 2 and levels 3-5, because to compare these systems apples-to-apples we would want to exclude all miles driven by human beings without any assistance. And ideally, I'd want to see data per mile driven in each road class (highway vs city, rural vs urban, etc.).

Agreed, your suggestion is a much better way to look at this.

Split out with and without driver assistance is fine, but I only care about the combined total. If there's higher accidents in X make/model vs Y make/model, that's a more useful data point. I don't need every detail on extenuating or contributing circumstances split out to decide the overall risk of the car itself versus the driver. Plus I don't trust any of the manufacturers to categorize accidents either fairly or consistently among each other.

Of course -- that's what the overall stats and safety ratings are meant do. But here we're talking about research comparing level 2 and level 3-5 systems.

I know there were multiple incidents where FSD disengaged less than a second before impact, so this may not be the most meaningful cut of the data.

Yup. A lot, relatively. "On Thursday, NHTSA said it had discovered in 16 separate instances when this occurred that Autopilot “aborted vehicle control less than one second prior to the first impact,” suggesting the driver was not prepared to assume full control over the vehicle. "

Exactly - they need to measure was FSD engaged at any point during the 30s prior to the crash. Tesla's approach of seeing that a crash is imminent and so then disabling FSD just before impact to hide their crash numbers is extremely scummy.

Waymo/Transdev (same thing): in over 50% of accidents car was stationary. 0 serious injuries or fatalities.

Tesla: 3 of 550 accidents the car was stationary. 9 fatalities, 6 serious injuries.

I'm sure Tesla are testing L3 systems, but I didn't think any currently offered to the public provided any more than L2? Even so-called "FSD" is still only a L2 system and requires the driver to maintain attention and be prepared to take control with no notice.

The NHTSA categorized the systems as level 2 and levels 3-5. You can read their rationale in the report.

To drive home the point about how unsafe Tesla' Autopilot is, Nissan has sold more than 560,000 vehicles with an L2 system...and had zero reportable accidents, which makes Tesla infinitely more dangerous than a Nissan.

Even viewing the data in the light most charitable for Tesla, Autopilot is 3x more dangerous than the next most dangerous L2 system, and 54x more dangerous than the companies that sell the most vehicles in the world.

Also, FSD is not an L3 system. It's an L2 system, so Tesla should not have any data in the L3-L5 section.

EDIT: Ars notes that Tesla's 273 crashes are only for the past year, since the NHTSA required automakers to report crashes, but the numbers for all other automakers are total numbers (meaning, since they started selling L2 systems). This makes the Tesla data even worse.

Exactly.

Im not sure how the Cult can justify this.

I get that Tesla are aspiring to lofty goals, but if your L1000 system (which is called L2 due to the need for driver supervision) can't compete with an L2 Nissan.... you may as well pack up and go home.

Tesla should have just stuck to what they're fantastic at. Driving the cultural and technological adoption of electric vehicles.

Anything else, from FSD to basic build quality is something I wouldn't touch with your barge pole.

And that's even before considering its man-child branding problem.

The number of incident per brand doesn't matter if you don't take into account how many cars they have on the road...

And Tesla could probably tell you how many crash per driving hour among other things

Thank you for linking to all of the data. Consumers have different priorities than society when it comes to these systems, "convenience" vs "safety" respectively.

If these systems are not improving safety above baseline (rate of accidents for non-assisted driving) it would be interesting to find out how this difference appears with each system's level of assistance. Then within those levels, are there certain platforms which pull the numbers in different directions (e.g is a Tesla more or less safe than other DAS of the same level).

If these systems are not resulting in a reduction of accidents, should we rethink what kind of 'driver assistance' should systems provide? If higher level systems are less safe, should the industry hold off on pushing those systems to market until they're developed further? If a specific platform is a significant source for the reported accidents, then what action should be taken to address? (disable those specific DAS already in market?)

