Slashdot Mirror
Self-Driving Cars Should Be Legal Because They Pass Safety Tests, Argues Google (theverge.com)
An anonymous reader quotes an article on The Verge: Chris Urmson, director of Google's self-driving car project, has sent a letter to US Transportation Secretary Anthony Foxx today with a plan for selling autonomous vehicles that have no steering wheels or pedals. The plan appears to be pretty straightforward: Urmson argues that if a self-driving car can pass standardized federal safety tests, they should be road-legal. Urmson adds that regulators could 'set conditions that limit use based on safety concerns.'
About three years ago I accidentally let my license expire and thus had to re-take the driving component of the exam.
I am somewhat convinced you could pass it with a non-autonomous vehicle having no steering wheels or pedals.
We can't write a 100% working OS for a phone. Please trust our software with your life.
..and lots of them have been proven later to be unsafe anyway. The law cannot account for everything.
Any ordinary car driven by a raging retarded monkey would pass the safety tests as well.
BECAUSE THE SAFETY TESTS ON CARS DOESN'T TEST DRIVING OR COGNITIVE SKILLS!!!!
There are plenty of computers in use (a lot of the better ones are running Linux or an RTOS and hell, even Windows NT/CE/XP) that people trust their lives to implicitly on a daily basis in a lot more delicate situations than driving a car. Commercial planes do most of the flying fully autonomous, most of both your debt and savings is being invested fully automated, any machine in a hospital parses a lot more data than a few dozen sensor and requires much more precision.
Custom electronics and digital signage for your business: www.evcircuits.com
As long as Google is willing to take responsibility for the damages, I'm fine with it. Basically, Google is the driver of all driverless cars they sell.
No, the owner of the car should get regular car insurance, and then the insurance company will take up the responsibility for the damages. Of course, if the insurance company can argue that google has been grossly negligent in some case, they can take them to court.
Have you seen a commercial plane without human pilots ? I thought so.
That's not the point. The computers on board a commercial plane have the potential to cause major accidents that the pilots would be unable to prevent. And still we trust them.
Now, I can just about grasp that a self driving car can be constructed that will navigate on the road, but that is not all that a car has to do. Let's look at a couple of examples:
1) Suppose I live on a small farm or ranch and you are coming to visit me in your car. I might say "When you get here, come up the drive, turn left at the old tractor and park behind the barn next to the chickens". With a conventional car this should be easy, but what if you have one of these Google cars with no controls. Presumably it will find my address and arrive at the end of the drive. Given that there are no manual controls, how would you tell it the bit about the tractor and chickens? Will you just be able to type that in and it will be clever enough to follow those instructions?
2) What about parking at work? I work on a big site with several car parks. How will I describe to the car which one I want to park in. They don't have separate Zip codes.
How will a computer respond to a tire blow out on the highway at 60mph plus?
Within one millisecond after the blow out, the computer will get data from a tire pressure sensor that indicates what happened, and which tire blew out, and take the appropriate action. They'll be in a much better position to handle this gracefully compared to a human driver that has never experienced this, and maybe requires a few seconds to realize what the hell is going on, and has no clue how to react safely.
Has any of these scenarios been tested? I don't see any crushed google cars so I am going to guess NONE.
Why would you assume Google would notify you of any experiments they have done with blown out tires ? And if they did it right, there wouldn't be any crushed cars.
The point of GP was that computers can't be trusted with life or death decisions, because even our phones have bugs. To which somebody else responded that we already trust computers with our life, such as those in airplanes or medical equipment. The fact a modern plane still has human pilots does not negate that, as the humans can't override every single computer action.
"So if it makes too many mistakes or gets into too many accidents it looses its license."
You now have one point on your spelling license.
BeauHD. Worst editor since kdawson.
An algorithm? Just one?
You must not know what an algorithm is. Hint: 1 algorithm + 1 algorithm = 1 algorithm.
I can tell you have never experienced a blow out that's for sure.
That's right. And I probably wouldn't handle it correctly, even though I'm human.
lol, now put the driverless car on a highway with thousands of other cars, fog, rain, etc... and let the fun begin.
