Slashdot Mirror
VW Raises the Bar for Self-Driving Vehicles
Old Man Kensey writes "According to the UK Daily Mail, VW has produced a prototype Golf (code-named "53 plus 1" in a reference to Herbie the Love Bug) that successfully steers and accelerates itself at speeds up to 150 MPH on tracks designed on the spot without pre-programming. It sounds almost too good to be true given some of the problems CMU's prototype has had over the years, but perhaps VW has learned from and extended CMU's research (and within-an-inch GPS positioning probably helps too)."
This is a good illustration of why research funded by a corperation is more likely to achieve results than that of academics. Academics are free to pursue whatever is most interesting as they work, and it is ok to get off on a tangent as long as some papers come out of it. However if you work for a company you need to get results, hence this car. Of course this model doesn't work quite as well for theoretical physics, but well enough for the computer science. I suspect we would have AI already if it could be turned directly into a product.
Philosophy.
And that bug is probably fixed by now, but the problem is, how do we determine we worked out all the bugs? We can't even do that with Linux/Windows/Anything. The closest we come to that in the OS world is a microkernel with only a few thousands lines of code and controlled input.
But how do we ever determine a program that learns and is subject to varying, uncontrolled data inputs is bug free? You can't and I wouldn't want to see the first literal blue screen of death when it happens.
I don't want to sound like a luddite, but the article mentions that planes have been flying autopilot (did they forget to mention landing/taking off is still done by the pilot) since the 1970s. But I believe we'll have flying cars before self-driving* cars because the problem is several hundreds of a magnitude easier in empty 3D space where all you have to do is stay high enough off the ground and avoid collisions via radar/whatnot.
*The only way is I see anything coming close to a self-driving car is on highways where lanes get marked magnetically and driving problem gets reduce to the car having to stay X feet behind the car in front of it.
Travelling at 150mph on a circuit is easy. Well, relatively anyways.
Now if they managed to get this car travelling at 20mph down a city street during rush hour, we'd really have something useful on our hands.
Don't get me wrong, I'm impressed. But a self-driving car on an empty track is a million miles away from the everyday driving conditions we encounter.
I don't think, that the satellite positions vary randomly in the sense, that they have gaussian variance in a deliberatly short intervall of time. But their positions contain a systematic error, which can be determined via a fixed known position (actually more, but who cares) and thereby be corrected. This, in general, is the principle behind DGPS. The accuracy does not depend as much on the position of the satelites, but the discrepancy between the systematic error between the fixed known position and the unknown one.
That said, I'm still sceptical concerning the quoted accuracy. Especially for a moving object, like a car.
"Between strong and weak, between rich and poor [...], it is freedom which oppresses and the law which sets free"
This is surely the perfect car. I can just imagine the scene:
Car pulled over by the highway patrol for doing 150 in a 65 zone.
Officer is puzzled by the fact that the only person in the car is asleep, in the back seat.
"Did you know what speed you were doing, sir?"
"Huh, um, wha? Oh - the car was driving, Officer".
Car has to appear in court next Wednesday.
The holy grail is cars that talk to each other to get around more efficiently yet. If the traffic up ahead narrows from four lanes to two because of construction, and car computers can talk to each other and say "Hey, you're two miles back but get ready for this", then orderly traffic flow can be maintained as the cars merge into the remaining lanes and decelerate. This in turn saves gas, etc.
Hell, think how much money you'd save if you car just automatically avoided potholes if it could. Tires, struts, shocks, suspension, all those would last much longer. Look at the figures on how much money it costs drivers annually in a city like Baltimore that's infested with chuckholed roads.
-- Old Man Kensey
And your point is? These types of errors you list (including satelite path deviations) are presisely what DGPS corrects for if the well known GPS is in relatively close proximity to the onboard GPS. In that case, these unknown variabilities will be reasonably well correlated so they will be removed when taking the differential. (ie. satelite orbit changes, ionospheric refraction etc., though unpredictable, will be nearly exaxtly the same for both GPSs so it gets subtracted out). In my work we generally get DGPS accuracies of less than half a meter which is well below your quoted error of 3 meters RMS.
For use with autonomous vehicles, one can generally do a lot better when the DGPS is augmented with a ground based equivalent of GPS like Terralites (XPS) which can and do routinely give accurate positions in the 1cm range.