Great questions. I don't think anyone has good (defensible) answers for them yet. This new effort by the NHTSA to compare ADAS and DAS systems apples-to-apples, in combination with more traditional efforts to collect overall data per mile driven (for both machine and human drivers), should eventually give us the answers :-)

This one from The AP [0] says Tesla’s figure and its crash rate per 1,000 vehicles was substantially higher than the corresponding numbers for other automakers that provided such data, though Tesla is the only one that does it real time and "Other automakers, by contrast, must wait for reports to arrive from the field and sometimes don’t learn about crashes for months." There's quite a few more details in there, the comparisons make Tesla look worse than the others, though I'm not quite sure by how much when you consider all the differences in reporting? (this might also be in the WAPO article, I couldn't read it)

0: https://apnews.com/article/technology-business-5e6c354622582...

Faster reporting shouldn't affect long time averages. If the delay is x months 100*x/(n+x)% of crashes go unreported at month n with delayed reporting under simplified assumptions.

The challenge is that the availability of these systems is increasing exponentially at the moment. Long term it matters less, as the long term reporting can include mileage. Short term, the lags introduce a lot of bias.

I'm thinking largely of how quickly Super Cruise and Blue Cruise are expanding here. Other systems may not be expanding as quickly.

An exponential also gets squashed by an asymptotic. That math gets a bit more complicated (You need to evaluate an integral) but that does not categorically change the math. Also i find the notion of an exponential ramp up quiet dubious, polynomial would make more sense.

That's fair, but going back to the report I had misread it. It actually includes a lot more cars than I thought.

However, it relies on reporting of accidents that is likely missing a lot of them. It is very likely that only Tesla's data is close to accurate.

One of the most contentious issues is that Tesla is able to provide a figure in terms of crashes per miles driven on Autopilot whereas other automakers can't do the same.

So they report crashes per 1000 vehicles, but that doesn't take into consideration how much people are using Autopilot vs how much people are using GM SuperCruise. If people with GM cars are hardly ever using SuperCruise but people with Teslas use Autopilot often, that's going to really skew the figure.

Sure, but is there a reason to assume (mere speculation allowed) that Tesla's software is being used more than GM's?

Maybe it sucks less?

I have no opinion, informed or otherwise on the matter. This is pure speculation. But it seems plausible.

It's more capable. Autopilot can basically do 100% of highway driving from on ramp to off ramp. It can switch lanes, and it can also take an exit from one highway to another highway.

Either way, without data about how often people use each driver assist feature, it's not scientifically rigorous to assume that each system is used exactly in the same proportion. The NHTSA should have made this caveat very clear, and should have released data from Tesla about crashes per miles driven if it's available (even if it's not available for other automakers).

Unfortunately the mile-level data simply doesn't exist. From the NHTSA: "Reporting entities are not required to submit information regarding the number of vehicles they have manufactured, the number of vehicles they are operating, or the distances traveled by those vehicles. Data required to contextualize the incident rates are limited."

Yes; the quality of autopilot greatly influences how much I use it. 5 years ago Tesla's was kind of nerve-wracking and I rarely used it. Now it's pretty good on highways but I don't use it on city streets. I probably will when the quality goes up. So I've gone through 10x changes in usage as quality improved.

I don't know what GM's autopilot is like, but I wouldn't assume that usage would be similar even within an order of magnitude.

If by "pretty good" you mean an absolute f'ing nightmare for the other drivers on the road, especially the ones driving behind the Tesla who must be prepared to react to it slamming on the brakes at random intervals because Autopilot misidentified a shadow on the road.

Autopilot is about as good as a teenage driver getting in the car for the second time.

I’ve used Autopilot (EAP, lane keeping, auto lane changes and highway transitions on intestates) more miles in four years (without an accident or death) than some people drive their car in it’s entire lifetime, and yet, my Tesla will only get counted as a single vehicle in the stats. Including miles is a necessary datapoint, statistically speaking.

I have two cars with technologies like this, Nissan ProPilot and Tesla Autopilot. ProPilot steering assist is useless on my most common two lane curvy roads and just annoying enough to not be worth it on some four lane divided highways.