Highway driving is the simplest of all. The car can see through fog and rain using radar and infrared, and can accurately measure the speed of all the cars using doppler, measure the slickness of the road, and accurately determine safe speed and driving distances. These extreme circumstances are actually much harder to handle for humans, as you can clearly see every time there's heavy snowfall, and cars just pile up left and right.
Oh, I love this game! The Luddites bring it up every time autonomous cars are discussed.
How will a computer respond to a tire blow out on the highway at 60mph plus? Some other emergency?
Within a few milliseconds of the emergency being detected by sensors, the computer will have fully assessed the situation and determined the safest course of action. A blown tire is simple, because it only really affects the vehicle handling parameters. At over 60MPH on a highway, the vehicle is going to have very minimal handling needs. The steering system can be told (within those milliseconds) that it will need to adjust, and in a few rotations of the tire, it can analyse the shape of effects of the new tire's shape. At minimum, it will know that it needs to steer a few degrees to the side of its intended course, allowing it to stay on course and maneuver safely to the shoulder.
How will the car that gets slammed into by an autonomous vehicle with a blown tire respond?
If by some absurd accident that does occur, it'd be treated like any other unavoided collision. As soon as the vehicle determines that a collision is unavoidable, it will attempt to minimize the damage. There has been research into having algorithms adjust the vehicle speed to change impact position, relying on a database of the vehicle's crush characteristics to reduce the chance of injury. Results show that the computer can do that faster and more successfully than any human driver.
How will the cars behind it react to the event in front? How fast?
When the front car detects the blowout, or when the cars behind it detect the debris or change in driving characteristics, they will consider the car to be a risk, and avoid it. They will start slowing down, changing lanes, and otherwise avoiding the affected vehicle. Again, this has all been tested.
As for the speed of this decision, everything happens in a few dozen milliseconds. At 60MPH highway speeds, that means that the computer will process and understand a situation before the vehicle has traveled a few feet. In comparison, a human brain reacts in about 150 to 250 ms, depending mostly on the type of stimulus. It doesn't matter how good of a driver you are, or how much attention you're paying to what's going on. An autonomous vehicle can observe, consider, and begin reacting to an emergency in front of it before your brain can even understand what your eyes are seeing.
What will happen to the hacker that intercepted and manipulated those signals the other cars are sending to each other? ( assumption made )
Oh, don't be so coy. You've made an awful lot of assumptions without actually understanding the current state of the art.
Legally, probably nothing will happen to the hacker, because it'd be difficult to find and catch such an attack, but that has nothing to do with the cars. It's equally hard to find someone using a cell phone jammer today. There have been a few cases, but they were caught due to prolonged or repeated use.
From the perspective of the cars on the road, losing communications with the other vehicles is a known and tested risk. Similarly, mismatched information is a risk condition. The easiest response is to slow down and try to move out of traffic.
Has any of these scenarios been tested? I don't see any crushed google cars so I am going to guess NONE.
Yes, these scenarios have been tested extensively, but you don't know about it, because you're not bothering to do research. One of the fascinating aspects of robotics research is that researchers can control all of the inputs to the algorithms. We can put a sensor in a tire, then drive it down a test track and make it blow. We can take that data, mix it with data from a dozen other test runs, and use that to build thousands of simulation input cases. Those inputs are run in simulated environments, with and without vehicle-to-vehicle
You do not have a moral or legal right to do absolutely anything you want.
The thing is that a third party would actually be responsible for the accident, and as such you would need additional coverage specifically for the case. Insurance policies spell out such things. They want to even know how much you drive it on average, where you drive and in what conditions. They take into account things like if it is a motor cycle all the way up to a tractor trailer. The price of insurance is based on these factors.
Now add the concept of hardware driving the car. Sure, we can demonstrate that this car with this software is this safe, but then comes a software update. This makes the insurance company's management of their risk much more difficult. All it takes is one bad update and lets say, hypothetically, a vehicle or two careens into the side of a bus. Some analysis and a patch later and all of a sudden we are at a new version without the bus smashing feature, but did that update add a toddler avoidance error?
AI "drivers" receiving software updates only add to the chaos that is our roads. No one has died yet, but someday someone will.
