Why Teslas Keep Striking Parked Firetrucks and Police Cars
slate.com
slate.com
People shouldn't really have to. The phrase "self driving car" is a binary yes/no in common, regular English. I don't know if this is Tesla's fault or some committee's fault (SAE?), but it's really terrible marketing to call things "self driving cars" when they can't drive by themselves.
Cars shouldn't be allowed to be marketed self-driving, or autonomous, etc., until a driver isn't required.
"Self-driving car" is just woefully inadequate to describe the capabilities it encompasses. A century old Model T with a brick on the accelerator could be accurately described as a "self-driving car". It's not very good at driving, but neither are drunk people and we still say that they're driving.
Likewise, I would argue that a car that can maintain speed and steer itself to stay in it's lane is conceptually closer to a "self-driving car" than it is a "fully manual car".
> Cars shouldn't be allowed to be marketed self-driving, or autonomous, etc., until a driver isn't required.
That's going to be a licensing thing. Frankly it probably is time for the government to step in and create some kind of licensing for cars that indicates they're fully autonomous. Even if they aren't going to issue any at the moment, it would be good to set the direction that only cars with whatever government seal are actually able to operate without a driver. Then it'd be a little harder for companies to play word games.
It's Tesla's, they're the ones calling their system "autopilot" (conjuring the idea of the car driving itself in the 99.99% of the population which has never flown a plane with an autopilot and has no idea how limited they can be) and literally selling a "Full Self-Driving Capability".
Other companies have generally been careful to bill their tech as assistive (Volvo's automation system is even carefully called "Pilot Assist").
And SAE is levels of automation, describing how little / how much automation there is in the driving system, it doesn't intrinsically imply self-driving capability, only that some aspects of driving can conditionally be handled automatically. SAE levels 1 and 2 are specifically worded in terms of assistance as well, here's SAE2:
> The driving mode-specific execution by one or more driver assistance systems of both steering and acceleration/deceleration using information about the driving environment and with the expectation that the human driver performs all remaining aspects of the dynamic driving task
One of these douchenozzles has a sticker that says 'I'm probably not driving'. This particular tesla has whatever 'drive like an asshole' package where the car is accelerating unsafely, weaving between cars/lanes... as this lady is reading a book.
I followed it for a while just out of curiosity. Once it got to a shopping center with dog-legs to turn left, the tesla lost its shit. It began for no reason to ride the curb, and I mean ride it. Brand new tesla just ate up both its driver's side rims to the point sparks could be seen. The lady reached over and did something that did not work, and it just kept going until she slid over and regained control.
If tesla can't even get the simplest of shit right how tf can they claim self-driving? And does it matter that if this person kills someone they are at fault? No one would have died if Tesla hadn't lied about their capabilities.
How many people must die before they are stopped from doing this shit? Tesla says they aren't misleading anyone, but then why do so many of their customers drive hands-free?
Generally, level 4 and 5 would be considered "self-driving" (read more here: https://www.nhtsa.gov/technology-innovation/automated-vehicl...).
The leaders in vehicle automation have self-driving cars, but Tesla is far behind.
Simply detect the presence of "flashing emergency lights" in the oncoming lane and disable autopilot when present. No object detector needed. The signal is so strong it's literally flashing extremely brightly in a regular, predictable pattern - any vision grad student should be able to figure this out.
Could there be false positives? Yep, but very few things will flash quite like that, and most half baked vision engineers can do this. The worst case is literally simply the driver taking over on occasion at night (maybe near strip clubs? lol)
Disclaimer: I'm an idiot so take my view with a boeing 747 load of salt
So can be controlled collisions as such.
Simplest FMCW radars are more than enough for every emergency braking system on the market.
It's not uncommon to even use them in parallel with some more "brainy" radar imagers like on Mercedes cars. That is to make a collision preventable in case the main imaging computer hangs, or crashes.