Tesla Autopilot handles both of those same streets, which are my most commonly driven streets, very well. It gets used much more frequently. Anecdotally, I hear similar reports about other systems.

Great context. Impossible to know if this is good or bad without a point of comparison.

I can’t help but feel discouraged that other automakers will improve their reporting. It’s all downside, because you get articles like this, that inevitably build into narratives - true or not - that your cara are unsafe. Meanwhile, why bother when your deficiency is reflecting poorly on your competitor?

It's also worth noting that if Tesla is automatically collecting these crash stats from vehicle telemetry and other manufacturers are collecting them from customer reports or vechicle service calls, Tesla's numbers are probably inflated by a factor of three (or other manufacturers are deflated by three, however one wants to call that).

The first thing Waymo learned starting out on public roads from observing their own telemetry was that NHTSA numbers (based on reported crashes and service calls) are three times lower than actual collisions because there's huge incentive to not report a collision (insurance premiums). Comparing numbers from differing data-collection approaches is an apples-to-oranges comparison.

I care about nine numbers, in the order of the groups given.

    Group A
    1. How many deaths and serious injuries per standard distance does AP have?
    2. How many deaths and serious injuries per standard distance do other self-driving cars have?
    3. How many deaths and serious injuries per standard distance do human drivers have?

    Group B
    1. What is the financial impact of crashes per distance driven with AP?
    2. What is the financial impact of crashes per distance driven with other self-driving cars?
    3. What is the financial impact of crashes per distance driven with human drivers?

    Group C
    1. How many crashes per distance does AP have?
    2. How many crashes per distance do other self-driving systems have?
    3. How many crashes per distance do human drivers have?

Anything not comparing these numbers one to the other is incomplete and potentially propagandistic. Tell me about the human impact. Tell me about the monetary impact. Tell me about the number of total crashes.

Without all three numbers in each group and all three groups, we can't tell if 273 is a high number or a low one. We can't tell how it's impacted by total distance driven with the systems. We can't tell if the accidents are more or less serious. We can't tell if these systems are dangerous convenience features or if they're actually improving on numbers from human drivers. We need real data, and real comparisons.

That's a very clean way of dividing up the data, I'd be curious about those figures too.

I was going to question the relevance of comparing tech-caused accidents with human-caused accidents since, despite the growing evidence that tech is the better driver under normal circumstances, we have this cognitive bias around lower perceived safety when someone/something else is driving. People just want to feel in control, even though they're more likely to crash, and my guess is that this will slow down adoption of self-driving technology more than necessary. But then I realized that evidence like this would probably be the best argument no matter what, so in the end, it makes even more sense to compare those numbers!

Humans have a fatal crash about once per million miles. Self Driving is complicated by the fact that humans can intervene. Human drivers have no such backup.

So comparison may require compensating for the fact that automated drivers have that human backup preventing them from otherwise killing more people.

> Self Driving is complicated by the fact that humans can intervene.

And that self driving "opts out" of any conditions it finds too challenging, requiring the human to take control. Meaning self driving has more miles in relatively safe conditions.

You and parent make very good points. These would still be relevant if we had all the data. It'd be nice to have the data and this context for it.

Aren’t level 2 systems supposed to be exactly this backup for humans?

Yes, that's important to keep in mind.

If the overall system of auto driving plus human turns out to be safer than a human alone, that could still be a good thing.

I'd also like to see data, come to think of it, for systems with the safety systems that aren't self driving like lane departure warnings, front and rear collision warnings, adaptive cruise control, automatic braking, driver awareness warnings, head position warnings, directional headlights, night-vision HUDs, and such.

lol thanks for the broken down version. I was saying the exact same thing but too lazy to break it out into all the rates. Seems like the dumbed down version is "what is the rate vs other equivalent autopilots" and "what is the rate vs your average fleshbag"

Oh, you're a tesla apologist then? For each of those groups, the AP case is much more forgiving as both the self-driving case and human driving case need to handle all driving conditions whereas AP only needs to handle the safest of conditions (especially safest on a per-mile basis).

