How do cars fare in crash tests they're not specifically optimized for? (2020)
danluu.com
danluu.com
I'm smitten with the stupid idea of teaching your self-driving car that when a crash can't be avoided it should attempt to make the crash as close as possible to the nearest crash test protocol of the jurisdiction it's operating in.
Like quickly put a dummy in your seat and throw you out?
If you still don't hit anything, you can often even walk away, or end up with minor arm/foot injury. Unfortunately I've seen videos of contrary, ie couple riding same bike, hitting car, flying over it, sliding maybe 50m and still hitting curb so hard their bodies jumped quite a bit in the air. Both dead on spot.
As I said, it could entirely have been chance that this worked out for the best. I'm not a rider myself, and my brother is not an expert just a rando like me who has had a fair amount of interesting experiences in his life (how he is still alive today, I'm not quite sure, but luck has to be the biggest factor).
Another completely insane motorcyclist meme is if you are inevitably going to strike an animal in the road, it's better to hit it in a wheelie, that is, with the front wheel raised off the ground. Again: insane.
Survivor bias. The people that it didn't work for don't get to tell you how well it worked.
I'm reminded of a talk at a science fiction convention years ago mostly targeted at writers (and done by a physician) - "Maim 'em Right." Basic message was that when writing injuries, make them realistic. If someone takes a blow to the head that knocks them unconscious for a significant amount of time, they're not resting for a day then getting up and going on to lots of exertion and winning the day. They're suffering from concussion, maybe subdural hematoma/brain bleed, etc.
Also discussion of TBI, including things like shaken baby syndrome - in the skull, the brain has a consistency somewhat more durable than thick jello. Get one of those brain jello molds available around Halloween, make a batch, then instead of unmolding it keep it sealed and give it a good shaking. THEN unmold it and see how much damage there's been.
Whoever is running the teleporter would have to have "collision avoidance" strategies I guess. :-)
Which decelerates your face more gently, 14" of air space for it to flop around in or 5" of air followed by 9" of airbag? Engineering tradeoff are everywhere.
You're absolutely right that slowing down slowly is the name of the game. Failing that, evenly spreading the force is the next best thing.
There are only two ways to go down on a motorcycle. High-side or low-side. Pretty sure most people would prefer low-side if they ever had to pick. I had a high-side crash on a bicycle once. Luckily it was a grass field with soft dirt. Had to go home and change my pants after that one.
I'm very lucky, in both cases, I didn't hit anything but the road. Especially not the car.
I still rode MX bike on tracks for a while then eventually quite entirely.
Now I've got a mountain bike and it's scary enough already!
And did you stop riding?
Getting run over by a car after going down is a bad day.
The real technique, in my opinion (I have been down a few times), is to use the brakes to take off as much speed as possible while, if you can, bringing both the bike and yourself off the roadway and onto a shoulder or median.
And always: legal or not, lane split into stopped traffic so that it's an instinct. More than once the car in front of me has slammed his brakes on and I've split in next to him, only for him to be rear-ended by the car which was behind me.
Much better to be next to a car that is rear-ended than to be behind one.
I didn’t know it was a feature until I nearly got rear ended waiting on a left turn and the car went into rear impact protection mode. (Other driver veered off at the last second). It was quite impressive tbh.
Edit: For the curious:
https://www.volvocars.com/lb/support/manuals/xc90/2018w46/dr...
https://www.volvocars.com/en-th/support/manuals/v60/2017w17/...
If the self-driving tries, and it worsens the impact somehow, bigger lawsuit. If the self-driving fails to do so, bigger lawsuit.
And it might sound "insane" that a best attempt could result in worse liability than no attempt, but unfortunately there is a lot of legal precedent to back up the idea.
PS - Although legislation could make self-driving manufacturers immune from increased liability for these "best effort" systems.
You jest but this is what humans and animals do instinctively. When you fall, you try to fall in a way that minimizes damage.
For example, if you trip over something and your hands are occupied so you can't use them to catch yourself, you will instinctively try to roll onto your shoulder instead of faceplanting into the ground.
You see this in freestyle motorcycle and mountain bike riding, and skateboarding etc as well. If you see you aren't going to land the trick, you try to throw the bike or skateboard away from you and land feet first down the ramp. Plenty of videos of pros doing this on youtube.
An AI driver should have those instincts too imo. I’d expect my car to crash in the safest possible way if it can’t not crash.
https://www.mercedesbenzofnatick.com/new-features-mercedes-b...
What if the nearest crash test protocol is an absolute "worst case" scenario like the "small overlap test" [0]. Here, getting any additional coverage would improve the outcome, even though it makes it less similar to the most-relevant IIHS test.
For example he's grouping BMW and Mercedes right below Volvo.. at least in the case of the Frontal Small Overlap test that doesn't seem right at all. The Small Overlap Frontal test was introduced in 2012.
The Honda Accord in 2013 was acing that test, BMW and Mercedes were failing it miserably for a few more years. The 2013 Accord IIRC should have been rated at least as high as the Volvos.