Look up TCAS sometime. The planes have sensors to detect each other. If the TCAS system detects a possible collision situation, the planes determine, all by themselves, the correct course of action, and then relay that information to the pilot. Commands like CLIMB or DESCEND or STAY LEVEL. In this situation, the pilot has absolutely no say in the matter. They are required to obey the computer because in the past, pilots ignoring this input have cause planes to crash into each other in mid-air because the pilot thought he knew better. The TCAS commands even override Air Traffic Control commands. How's that for trusting your life to a computer?
"Freedom in the USA is not the ability to do what you want. It is the ability to stop others from doing what THEY want"
Unless the weather isn't conducive to VFR, in which case, guess what? The autopilot lands the plane. In situations where the pilot cannot be trusted to land due to poor visibility or other issues, the autopilot is king. I've taken an A320 class D flight simulator from ATL to JFK without touching the controls except to set the throttles to the climb position from TOGA, setting the flaps, and landing gear position.
"Freedom in the USA is not the ability to do what you want. It is the ability to stop others from doing what THEY want"
An algorithm? Just one?
Yes, just one algorithm. As slashping points out, it's the combination of other smaller algorithms:
All of those parts get bundled together into a single unit for designers to track, and it'll probably be documented somewhere as the "tire blowout algorithm".
Lets crank it up a notch. The tire didn't blow out, but the tread separated and wrapped itself round some structure of the undercarriage and locks up the rear end. On asphalt pavement with a ton of road grit and gravel.
Lol, now put the driverless car on a highway with thousands of other cars, fog, rain, etc... and let the fun begin.
Put those cars on the r[o]ad and let the slaughter begin.
Ok. Where's the hard part? What you're describing is a worst-case situation, and most of it doesn't matter. Fog and rain only affect one out of several sensors, and a road with "a ton of road grit and gravel" isn't unusual. That's expected for a road. Every time the software checks its own state, it considers the road condition. If it doesn't have "a ton of road grit and gravel", then the vehicle doesn't need to compensate for that.
What is being built is a perfect super-human driver. It has access to senses that humans don't, never gets distracted, never panics, and never deviates from best practices. It has spent several decades of (sim) time in simulations, dealing with hundreds of thousands of hazards, and now you're worried about rain?
You do not have a moral or legal right to do absolutely anything you want.
Yes, this is real life, not your if-it-ain't-broke-don't-fix-it fantasy.
You can choose to ignore what I've written, as you've chosen to ignore the other research and progress in the field. Looking through your comment history, you seem to do that a lot. You also seem to have a bias against modern automotive safety, based on myths that have been continually disproven.
Yes, it's true that there are faults to modern technology, especially with regards to safety equipment. However, you don't seem to understand that the risk of new technology is far less than the risk presented by older technology. A driver in a collision who has an airbag may break a collarbone, but he won't have frontal-lobe trauma as he would if he were allowed to hit the steering wheel.
These things can (and are) tested. What's important is that the circumstances of the test be identical, so the difference in effectiveness can be measured. One good way to do that for vehicles is to take an old vehicle, and crash it into a new vehicle. That way, thanks to Newton's third law, we know the impact force is exactly identical. In fact, here's a good demonstration.
You do not have a moral or legal right to do absolutely anything you want.
Why do you move to the things that humans are worst at? Everything you're describing is *super easy* for a computer to do better than humans. You've picked scenarios with an incredibly low bar for software to cross.
What you should be picking are things humans find easy and safe, but which are nonetheless unusual. Like "the traffic lights have failed and a private citizen has taken it upon himself to direct traffic" or "driving into a restricted off-road area which is closed to the general public and blacked out by satellites etc.". Until that's solved (and I have little doubt that they will be solved), self-driving cars will likely need a human backup driver.
You're kind of like the person from the 80s who thinks a computer may recognize writing someday, but can never beat a world champion human at chess or go due to a lack of creativity. Turns out, beating humans at games is relatively easy for computers. Solving captchas is hard. In your "self-driving cars will never work" rants, you should look for scenarios that are like captchas, not scenarios that are like chess.
An automobile not only has many, many more variables but it has variables that apply to it as well as variables that apply to other vehicles and terrain types. Call me back when your AV can get me to my home, in a snowstorm, without having been there before. Then, it needs to be safer than I am.