Oh, and bonus: "other self-driving systems", Tesla's AP and FSD (despite it's name literally being "full self-driving") are not self-driving systems.

Your comment ignores the parent and therefore misses their point. It is _irrelevant_ which conditions need to be handled by AP. If we don't have even basic data, we can't make any useful comparisons.

No need for the strawman.

Literally just omit the word "other" and the point stands. Without data we don't know numbers for Tesla or other manufacturers. That's literally what data means - the quantitative and possibly qualitative information about the topic.

A "crash" in this case is any crash where the automated driving aids were in use less than 30 seconds prior to the crash. So for example, if the car realized it couldn't deal with something 29 seconds up the road, gives a 10 second warning to the human that the human needs to take over, and then the human drives into a crash, this would count as an Autopilot crash.

I'm no Tesla fan, but that's a bit disingenuous.

It does seem like there should be some happy medium between this and Tesla's policy of not counting it as an Autopilot crash if it disengages one second before impact.

I guess at the end of the day I mostly care about the overall accident rate, AP or not. Since the premise is that this kind of car will be safer, you would expect that rate to be lower.

Tesla uses a 5 second threshold, not 1 second.

How does it not count? 10 seconds doesn't seem like much time for a human not really paying attention to the road, as would probably be the case, to assess the situation and respond. Although the cutoff is at 30 seconds, I wonder what the distribution is actually like.

10s is basically an eternity in car time. Pretty sure I could wake up from a nap and boot my brain enough to drive in 5s or less. A car going 80mph can stop in around 4s.

> Pretty sure I could wake up from a nap and boot my brain enough to drive in 5s or less.

Citation needed, certainly for the case at hand, where you’re suddenly required to drive a car in an emergency situation.

(Edited to add the claim I responded to, because it looks like a few replies didn’t look at that context)

On the contrary, I’d challenge anyone to find a non-contrived example of a crash that is unavoidable with 10 seconds of warning. Driving off a 1600 foot cliff does not count.

Your car has gone the wrong way down an express lane some distance, and deactivated when a semi trailer traveling 75mph in the opposite direction is detected. There is no shoulder, and it's stopping distance plus yours is greater than the remaining distance between you (.1 mile is way insufficient)

”Wrong way down an express lane” is super contrived. Those are the kinds of clear cut issues that self driving is guaranteed to be better at than humans.

If semi can stop in 6 seconds sounding horn would do it.

Every crash is avoidable by definition.

Human error/poor decisions, mechanical/electronic failures, acts of god could all reasonably be labelled as reasons.

Regardless of time.

Not every crash is physically avoidable _given a time constraint_. When you're a millisecond from impact, there's nothing you can do. One second, there exist many cases where there's nothing you can do.

10 seconds before impending collision, and it's hard to imagine the case where there's no time to react in some way that avoids the accident.

You are always required to drive the car. Autopilot does not replace an attentive human driver, there is no excuse for being distracted or spacing out.

Pretty sure you could not, and your confidence is terrifying.

Set a timer and watch it count to 10s. It’s a long time.

Remember, people fall asleep while driving cars pretty regularly. I know people who have fallen asleep driving on the freeway, woken up in the median, and just guided the car back on the road to continue.

Are they stupid for not taking breaks, etc? Yeah. But is it unreasonable to think 10s for taking over control from sleep is doable? No, it’s doable for people with the right mental state. The most common mistake is overcorrection on waking up.

Are you talking from sleep or in the back seat of the car (both illegal)?! 10 seconds is long as hell - i have no clue what you are talking about.

Definitely sleeping in a cat is illegal.

With level 2 autonomous driving systems the human is required to always be watching and ready to take over. 10 seconds is a reasonable amount of time for this transition according to Mercedes Drive Pilot system: https://www.roadandtrack.com/news/a39481699/what-happens-if-...

That 10s is for a L3 system -- L2 systems require the driver to maintain attention at all times. "FSD" is still a level two system.

