Before artificial intelligence takes over the world, it might first try to beat humans at table tennis. Google reports that a robot powered by DeepMind AI can now play table tennis at “amateur human level.”
DeepMind Labs is where Google works on some of its most cutting-edge AI techniques, and we've seen plenty of DeepMind hands-on experiments in the past, from AI adding audio to silent videos to discovering new material.
In a new paper, Google researchers say their table tennis robot won 13 of 29 matches against humans, with success rates varying depending on the ability of the opposing player, which ranged from beginner to advanced.
“This is the first robotic agent capable of performing a sport at human-level with a human and marks a milestone in robotic learning and control,” the researchers wrote, but added that it represents only a small step forward in the more general field of enabling robots to perform useful skills in the real world.
There's no need to rush
Demonstration – Robot Table Tennis Achieves Human-Level Competitive Skills – Watch on YouTube
Table tennis was chosen as the DeepMind team's project because of the many different factors involved, from the complex physics of movement to the hand-eye coordination needed to successfully hit the ball back.
The robot was trained to focus on each shot type individually, from backhand spin to forehand serve, and this training was then combined with more advanced algorithms designed to select the desired shot type every time.
As expected, the robot struggled most with faster shots (which give the AI less time to think about what to do), and the researchers are already looking into ways to improve the system, such as making plays more unpredictable.
It also has the ability to learn from the strategies of human opponents and weigh their strengths and weaknesses built in. If you're interested in the challenges of training and scaling AI, and how we develop the combination of skills needed for robots to be useful in physical tasks, the paper is worth a read in its entirety.
