Once this algorithm, which learns by trial and error, could reliably return the ball, the researchers set it up to control the movement of a real robot arm positioned next to a real table (pictured).