In level two, it's not about "handing back control" -- the driver is assumed to be in control at all times, with the car assisting. If the car assists too much then people will start treating it as if it were an L3 system, but it's really not.

Why? 10 seconds is not enough. It seems perfectly reasonable to assume Autopilot needs to do better here.

I disagree. How could a system look much further ahead in a dynamic environment?

If "as good as we can do" isn't good enough than it just isn't. If it can't offer enough time for a driver to reasonably take over then it can't meaningfully function in its intended role.

I'm not informed enough about this to say whether it falls under that threshold or even what that threshold is. But the limits being clear technical ones does not change the degree to which they are limits.

What I mean is, can you imagine any system that can see 20s (say) ahead in terms of other people's driving? It's an eternity.

This is exactly why level 2 systems will never be safe. The constraints set the human up to fail.

"Level 2" is just lane centering + adaptive cruise control. There's nothing about this combination that "sets the driver up to fail". Some drivers like to pretend that the system is way smarter than it really is, but that's a case of "play stupid games, win stupid prizes". You can't fix that with changes on the system side.

It's not disingenuous,

The whole point is "my new system reduces crashes", if it doesn't, in the bottom line, then it doesn't.

I think OP is saying something slightly different: the Tesla numbers could be substantially worse if Autopilot simply disengages 1 millisecond before impact, it would not count as an Autopilot crash.

And there's recent articles where almost exactly this has been found. Autopilot deactivating 1 second before impact.

In these kinds of cases, then yes, Autopilot is definitely partly at fault, but also the human is because Autopilot is only a level 2 system, the human is still responsible.

> Autopilot is only a level 2 system, the human is still responsible.

True.

> In these kinds of cases, then yes, Autopilot is definitely partly at fault,

Partly? Now that's called shilling.

1 ms < 30 seconds, so that data is there as well.

Also,

>I'm no Tesla fan, but ...

Sounds like he's vouching for them.

Anyway, the argument is not against @bradfa, it is about whether AP makes driving safer or not.

The WaPo article addresses that and counts them as AutoPilot engaged crashes. It seems to disengage 1 second before many of the crashes. It is nearly impossible to draw any conclusions about anything from this article. Other than, people use AutoPilot and sometimes they crash using it.

The one interesting bit is where the damage happened on the cars and the hitting of stationary objects. Those seem like unforced autopilot errors.

But with a raw number in the headline -- instead of some kind of ratio that could compare with/without the new system -- it's not that honest.

As has been stated, though, simply saying there have been 273 crashes by itself means nothing.

If I said to you "following the introduction of a new drug there have been 1000 heart attacks in the USA this year so far", you may think, shit, 1000 people have had heart attacks, but when you realise that there are 800,000 heart attacks per year in the USA, making it about 400,000 pro-rata for June, reducing that down to 1000 would be amazing.

So, 273 crashes. But what would that figure be without the system?

We need the data to compare this to groups of drivers who don't use these systems. Otherwise, it's meaningless. The numbers are meaningless. Like values without units.

Huh?

But we DO have the data for "groups of drivers who don't use these systems" ...

> A "crash" in this case is any crash where the automated driving aids were in use less than 30 seconds prior to the crash.

It's 5 seconds, not 30 seconds. From Tesla safety report [1]:

> To ensure our statistics are conservative, we count any crash in which Autopilot was deactivated within 5 seconds before impact, and we count all crashes in which the incident alert indicated an airbag or other active restraint deployed.

[1] https://www.tesla.com/VehicleSafetyReport

The article states: "The NHTSA order required manufacturers to disclose crashes where the software was in use within 30 seconds of the crash, in part to mitigate the concern that manufacturers would hide crashes by claiming the software wasn’t in use at the time of the impact."

Ah, thanks for clarifying. I thought you referred to Tesla's counting methodology.

I'm just going to interpret "[autopilot] couldn't deal with something" as "autopilot would also have crashed the car in that situation".