How about emergency vehicles broadcast on 5Ghz their intended route (for the next 300 meters) or just that they are blue light in the area.
This would only be active in a blue light situation. Manufacturers can then add a detection and warn the driver.
A while back I heard about car to car communications, but I heard next to nothing these days, this is a great use case if you ask me. If someone has more details of recent developments, I'd love to read more.
How would the vehicle know what the driver intends to do in the next 15-30 seconds (depending on speed)?
There are literally tense of thousands of these emergency departments and even more when you consider other vehicles which may be stopped roadside.
Perhaps something great to discuss for the future but not at all practical for the next few years or even decade.
[0] https://journals.sagepub.com/doi/full/10.1177/00187208187861...
But when a piece of software deployed by an OEM has a problem, then it is a systematic failure where all cars of that OEM are having the very same problem and then the OEM has to do something.
It's probably hard to intuit, and no one explicitly thinks about this (except AI engineers and researchers), but emergency vehicles are parked in a way that is "obvious" to human drivers.
Every situation is different -- how much shoulder there is, etc. What the visibility is. The driver unconsciously makes a judgement call to be visible (to humans). There is no way they can judge how an AI is going to view them (which changes from time to time, and across AIs).
So the only solution is for the AI to err in the same way that humans do, or not at all.
It doesn't matter the question of "Oh, what was the accident rate when people where unexpectedly speeding?"
What matters was that single type of car by one OEM. If that OEM has some type of problem with its cars, then its a systematic failure which needs to be investigated and resolved by that OEM.
Similarly, if the AI driver crashes in that situation in a similar (or even less) rate to the base rate (manual driving), you would likely not reasonably bring a suit.
Either way you should arrive to improve the AI, but the base rate does matter, and telling someone their question does not matter because you personally don't care about it is bullshit.
It is the same here. Comparing individual human choices and errors leading to striking emergency vehicles with an equipment failure (Tesla’s detection and control software) leading to striking emergency vehicles is not a comparison and it is no more an excuse to not hold Tesla accountable than it would be to not hold Toyota accountable because their accelerators stick less often than a teenager decided to mash the gas on his new car and wrecks it.
If AI driving is not worse than the base rate in that specific circumstance, that would be very relevant information.
FWIW, I think we're very far from ethically introducing AI driving into the mainstream.
A systematic failure would be striking every emergency vehicle for instance.
So it needs to be investigated what are the circumstances and the OEM needs to resolve the issue.
This is really the glory of automated driving - once an issue is identified, it can be addressed in all the vehicles of that type simultaneously.
In any case the article clearly suggests that it's a systemic issue due to using too few sensor modalities in conjunction with too little training data. Hence it's not restricted to firetrucks, but many new, unlearned situations.
Tesla is so many orders of magnitude worse in this area than human drivers that it will be laughable for them to try to resort to Musk's usual lie -- that Teslas are better drivers than humans.
My remark on that is : non-stationary emergency vehicles also cause accidents, so the number is not comparable with Teslas/SmashedEmergencyVehicle.
I believe your anecdote, but it's uncommom. Colliding with parked emergency vehicles is extremely rare, almost non-existent.
Source: working in auto insurance
There's just so many unanswered questions with automated driving that a lot of people don't really want to solve, they just want the technology out there and we'll worry about the edge cases later...as in, we'll worry about it when people start dying. "Move fast and break things", that retarded shit. I hate it. This SV frat programmer mentality is now bleeding over into automobiles, an industry that does indeed need to be shaken up, but not when it's at the expense of lives.
Treating driving software like it's a fucking web application that can be fixed later is malicious negligence.
So your comparison doesn't make sense.
(Have some history in elevation control, because we made some standards in the industry.)
That has no bearing on how common these types of collisions are. The original issue is whether Teslas are more likely to crash into a parked emergency vehicle than a human.