My home is in NW Maine, near the edge of cellular connection but there are actually two towers to work with, it's on the side of a mountain, the driveway is paved, it's about a half-mile long. It's steep but not too steep. The road to get there varies, it's in the mountains and foothills. It's not unknown to have 12" of snow. Your best bet would be something in a ruggedized SUV if you want to make this work.
I need to be private, completely private. The vehicle has some serious drawbacks due to the required size and resistance so it's inherently less efficient than it can be so that's a serious concern. The house is 24 miles from the village, the village is Rangeley, Maine if you want to see it on a map. I've actually needed to put a vehicle into reverse to keep from continuing forward - it was ice on the side of a mountain and hadn't been sanded yet.
I need to be able to operate this vehicle in the city as well. It has to meld with and ensure it does not interfere with local traffic, shipping traffic, and conditional traffic for those who might drive for the purpose of enjoying the trip and not just getting from one location to another. I want to do this privately. I can travel, with relative anonymity, currently - I do not live in an area where this is tracking - I do not want to.
When I am at my home, I can go through the woods and be in Canada. It'd be against the law but I will not cross a routed highway to get there. I have vehicles that will accomplish this task. It is a mere 40 miles, it's on logging roads, a old dirt road (without power lines or anything - some dude does live up there though and he's gonna need one too), and then I'm on snowmobile and ATV trails all the way into Canada. It crosses the border just above the actual border crossing. You can actually see it on the map - just use the address I mentioned and Google.
I don't want to be tracked. That's a requirement. I'd go so far as to say it is a right. More so considering I'd be needing to use it in areas where I have ever expectation of privacy - I'm quite frequently on private property. I should not be mandated to use it to the exclusion of all other vehicles. Access to the federal highway system should not be restricted for me - nor should state highway systems - even if tracking occurs on those sections.
There are many others who are in more interesting situations than I. I've not even begun to get into some of the many, many difficulties that need to be overcome in order to make this happen and happen in a reasonable manner. Hand-waving it away with airplanes, things that land under limited conditions, with limited surfaces, in limited places, and with limited locations, doesn't cut it. Let me know when the plane picks up passengers, departs, taxis, and makes it over to the mechanic's shop on its own - without human intervention. 'Cause having humans drive at the beginning and at the end is not autonomous, it's partially autonomous and that's not what we're discussing here.
Full autonomy is going to be needed, otherwise you end up with idiots who don't know how to drive and they're operating it only under conditions where computer is able to deal with it. Do you know where the largest number of accidents occur? In parking lots. They're not serious but, as a percentile, they're something crazy like more than twice as many - and that's not even extending it out to roads that are less than 10 miles from the destination. That's where, that and parking lots, crash - and those are often the areas where people live - or not far from them. You can't just go, "Your turn now!" and expect it to work out well in the end. That's just retarded.
So, sure... Make me one. Make me one that's affordable. Hop right on it Go
"So long and thanks for all the fish."
Because a serious failure in judgment of an AI car would be vehicular manslaughter if it killed someone.
Laws are rules for the court, but merely a bottom bar to hit for life. Think beyond laws in your actions always.
After almost two decades on Slashdot and a decade in industry it's dead accurate. You guys whine constantly about new technology. I can almost guarantee most of you are the guys in the office that insist on doing things 'the way you learned them' 20 or 30 years ago. Where I work you're the barriers to progress. Up to a point it's just cheaper to keep you around until it's easier to start over with a fresh grad or a H1B and train them on new tools.
This isn't a CS problem, this is a mechanical engineering / dynamics problem. Sometime along the way they figured out it would be easier to teach engineers to code than it was to teach coders to engineer. We use tools that most of Slashdot thinks don't even exist, like Simulink (a drag and drop coding language of sorts).
but lets do some *real* due diligence before we so blindly strap ourselves in.
We are. But the test cases that you've come up with are quite literally the lowest hanging fruit there is. A flat tire? Most cars auto driving is going to come with first come with run flats. (BMW, Benz). For those that don't it's still one of the simplest FMEAs to possibly do. You put a car in the test cell and blow the tires over and over and over, collect data, toss it in the HIL and test your controller. Toss the controller on a car. Repeat.
And I can guarantee that the self driving car is going to handle a blown tire better than you ever will. It'll handle it better than race car drivers will.
and you don't grow old by being stupid.
And you don't stay gainfully employed by doing things the old way.