To me a situation in which autopilot would have crashed, gave control to the driver, who couldn't "save" the bad situation, is just an autopilot crash.

Autopilot is a SAE level 2 driver assist system; it should never be trusted to keep you safe. Whether or not the system bails out completely prior to a crash is immaterial.

That's a definition that no other car company would accept either, as their systems would score very poorly.

1. While that’s true there’s no reason to think that extreme case is normal for these.

2. Even if it was, it’s sort of irrelevant. When considering the safety of the feature, the whole process, including allowing the driver to become distracted, must be considered. This isn’t a clean room where you can just focus on the performance of the models, this is a real life car interacting with other unpredictable cars, being driven by an imperfect human. If the nature of these systems is that it makes the drivers complacent, that’s absolutely something that needs to be taken into account when thinking about the overall experience of a self driving system.

What's your alternative? Tesla has also faced criticism for turning off autopilot less than a second before the crash, and then claiming that because the car was not on autopilot it was driver error.

"On Thursday, NHTSA said it had discovered in 16 separate instances when this occurred that Autopilot “aborted vehicle control less than one second prior to the first impact,” suggesting the driver was not prepared to assume full control over the vehicle.

CEO Elon Musk has often claimed that accidents cannot be the fault of the company, as data it extracted invariably showed Autopilot was not active in the moment of the collision. "

* https://fortune.com/2022/06/10/elon-musk-tesla-nhtsa-investi...

My alternative is to present the data including how long before the crash the system deactivated and not just report all of the incidents into one category. This requires nuance and judgement in reporting.

I'm not blaming Tesla or the NHTSA for this, I'm blaming the WP as their reporting on this lacks any information and seems rather to just paint the data in such a way to make Tesla look bad. There's no room for interpretation by the reader because no actual useful data is presented.

Edited to correct my naming of the article source.

Even Tesla uses 5 seconds and not 1 second. The Fortune article was misleading.

FYI: there are over ~17,000 crashes per day in the US.

https://amaresq.com/blog/auto-accidents/how-many-car-acciden...

Yep, by itself the data in the article is kind of useless unless we can get that normalised to crashes per km driven, and to also include stats on other manufacturers' cars' use of autonomous systems at the time of crashes.

The author alluded to it by saying Tesla collects more data, however it doesn't state that no other manufacturer collects what we want to know about (E.g. dynamic cruise control, active lane keep assist wheel nudges, AEB activation, etc).

So, actual data please, then let's compare apples to apples rather than this sensationalist clickbait. (I would really like to see crashes per km for both AP on, AP off, and FSD beta enabled in the case of Tesla.)

I agree. I also want to note that Tesla has treated this subject like an opportunity to market and really done nothing to deserve the benefit of the doubt.

> normalised to crashes per km driven

Preferably normalised by type of road too. I.e. if one system only drives on highways it might not compare to one that drives on more dangerous roads.

Good point.

It's so ridiculous the denominator doesnt matter to anyone ...

Was this headline written by an unpaid teenage intern? It should be obvious that '273 crashes' is meaningless without a comparison to how many are expected among other similar vehicles.

Why? A human driver and a computer “driver” that is misleadingly marketed are two, totally different things.

No this is just the new way of doing things. COVID case numbers were notoriously absolute during the whole pandemic (in some cases govts insisted the media/dashboads only gave absolutes), which is so useless as a data point and so easy to make useful, but no...

This is the new journalism/government: we love absolutes.

Your first sentence is more of a commentary than anything. Other than that I have to agree with you, just giving the numbers like that is super meaningless and almost deceptive, especially for infection data. It's a disgrace and shouldn't be tolerated, it makes people with poor deductive skills draw the wrong conclusions and operate with emotions instead of logic.

If you bothered to read the article, this information is presented.

There's a whole article motte to justify the headline bailey!

It could have been done in a couple paragraphs and charts/tables with informative "rates of..." but they have to mix it in with a ton of fluff. Modern journalists always have to interject their "take" rather than just make a statement and show some data to support it.

