Why it matters: Google’s artificial intelligence company DeepMind has developed a robotic arm that can compete with the best amateur table tennis players. It can perform backhands, forehands, a good dose of spin and even shots that skim the net, all with remarkable agility.
In a recent research paper, Google subsidiary DeepMind revealed that its robot beat amateur-level opponents in full table tennis matches in 13 out of 29 games. Granted, it’s still not up to par with real pros, but being able to reach amateur-level skill is nonetheless an impressive feat for an AI system.
MIT Technology Review noted that human players who played against the robot enjoyed the matches. They said it was an engaging challenge that could help them improve their game as a training partner. The video shows the robot deftly handling a variety of volleys and playing styles. It even appears to “jump” like a human during one particularly intense game, even though it has no legs.
“A few months ago, we thought the robot might not be able to win against opponents it had never faced before,” said Pannag Sanketi, the DeepMind engineer who started the project. “The system has certainly exceeded our expectations. The way the robot outperformed even powerful opponents was mind-blowing.”
DeepMind used a two-pronged approach to train its ping-pong automaton. First, the system mastered its hitting skills using computer simulations that mimicked the realistic physics and gameplay of table tennis. Then, the team honed those skills by having it learn from real-world data.
During live matches, the robot uses a pair of cameras to track the positioning of the ball. It also uses motion capture technology to track the movements of its human opponent via an LED-equipped racket to help it identify them and their playing style. All of this information is collected and fed back into the simulations, constantly improving tactics through a continuous feedback loop. In other words, it gets better as the matches go on.
The system does have some limitations, though. The robot struggles to return extremely fast balls, balls that are very far off the table, or balls that slide low. It has also struggled against players who can put exceptional spin on the ball, because it can’t measure the ball’s spin — at least not yet. DeepMind believes that improved AI predictive modeling and smarter collision detection could help solve these problems.
The project sounds like fun and has few practical applications. Still, the report emphasizes that it represents an important step toward creating AI that can perform complex physical tasks safely in natural environments like homes or warehouses.
