Slashdot Mirror
California Legalizes Self Driving Cars
Hugh Pickens writes writes "The Seattle PI reports that California has become the third state to explicitly legalize driverless vehicles, setting the stage for computers to take the wheel along the state's highways and roads ... 'Today we're looking at science fiction becoming tomorrow's reality,' said Gov. Brown. 'This self-driving car is another step forward in this long, march of California pioneering the future and leading not just the country, but the whole world.' The law immediately allows for testing of the vehicles on public roadways, so long as properly licensed drivers are seated at the wheel and able to take over. It also lays out a roadmap for manufacturers to seek permits from the DMV to build and sell driverless cars to consumers. Bryant Walker Smith, a fellow at Stanford's Center for Automotive Research points to a statistical basis for safety that the DMV might consider as it begins to develop standards: 'Google's cars would need to drive themselves (by themselves) more than 725,000 representative miles without incident for us to say with 99 percent confidence that they crash less frequently than conventional cars. If we look only at fatal crashes, this minimum skyrockets to 300 million miles. To my knowledge, Google has yet to reach these milestones.'"
So the self-driving wonder swerves right to avoid the other car and zooms off the cliff. A human driver would recognize that hitting the other car in this instance is the safer solution then to go careening off the steep cliff.
Someone has never, ever taken an AI class. Or even an algorithm class dealing with risk. Here's how the calculation actually works (and by the way, that approach is about 20-30 years old).
Every situation is assessed an impact value: driving into oncoming traffic, 0 (very bad); driving into the right ditch, 10; swerving into a legal lane, 50; etc. Every situation is given a set of possible actions, with each action having a probability of being completed successfully. The algorithm multiplies the outcome with the odds of achieving that outcome, and picks the highest value. You can set it up in different ways, but the idea is the same: multiply outcome severity with odds of achieving outcome, pick lowest combined risk/outcome. In your situation, driving off the cliff (which is assumed to be very bad, since the car can see a very steep drop-off with no bottom) is going to have a much worse outcome than hitting the car in front of it. Hitting the car in front of it is guaranteed, but so is driving off the cliff. As a result, the algorithm will make the automated car hit the car in front of it, rather than drive off the cliff.
Not to mention that cars don't sleep, always behave optimally (according to the algorithms in place), and have no blind spots.
Basically what I am waiting for is the inevitable 100 car pile up with massive fatalities that WILL occur at some point in time where investigation will identify that a self-driven car, or cars, was the cause of it.
You mean like the ones that regularly happen in fog and icy/rainy conditions?
Any company involved in programming or manufacturing that self-driven car will be sued out of existence and the "love affair" everyone seems to have about auto-driving cars will end quickly.
That is a very real risk. Not sure how the laws will deal with it. But until that question is addressed, we won't see large-scale sales of automated cars. I suspect that we'll see the equivalent of ToS: by using this car, you agree to be fully responsible for all its actions and accidents.
Those who can, do. Those who can't, sue.
Self driving cars *never* swerve. They brake. Statistically they know that swerving almost always is worse than the incoming accident. Humans on the other hand will swerve. See all the accidents that occur when attempting to miss an animal crossing the road.
1) You reaction time is far worse than a computer.
2) Your estimation of distances is far worse than machines absolute measurements.
3) You are limited to two forward facing eyes, augmented by 3 small mirrors. And you share some of the vision time with looking at the dash. An auto-car can look in all directions at once, and monitor all dashboard information and more at the same time.
4) An auto-driver will be better at maintaining a safe speed. Able to stop in the distance it knows to be clear far more often than a human driver.
5) I'd expect an auto-driver system to be seperate from any other computing devices in the car, and connected to the internet or any other vector for hacking. I'd expect them to be as immune to hacking as an auto-pilot system in a plane.
Why do Luddites chose to come to Slashdot?
Did you mean this forecast to sound neutral, damming, or hopeful?
I see it as hopeful. Driving a car on a public road isn't a right. If you want to drive with manual control, do it on a road you paid for yourself.
Grammer Nazis - I mod you "troll" unless you actually add something on-topic. Yes, I know I have mispellings in my sig.
That brings up another thing autocars will be better about than humans. Individual humans can learn from their mistakes, but that knowledge is not directly transferable to other humans. Any mistake a self-driving car makes, however, can have its solution incorporated into all self-driving cars (or at least all the ones of that model.) So, lots and lots of testing should ultimately give us very safe and effective cars.
Check out my world simulator thingy.