From the article... Tesla vehicles made up nearly 70 percent of the 392 crashes involving advanced driver-assistance systems reported since last July

Do Teslas make up 70% of vehicles with advanced driver assistance? Here's a list of cars with these systems.[1] That's a lot of models, so it certainly appears Tesla has an outsized proportion of crashes.

1 - https://www.consumerreports.org/car-safety/cars-with-advance...

Poor comparison. The comparison is what percentage of vehicles that operate with functional assist are Tesla vehicles.

This why we would use rates per company vs absolute numbers. Takes most of the speculation out.

Not all of the vehicles listed qualify as "advanced driver assistance" by the NHTSA definition.

A comparison of miles driven with the systems active, normalized for road conditions, is not currently possible, but would be the most useful metric.

Well, considering these happened when (I assume) the cars are using 'state of the art' driver assistance software, any number >1 is something to discuss.

Tesla could have been a fantastic brand if they'd managed expectations better on AP/FSD and the douche-in-chief kept his twitter trigger hands in check a bit.

Their brand's jumped the shark now.

The numbers are a bit meaningless without context. How many Teslas total on the road? How many accidents total?

For me the interesting question is: does Autopilot crash more often than human drivers?

Short answer, yes. Based on the data so far released by the NHTSA, comparing similar cohorts (drivers of luxury cars with advanced driving systems), Tesla's using Autopilot crash at least 3x more frequently than cars from other automakers using their respective advanced driving systems. (3x compared to Honda, 50+x compared to Toyota, GM, or Ford L2 systems)

Or to put it more directly: Nissan has sold over 560k cars in the U.S. with an L2 system, and has no reportable accidents. Tesla has sold approximately 1.3 million cars in the U.S. with Autopilot, and has 273 accidents. GM has been selling SuperCruise since last year (approx 100k vehicles), and has 0 L2-level accidents. (And testing by Consumer Reports, et al, report fewer issues with SuperCruise than with Autopilot in similar situations.) TLDR: there is no way to massage the data to make Tesla look even remotely safe compared to its competitors.

Can't read the article because I'm poor, but... Is that a lot? How high is this compared to non-autopilot cars? How many of these are fault of the autopilot? Where are the stats coming from? What counts as a crash - is every one of those fatal, or does this include running into a pole while parking?

Am poor as well, can't read the WP articles coz of my location.

If https://wapo.st/3mS7whs doesn't work, try https://archive.ph/ycgCV

https://wapo.st/3mS7whs <- "gift article" link, no account required

Non-clickbait title: Teslas running Autopilot have been in 0.0000273 crashes/km in less than a year!

I've been alive long enough to see a considerable sample set of human drivers and I've got to say the bar Autopilot needs to exceed in order to save limb and life is incredibly low. Couple that with the fact that just here in SoCal Tesla has become a common sight, 273 crashes sounds like a hell of a lot of human suffering was avoided.

I don’t get the point of fsd. I get the point of cc-take a load off your foot; and lh-wake you up if you drift. But all that fsd does is stress me out having to hyper-vig over the wheel and break in case the fsd fsd’s up. (And those stupid flight-yolk style steering wheels suck!)

Compared to how many hundreds of thousands or millions of accidents that have happened to human drivers this year. Even adjusting for population size of Tesla vs normal cars, that still seems better than a human.

What's the crash rate per mile for Tesla vehicles in production? Compare that to the crash rate per mile for the rest of the market and you probably have a better measurement. A more informed headline.

That number really says very little if it's not expressed as crashes per distance driven. The article is paywalled, so it's hard to say if the clickbait title is substantiated in the article itself.

As always, if the headline sounds juicy, put your skeptical spectacles on.

As a contrast, is it possible to produce data on how many crashes, if any, the software has prevented?

A Tesla was in a junkyard for three weeks. Then it burst into flames (washingtonpost.com)

Washingtonpost

Absolute numbers mean nothing in a country of 330,000,000 in such low values. What is the crash rate vs your average driver per mile driven having crashes? We know that autopilot isn't perfect.

