Here we see what Petar Kormushev, Sylvain Calinon, Ryo Saegusa, and Giorgio Metta get up to when they’re not teaching robot arms to flip pancakes at the Italian Institute of Technology. They’ve taught the iCub how to hit a bull’s eye with a bow and arrow using a learning algorithm cleverly named ARCHER (Augmented Reward Chained Regression). The robot uses image processing to detect the distance between the arrow’s tip and the target’s center at impact. The center of the target gives the highest reward, so the robot adjusts its aim based on the previous shot in an effort to hit the bull’s eye. This requires the robot to not only position the bow, but also hold the arrow in its other hand between its dainty little fingers. After just a few trial runs, the iCub has learned to reliably hit the target dead-center. The robot is only about 3.5 meters away, but hey it’s a start.
The project will be presented at the Humanoids 2010 conference. More photos can be accessed at the Flickr set.