The training has a better explanation: while "number of collisions per mile driven" is vanishingly small, "percentage of law-enforcement deaths as a result of collisions" is very high. Out of the ~1,200 LEOs killed between 2000 and 2008, about half were due to collisions[1].
That alone tells you how such collisions can be rare, but still a top-of-mind issue for law enforcement.
1. https://journals.sagepub.com/doi/full/10.1177/00187208187861...
(YouTube search results are not useful data about national trends, and YouTube would be intensely boring if they were.)
1. https://www.youtube.com/results?search_query=shark+bite
2. https://www.youtube.com/results?search_query=dog+cheetah+fri...
3. https://www.youtube.com/results?search_query=slot+machine+ja...
I wonder if you could look at accidents that involved a stationary police vehicle. I’d be curious if it’s still very uncommon.
Anyway, I heard that so many times from cops directly. Squad cars in the night are like magnets for drunk drivers.
Accidents involving any emergency vehicle traveling at any speed (including very fast through traffic) are less than 25,000/year in the US.
As I mentioned in another comment, humans drive more than 3 trillion miles per year in the US, so the rate at which Teslas are hitting stationary police vehicles even if they have only ever done it once would be much higher than the rate at which humans do.
Maybe, but the very first thing i was taught for how to work a MVA was that under no circumstances were you allowed to walk between parked emergency vehicles, because fenders are just a very blunt pair of shears at knee height that will happily amputate you when some dingus rear-ends the parked squad car.
Source: EMT
Note that I said generally recognize and not always. Humans are bad drivers and never should be in control of a car. (which is why I'm so in favor of mass transit)
When Mr. Rajkumar details how a camera works (he did dumb it down for the layperson quite a bit), it seems almost laughable that a technology like that can even work in a real-world scenario like vehicular traffic.
I've always been of the mind that this technology can only work if you reduce the number of variables at play... that means removing the human element altogether (and even then, mother nature may have other plans), meaning self-driving cars should have their own roads. Pipe dream, I know.
Why a pipe dream? It's one of the most sound solutions to the problem at the moment.
This is the model China went with its self-driving pilot projects for example.
Or retrofit our roads with Machine friendly signals? Signs that can ping information to the car, other cars pinging information to the car...Those sound like cheaper alternatives and the models can focus more on the difficult problems/edge cases like overtaking and "oh human like dog is crossing the street" on the freeway.
That’s not the only way of reducing variables. I think the key is rethinking how self-driving cars can fit into a broader ecosystem of transportation, instead of narrowly thinking about self-driving cars as a one to one replacement for human driven cars.
Self-driving or not, cars are a terribly inefficient way to move people. 10 buses that are only a third full replaces 260 single commuter cars. A light rail running at 50% capacity carries more people per hour that 4 full lanes of highway [1]. Even if we build more and more roads, that just causes more traffic, not less [2]. The problem with public transport is the “first mile, last mile” problem. Unless you are in a dense well-designed environment, it is a huge pain getting from your house to the public transport, and from the public transport to your office/store/ballet recital. People don’t want to drive 3 miles and park their car at the light rail station, ride to the closest stop, take a bus 2.5 miles, and then walk the last half mile to the office. But if a self-driving car could pick you up at the curb and take you right to the lightrail, and another could pick you up from your stop and drop you right at the office, suddenly that changes the calculus. Why sit on the highway for 90 mins in your self-driving car, surrounded by 100,000 other people in self driving cars, when you could make the same trip in 40 mins using the light rail and a couple Uber-driverless.
In that situation, you can reduce variables by having self driving cars drive at lower speeds (in residential and urban areas) and by having a few dedicated lanes on strategic artery roads to move people to central locations for transport on bus/lightrail/train, much like we have dedicated lanes for buses in some places.
Once you have some basic infrastructure like that, then it makes more sense to reduce the massive amount of streets and parking wasting space in dense urban environments. If you want a car for your suburban life, fine, but you have to park it outside downtown and take the lightrail in and grab a little self-driving car to get where you want to go.
