German researchers have created a ping-pong-playing robot that not only learns its initial move set from a human instructor, but can learn from its opponents and improvise new strategies.
The robot, built by Katharina Muelling and her team at Germany's Technical University of Darmstadt, is equipped with a single large arm and a camera that watches the field of play. The team gave it a set of elementary techniques by a process called "kinesthetic teach-in," in which they physically guided the arm through a move, and the system committed the motion to memory.
Equipped with these "primitives," the arm could watch for the position of the ball using its camera and perform the appropriate move to return it to the other side of the table.
At first, it failed to hit any balls, even ones sent reliably from a ball-launching machine. But the system was designed to learn how to adjust and combine movements, and before long, it was returning 79 percent of shots from the launcher.
Then came the true test of its ability: A real human opponent. People are less reliable than ball machines, and hit to a larger area than the system was "used to." However, it quickly adjusted, and after just one hour of playing against a human, was returning balls 88 percent of the time. It even hit back nine in a row, which anyone new to ping-pong can tell you is no easy feat.
A ping-pong bot may sound like rather a narrow field of research, but work along these lines can be generalized to other types of robots and devices, for instance medical or industrial devices that need to adjust to their environments. Muelling's work will be presented at an artificial intelligence symposium in Virginia next month, but the paper describing the latest advances can be read in its entirety here (PDF).
Devin Coldewey is a contributing writer for NBC News Digital. His personal website is coldewey.cc.
