DARPA‘s Urban Grand Challenge ran over the weekend, with CMU/GM taking first place, and Stanford/VW taking second. This situation was exactly reversed for the 2005 Grand Challenge, which means the rivalry between Stanford and Carnegie Mellon is building to epic proportions. More info here and here.
I’ve been following this challenge for a few years now, using the Nova special on the 2005 race for my classes. The big change for this race involves the car responding to other agents in its environment, including other moving cars (driven by professional stunt drivers), and for obeying all traffic laws, including right-of-way laws at intersections.
I went to a couple of the site visits and the first thing (one of) the vehicles did for me was a three-way turn. Now, imagine you’re watching this vehicle all by itself do a three-way turn and then come to an intersection, and there was a car there already and when it pulled up, another car pulled up after it. It knew enough to wait for the first car to go because by the rules, it knew that car had precedent. But it also knew that it had precedence over the other car that showed up after. It was stunning.
I mean it was spooky because they went down the road, they made a turn. And he turned to me and he said, ‘Now look, there’s nobody inside there right?’ I said, ‘No, no, there’s nobody inside there.’ He said, ‘Now, and there’s nobody controlling them remotely right’ because it looked like they were being driven by somebody. Now these were the two vehicles that got the furthest, by the way. |link|
There remains the critical problem that the robots are still treating other drivers merely as moving objects, and not as full-bodied agents. This allows the robots to make a host of assumptions about these other drivers, for instance that they will also obey precedent laws perfectly, but these assumptions will not reliably hold in the real world situations that the military desires. The link above suggests that the military is still on track to make 1/3rd of its ground vehicles autonomous by 2015, but there is still more work to do.