I have a 2013 Subaru, it got 5 stars in the frontal small overlap test as well, years ahead of the BMWs and Mercedes.
Something is just off about the whole thing if you remember back to when these tests were introduced and you were looking at buying a car.
Tesla pretty much never got anything but 5 stars in these 2 tests IIRC, not sure where that came from.
The article would have been much better if it linked to the specific crash data to support the claims. IIHS has all that data publicly available IIRC. It is good stuff to look at when buying a car.
BMW, Mercedes, and Toyota/Lexus were the real outliers who had appeared to try to "game" the old benchmarks and then failed these new tests the most spectacularly.
You’re not recalling correctly. The Model S got an “acceptable” rating on the small frontal offset test in 2017: https://www.slashgear.com/tesla-model-s-again-fails-to-earn-...
Also he's attributing poor passenger side tests to malice when the real reason might be that the engineering requirements are very different because the steering column doesn't exist on the passenger side.
And he's somehow rewarding M-B and BMW when they scored "Unacceptable" or "Poor" on the same tests for years when scoring "Excellent" on the other tests. Scoring Unacceptable/Marginal/Poor on a test when scoring Excellent on others is much more evidential that the car was specifically designed only to game the test standard.
It took Tesla longer to upgrade the car from "Acceptable" to "Excellent" on the new tests than it took BMW/Mercedes to upgrade their cars from "Poor" to "Excellent" on the tests, but that still means M-B and BMW put hundreds of thousands of cars with Poor crash performance on the road for years while Tesla was putting cars on the road with "Acceptable" performance in the same test. Somehow that makes Tesla worse in his rating system.
This is an especially concerning observation for Honda. I'm curious if there's any more detail on this or if there's a plausible reason aside from gaming the benchmarks (e.g., the positioning of firewall-forward components might naturally favor small-overlap protection on the driver's side, or a steering-wheel mounted airbag would naturally perform better than dash-mounted).
In both single and multi car collisions, we will assume there’s a driver in all cars, but less than 100% chance for there to be a front-seat passenger in all cars.
It also meant that changing the right hand side sparkplugs was a breeze, but the left side required either dropping the engine or drilling the firewall.
I'm not from the UK or any other RHD market, but if you are then are some cars, such as the Camaro/Firebird, sold LHD there? Or just not sold at all?
I'd love to know if there really are factory RHD Camaros, and if not then why did they have that unusually lopsided engine bay.
I thought about the following cases where you might encounter this type of impact:
1. Driver is not controlling the car properly (asleep, drunk, etc.) and leaves the carriageway and impacts some road furniture, etc. - in this case, single driver likely has around 50/50 chance of going left vs right, but an accompanied driver is less likely to have this type of incident (passenger can shout to wake the driver, reach for the wheel, etc.). As a result, total fatalities from this collision mode would probably be significantly higher for drivers than passengers.
2. Driver understeers around a corner. In this case, understeer would be more likely on a smaller radius corner than a larger one. As the drivers' side is closest to the other carriageway, this means that the driver would already need to be on the widest part of the curve in order for the impact to be on the passenger's side. This likely reduces the frequency of oversteering from the outside of a turn - although as a secondary consideration, it probably implies higher speed, and so potentially higher fatality rate. Therefore, as more oversteer-related incidents would happen from the inside of the curve, the driver is more likely to be impacted.
3. Impact occurs during overtaking. In this case, we assume the driver is in the opposite carriageway. Oncoming traffic is likely to evade, and during an overtaking manoeuvre, it is likely that evasion into the driver's original lane is not possible due to other traffic. As a result, both our driver and oncoming traffic are likely to try to evade off the carriageway (onto the hard shoulder, grass verge, etc.) Considering that this is a small overlap test, it would imply that evasion has been at least partially successful, implying reduced speeds, etc. It's probably not particularly more likely that either driver or passenger is subject to this impact.
I couldn't think of any other illustrative cases, but it does seem that the even considering that passengers are not present in all cases, the individual risk to a passenger is lower than the risk to a driver when there is one present.
Additionally, this is a really, really difficult test to pass, which is why everyone failed it. Even the revisions didn't help too much because there's not much that can be done to an existing chassis other than add a deflection bar or something. True success here requires ground up engineering.
The most popular Honda and Toyota models are just as American (or more so) than anything the Big 3 produces. Some of them are even designed in Detroit.
The small overlap tests started in 2013, most of the Hondas already scored "Good" in the small overlap tests by 2013.
The Author awards BMW and M-B excellent status for correcting their problems by 2017, but most of their cars were Poor/Marginal for 4 years.
Somehow Honda is worse for having most of their models "Good" in 2013 with only a few models having issues.
The Civic and the Accord for example have never been tested as anything but "Good" in these tests, whereas comparing them to the BMW 3-series and M-B C Class those cars tested out as "Poor" or "Marginal" for 4-5 years before they were corrected. The Pilot by comparison did have problems.
