Researchers from Google DeepMind have unveiled an A.I. powered robot that is not only able to play table tennis with humans but can win matches nearly half of the time. According a recent paper from the A.I. lab published in Arxiv, it is the first robot agent capable of playing sports with humans at a human level.
The system represents a significant milestone in robot learning and control, especially regarding the ability to scale “robot learning to complex physical tasks which may involve a human partner or adversary,” according to the paper. In addition to winning 45 percent of its games against 29 opponents, matches with the table tennis-playing robot were still enjoyable for more advanced players, who described the system as a promising practice partner.
Table tennis has been a fruitful area for robotics research since the 1980s, according to Google DeepMind, due to the sport’s qualities of high-speed motion, precise control, real-time decision-making and human-robot interaction. But up until now, “no prior work has tackled the competitive game in which a robot plays a full game of table tennis against a previously unseen human opponent,” the A.I. lab said.
The company’s robot, which consists of a robotic arm holding a 3D-printed paddle, was trained on a vast dataset of low-level skills that included aspects like forehand topspin, forehand serve or backhand targeting. Able to select the optimal skill for a particular situation in light of game statistics and the opponent’s capabilities, it continued learning as it played matches and therefore improved over time.
The robot played 29 matches against players with beginner, intermediate, advanced and advanced+ skill levels. It beat all beginner players but lost 55 percent of games against intermediate opponents and won zero matches against more skilled players, suggesting its own skill level is intermediate. More advanced players were also able to identify and exploit the robot’s weaknesses, such as its limitations in measuring spin, a weaker backhand and likelihood to miss low balls due to a collision avoidance protocol in its system.
“It was truly awesome to watch the robot play players of all levels and styles,” said Barney J. Reed, a competitive table tennis coach and one of the authors of Google DeepMind’s paper, in a statement on a website dedicated to the project. “Going in, our aim was to have the robot be at an intermediate level. Amazingly, it did just that. All the hard work paid off,” he said.
After completing three matches each with the robot, players described the games as “fun” and “engaging,” according to Google DeepMind. When asked how interested they would be in playing with the robot again on a scale of one to five, players gave a 4.87 average response.
What impact will the system have on future robotics research?
Games and sports have long been a fruitful area to experiment with A.I. and robotics, according to Google DeepMind. “Since the early days of artificial intelligence, chess and other competitive games have played a critical role in the development of new algorithms and technologies,” according to its paper, which noted that A.I. agents can be found across games like backgammon, poker, Dota 2 and Go. Google DeepMind has previously developed programs like AlphaGo, which in 2016 notably beat Lee Sedol, a top-ranked Go player.
Much work is needed to eventually build robots capable of performing useful tasks and safely interacting with humans, according to Google DeepMind. But its tennis table-playing robotic arm demonstrates a small step in the right direction and towards “a long-standing goal in robotics of achieving human-level performance on many useful real-world skills.”