Monday, April 1, 2013

Self driving cars, again

Another week, another study.
The latest study into self-driving cars has stated that "The thought is that autonomous cars will reduce the number of traffic deaths — more than 100 people per day, currently". Marketing doublespeak might make you think that they mean self-driving cars would save 100 lives per day. Actually, what it means is that currently 100 people a day are killed in car accidents, and that that number might be reduced, but they don't guesstimate by how much.
In truth the same could easily be achieved much more effectively by proper driver training and annual re-assessment. We don't license airline pilots and train drivers once and then leave them to it - they get annual or bi-annual re-assessments. They're human just like you and me, yet in America we're allowed to take an open-book written test in some states, and shockingly, people still fail even when they're given the answers. That's followed by 8 minutes of the dumbest driving test in the history of the universe and then no re-testing until we're dead.
The key phrase in that study was "the thought is". We thought the earth was flat at one point too, and millions of people think American Idol is important (hint: it's not).
In America, we have an appallingly easy mechanism of getting untalented, unprepared drivers on to the road, reflected in the accident rate here. 8.5 fatalities per billion road kilometres driven (a good metric - it takes into account population, number of cars, number of drivers and distances driven). In Sweden, where the driver licensing mechanism can take a year, that number is 5.1. In the UK where driver training takes months and the test is phased, it's 5.7. In fact, the only places in the world with a higher fatality rate than the US are places like Latvia, Estonia and Slovakia (Source).
So trying to sell the idea that automated cars will somehow "fix" the accident rate in the US ignores that fact that the driver licensing mechanism is totally broken. Fix that and the accident rate will drop naturally. It's a lot cheaper and doesn't require the wholesale replacement of every vehicle on the road.
That's a serious point - unless every car on the road has this technology, it simply won't work. Say you're in a self-driving car and it sees something and performs an emergency stop. I'm behind you in a normal car and I can't react anything like as quick as your computer. I slam into the back of you with enough speed that my airbags deploy and I shunt your car forwards into whatever it was trying to avoid. That's a very common accident right now and there's nothing your self-driving car can do to prevent that accident.
That assumes the tech works in the first place. Case in point: Volvo's new auto-braking system - exactly what I just described above.

It failed spectacularly at this press event because of "battery issues". Mercedes have problems too. The UK TV Show 'Fifth Gear' recently tested their auto emergency braking system and in every instance (slow, 18mph, 28mph and 60mph) it hit the car in front. Sure they can iron out the bugs - I'm not a total luddite - but ultimately, demonstrations like this prove that even when everything ought to have be set up perfectly, it can still go wrong. And that's my point - if you can't rely on this technology 100% of the time, it shouldn't be being put into cars. Heck - ABS has been around for decades and that still has problems (gravel, snow, ice, mud etc). The last generation Ford Mustang had ABS that would fail when it was raining - the one time you might actually need it.
Look at Google's self-driving cars - loaded up with sensors, cameras, radars, lasers and computers yet they can't merge on and off motorways (they can't figure out speed differences and sizes of prospective gaps for the car) and they can't drive any route that hasn't been pre-driven and stored in GPS. So they're not really autonomous at all - they're simply playing back pre-driven routes.
There's dozens if not hundreds of other examples where this technology, impressive as it is, simply isn't working.
Parallel parking a car is about the easiest thing you can do - it's a low speed geometrically simple problem and you ought to be able to do it pretty much in your sleep. Yet they haven't been able to make a car that can do that reliably. The ones that attempt it have a manual an inch thick telling you all the things you need to do to "prepare" the car for the maneuver and even then it'll fail 6 times out of 10. Do you really want to rely on a car to save you, to think for you, to make decisions for you? When they can't even make a car that can reliably parallel park, I don't.