[1] https://en.wikipedia.org/wiki/Light_rail [2] https://www.wired.com/2014/06/wuwt-traffic-induced-demand/
Almost certainly due to the fact that these vehicles tend to be surrounded by flashing lights in 'random' shapes and colors, and that is confusing the cameras and AI.
"The person in the driver's seat is only there for legal reasons."
Can't believe what they've been allowed to do until now.
What wavelengths do other cars' radars operate at? How big are the apertures?
Elon may not be a great engineer but he is a genius businessman in many aspects so that would also make sense.
To a layman the sole reliance on cameras seems like one of Tesla's biggest mistakes.
This is quite subjective. My opinion is not very representative, but I for one find teslas horrifically bulky and ugly (and most other cars, in fact). They all look like gigantic broken eggshells. On the other hand, a lidar rig looks very cool.
There is no reason to not to have them in a car priced $100k, even just as a failsafe backup to the videocamera.
In the future though, mass produced single chip radars have all the chances to beat every other sensing method on cost.
Lidar on the other hand has no problem with stationary objects, but cars can’t depend on it in all weather conditions. Which means full self driving cars need to be able to operate based on vision like we do. I would expect all level 5 systems to use Lidar especially if costs continue to drop, but it’s not as clear cut as you might think.
But you are right on the point of there being no silver bullet sensing technology. All of them have their own very glaring weaknesses.
Radars are superior for simple collision avoidance for them being able to directly sense the velocity, but they are obviously useless for lane keeping.
Lidars are super unreliable from both internal factors, and the environment. Basically, a splat of dirt, and a collision.
Cameras are subject to all uncertainties of a CV system, and smog, smoke, and weather as well.
Unless they remove them and rely on only one type of sensor, which is exactly what Tesla has done. It's like complaining that each sense has issues, then removing all but one. Multimodal integration is literally how we use our senses normally.
RADAR is great for collision detection and early breaking systems because drivers don’t depend on them. If they fail in an early breaking system noting worse happens. But self driving systems can’t accept failures like that. These systems would need to stop if their camera/Lidar system failed so what’s it actually useful for?
From my understanding, the problem is the software doesn't know if that radar object is a car or a sign or an overpass. So it assumes that moving objects are cars, and makes a good guess at stationary objects. Occasionally when it is wrong, there would be unusual braking for signs, or just plowing into stationary cars.
RADAR systems are rapidly improving, but there’s a reason everyone shows Lidar data over RADAR.
Now in the long run, for true driverless cars, LIDAR will be rendered obsolete by sufficiently accurate AI interpretations of camera feeds. But in the medium term, I think they are key to filling in huge weaknesses camera only systems have, and freeing up computing power currently being used by trying to judge “depth perception” using cameras, to be freed up to better deal with other things cameras are actually good at like color/text recognition for reading traffic signs, facial recognition for determining if a pedestrian sees you, differentiating between hard and soft objects, etc.
[0] https://www.automotiveworld.com/articles/lidars-for-self-dri... [1] https://asia.nikkei.com/Business/Automobiles/Cheaper-lidar-s... [2] https://news.itu.int/the-price-of-lidar-is-falling/
Yet, that's the very point. These very primitive radars are more than enough for a car to avoid an obvious collision threat.
System redundancy in safety-critical systems exists for a reason. Either their engineering team refused to see this or they were unable to make Product/Management/Regulatory people see it.
AEB is quite finicky and needs quite a fair bit of calibration and mechanical alignment, still like the article says can suffer from false positives.
But although LIDAR is still prohibitively expensive it is possible to design ADAS that can integrate stereo camera, RADAR and long range sonar, that would at least give you three sources of data for some margin.
> What wavelengths do other cars' radars operate at?
The spectrum used for AEB in most transceivers is 76 GHz to 81 GHz (so between about 3.9 and 3.7 mm) some older implementations also worked at 24GHz i think.