How this equates to BMW/M-B being great and Honda being problematic is terrible.
Note the author in his notes doesn't denote which cars/trucks/SUVs in which model years he's actually talking about, making it impossible to figure out how he's judging the brands. Most of doing well on these tests tended to have to do with where in the redesign cycle a manufacturer was with a particular car/truck/SUV.
https://kjonnsforskning.no/en/2019/06/cars-are-still-designe...
https://www.forbes.com/sites/stevetengler/2021/02/16/volvo-c...
Regulatory Frontal: 50th male in front-left and front-right, or 5th female in front-left and front-right
Regulatory Side, struck by SUV: 50th male
Regulatory Side, struck by pole: 50th male
NCAP Frontal: 50th male in front-left, 5th female in front-right
NCAP Side, struck by SUV: 50th male in front-left, 5th female in rear-left
NCAP Side, struck by pole: 50th male
As you can see, most tests are done with male dummies, and almost no females dummies are drivers. Fortunately, there's money that's just been allocated by Congress for doing more female dummy testing.
[1] Also important to note that, at least in the United States, the percentiles were determined in the 80s and Americans have gotten fatter.
Here's a list of current US dummies, including average males, extra small females, kids, etc:
But, they have no power to regulate. Vehicles that do poorly under IIHS testing are completely legal, they simply get named-and-shamed. Sometimes I wonder how lax NHTSA testing would be if it weren't for the comparison point which is the IIHS.
Like how them changing their whiplash tests caused neck pain and injuries for anyone significantly outside their test dummies. But because they're not caused by an impact, but rather because they fail cars that allow the user to adjust head restraints to be less effective for the median posture, they don't care.
Thankfully NHTSA has a much broader mandate.
What they claim to do there is to use their real world data to try to optimize for real crashs and not for crash tests.
If the average car would already (pre-modifications) do well in driver-side small overlap situations, then this type of crash would simply not get chosen for extra testing.
a. Electric Volvo?
b. Won't even need a car by then?
Electric cars are really amazing in some scenarios though - I was a passenger in a Tesla in a winter accident that would have been a rollover event in many cars… we ended up in a flat spin and in a ditch instead.
Volvo has always made safety a key value. I remember picking up my dad from a police station after he had been in a 50 car pileup in the 90s. His Volvo was unrecognizable, but he walked away, only needed help from a policeman to cut the seatbelt off as the buckle release wasn’t accessible.
Women are neglected in testing of every standard. Essentially they’re benchmarks of computer performance in human form, with the most hopeful data being used.
It’s much harder to account for women who have more variation such as hips, breasts, pregnancy status, corporate would rather rather tell you the best results with the cheapest and the best case results. The weight of the dummies, the height, and the lack of variation is both cost cutting, testing the most healthy models, and way to bypass the regulations with clever tricks that most people trust because there aren’t more data sources and it’s a lot of effort to learn more.
Most things that are tested are, especially pharmaceutical drugs. They use men’s dosages and at most change bu weight but ignore important factors that are women’s issues since it’s easier to neglect it lazily, then blame the patient’s biology rather than the doctor’s incompetence and lack of research, but the lack of information that drug companies give.
Paid drug trials are often signed up for by more risk taking males, and hormone cycles of women make drugs have more variation they’d rather sweep under the rug, you also see the difference in showdogs winners where a female dog’s hormone cycle like her period can destroy performance.
Another cost cutting example is the lack of long term effects, they use mice models rather than dogs because it’s cheaper and shows less long term effects, win/win for hiding these results until it’s discovered by real world usage.
Something encouraging though are computer physics simulations. Volvo has been at the forefront of safety, they’re the top rated, tested on women before regulations forced others to, making the safety belt and not patenting it, encouraging its usage due to their focus and morality. The results don’t surprise me.
I think it's fair to say that cars are incredibly safe compared to 40 years ago. But crash safety compromises new cars in so many ways, it's a bit of a pity.
The angle and positioning of the pillar in relation to the driver are huge factors.
Anyone telling you that energy absorbing structures are of significant value except at a very narrow range of speeds and impact type is peddling falsehood.
The bulk of the improvement since the 90s (well into the era of crumple zones, remember they are a prerequisite for airbags) has been from better airbags, side curtain airbags, pre-tensioners and better modeling software that allows those things as well as seats and seatbelts to be engineered to levels that were unheard of even a decade ago. Whether you hit hard objects in a crash has a lot less to do with luck these days. Stronger cabins are a plus (especially for non-frontal crashes) too.
How well do cars do in crash tests they're not optimized for? - https://news.ycombinator.com/item?id=23689538 - June 2020 (343 comments)
I think a better way to look at how cars fair in crash tests that they are not "specifically optimized" for is to look at results from real world crash data. Like, among various vehicles which received top marks in <crash test>, how did buyers of those vehicles fair in real world accidents that the crash tests are designed to simulate.
https://www.euroncap.com/en/vehicle-safety/the-ratings-expla...