Scientists from Germany have presented an algorithm that taught a robotic arm to play table tennis. The whole learning process took 1.5 hours.
Researchers at the University of Tübingen in Germany began by developing a computer simulation in which a robot arm equipped with a table tennis racket returned ping pong balls across a virtual table tennis table.
The researchers ran a simulation so that a machine learning algorithm could learn how the speed and orientation of the paddle affects the trajectory of the ball. After the algorithm, which learns by trial and error, was able to hit the balls in more situations, the researchers tuned it to control the movement of the arm of a real robot next to the table.
The system used two cameras to track the location of the ball every 7 milliseconds. The algorithm processed the signals and decided where to move the robotic arm to hit the ball.
The whole process, including training in simulation and in the real world, took only 1.5 hours. The scientists note that they did not expect the algorithm to learn so quickly.
However, despite the fact that the robot plays well even against humans, it is confused by too strong and weak hits. “If the ball flies slowly, paradoxically, the robot must develop high speed. When the robot did not succeed, the ball would often slide off the racket, ”the researchers noted.
Scientists were surprised that the algorithm did not cope with slow hits. They also find it interesting that sometimes the hitting difficulties were due to the mechanical limitations of the robotic system, rather than due to flaws in the algorithm.