Watch a robot playing table tennis after just 90 minutes of training
A table tennis-playing robot can keep up a rally against humans, but like many amateur players, it struggles when attempting fancier shots.
Yapeng Gao, Jonas Tebbe and Andreas Zell at the University of Tübingen in Germany began by designing a computer simulation in which a virtual robot arm equipped with a table tennis racket attempted to return ping pong balls across a virtual table tennis table.
The researchers ran this simulation so that a machine learning algorithm could learn how the velocity and orientation of the racket affects the path the ball takes.
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).
The system used two cameras to track the location of the real ball every 7 milliseconds, and the algorithm processed the signals and decided where to move the robotic arm to hit and return the ball.
The signals that the algorithm sent allowed the robot arm to accurately play shots to within an average of 24.9 centimetres of the intended location. This accuracy level was slightly worse than when the algorithm was working with a simulation – a common occurrence, says Tebbe, as computer simulations can’t accurately represent everything in real life.
The entire process – including training in the virtual simulation and in the real world – took just 1.5 hours, demonstrating how rapidly algorithms can learn to perform in a new situation.
However, although the robot performed well against human players, it was tripped up by fast shots – and, surprisingly, by slow ones. “If a ball is slow, the robot needs to generate more speed,” says Tebbe. Struggling to do that, the ball often slumped off the racket.
“By training the system for a relatively short period of time the robot is able to cope well with differences in serve, and capable of returning using a random policy,” says Jonathan Aitken at the University of Sheffield in the UK, who wasn’t involved in the study.
Aitken was surprised the algorithm flunked returning slow shots. He also finds it interesting that it sometimes struggled with making shots because of the mechanical limitations of the robot system, rather than because of shortcomings with the algorithm.
The robot arm has other limitations. For instance, it struggles to play backspin shots, says Zell, because the robot arm is unable to hold the racket at the required angle needed to perform such shots. But despite these issues, he believes the robot is a good player.
“It’s not worse than a regular human player,” he says. “It’s already on par with me.”
More on these topics: