The robocalypse may be nearing, but in general the bots you see are pretty much just following carefully-prepared human orders. Until now. Turning the fear-factor up to 11, a researcher has a robot that can learn and generalize.
Honda's Asimo android can, actually, perform some generalization actions, and even work out how best to dodge around you in one of those awkward head-to-head corridor encounters we all find ourselves in sometimes. But Asimo isn't half as smart as the arm-like manipulator robot PR2 that they've been working on in the University of Southern California.
The amazing algorithms powering PR2's machine-vision/object manipulator system can actually learn how to do complex tasks after being led through them by a human. But instead of just being able to dumbly reproduce the motions--and amusingly cocking things up when the real-world situation doesn't match up with the training one--PR2 can generalize the task and work out how to successfully complete it even if the situation changes dynamically. Check out the video:
Amazing huh? The whole point of this is to get robots smart enough to cope with all the tricksy variable conditions in a real world scenario--one where the cat knocks over stuff on shelves, we never put cutlery back in the right drawer, and where you can't park your car as near as you'd like to the gas pump in a filling station. This research will be vital in our upcoming robo-butler assisted life...and it's more than a little creepy (particularly the way it ever so slightly dents the can, hinting that on no more than a whim it could crush it into atoms).