The article, like the NHSTA, are comparing them to other “autonomous” features of other car companies.

But the real test should be, how much safer is this than drivers not using the feature

You can still compare against other car companies their autonomous features.

If car company A has 20 accidents for every 100 cars utilizing autonomous features and company B only has 5 accidents/100 cars, you can say that company B has better autonomous features and would be a wake-up call for company A.

Autopilot is just adaptive cruise control which most cars have now a days (depending on trim level). Do they mean full self driving?

No they mean just Autopilot which does lane following and lane switching on highways.

Enhanced AutoPilot switches lanes. Regular AutoPilot does not. I think most people have regular.

Is there a comparison with other top car models?

"Tesla vehicles made up nearly 70 percent of the 392 crashes involving advanced driver-assistance systems reported since last July"

"Honda reported 90 crashes during the same time period involving advanced driver-assistance systems, and Subaru reported 10."

Wow, that sounds a lot

Unless Tesla's Autopilot is far more reliable than the rest of the industry, so Tesla's drivers use it more. That cannot be inferred from the current data but Tesla does communicate this figure for their vehicles.

We need accident per mile driven on active ADAS to make a fair evaluation.

Tesla's Autopilot makes it easier for drivers to treat it as if it were L3 (even down to marketing it as "full self driving") which means they're not paying as much attention to actually driving the car as they're required to for a L2 system.

Do they make up 70 % of driver-assistance cars? Or how much? It's like saying 80% of bad search results are google results, if google had an 80% market share - it says nothing without the comparison

You're right it would be good to have the actual marketshare, but I'd have a hard time believing it would be anywhere near 70% marketshare of ADAS enabled vehicles. Even my 2017 Hyundai has ADAS functionality.

It would be great to see the 273 crashes in detail categorized by what kind of edge case crash. I hope there were very few fatalities. I suspect majority humans transitioning to control is responsible which means that needs to be looked at - which shouldn't be the problem.

Also it doesn't seem that high of a number relative to how many crashes have occurred. Obviously the closer to zero the better since they have a high bar to instill confidence in self driving.

https://archive.ph/ycgCV

How many were deemed the fault of autopilot. Many may be cars running into the back of them etc.

Don’t think AI drivers on the same roads as human drivers is viable in the long term

I agree. And I think insurance companies will lean this way also. Self-driving (meaning human control) will come at an insurance cost premium. Of course the whole concept of liability gets interesting. I'm expect that within a couple years that AI-driving will be markedly safer.

We need a per capita comparison of humans in crashes and autopilot in crashes.

It should not surprise me that the Washington Post, the paper that ran 2-3 hit pieces per day against Musk after he announced his intent to acquire twitter is now running a hit piece against Tesla with a misleading title.

How is the title misleading?

It's misleading because it doesn't fit their Agenda.

I would say that the wording used subtly implies that it is a negative thing. Using "less then a year" signals, to me, that the number is high for a short amount of time. (being behind a paywall I cannot read the full article to judge if it is actually misleading based on that)

It's misleading because it's a number of crashes that sounds bad, but there's no way of knowing if that's true from the headline if you haven't got a bunch of car statistics memorised

A less misleading title would be "Tesla autopilot involved in 70% of driving-assistance accidents this year"

Even that is potentially misleading if Tesla autopilot represents more than 70% of driver assisted miles on the road

Not matter what way the headline would have been phrased people would have complained. "Company X lost $money ..." headlines are printed all the time.

To me it doesn't look like unfair reporting but instead i see a group of people accusing with loaded language ("hit pieces" is loaded) normal reporting to be biased against Musk because "daddy can't be wrong" and criticism must there for be an attack.

I am not sure what is misleading in the title.

What is missing however is the percentage of crash per amount of cars registered with and without assistance equipment and per brand. I have no way from this article to know if it fare better or worse than other brands and/or traditionnally operated cars.

Reply on news.ycombinator.com