Fact Checked by 
While sports being played by robots sounds like a sci-fi film plot (anyone remember Real Steel?), it has quickly become a reality as Google’s AI company DeepMind has created a robotic arm that play table tennis.
And it’s pretty good too, having beaten amateur-level opponents in 13 out of 29 games. While the consistency cannot be confirmed, this move by DeepMind marks the first time a robot has been taught to play a sport with humans at their level.
An MIT Technology Review noted that the human players even enjoyed the matches, with a study participant saying: “I would definitely love to have it as a training partner, someone to play some matches from time to time.”
This was done by the team of researchers training a “skill library of low-level controllers, including service and rallying, and sim-to-sim adapters from a dataset of all states.
“Using the same ball states, we train a high-level controller for style selection. The policies are trained purely in simulation (but using real ball states) using Blackbox Gradient Sensing.”
The robot also has 20 cameras mounted on it to track the opponent’s paddle. When it’s all working together, the neural perception system tracks the ball, the motion capture system tracks the human player’s paddle pose, a state machine tracks the state of the game, and an observation module provides data such as ball position and velocity.
Google DeepMind robot beat all beginners it played against
The sports-based robot was trained on the knowledge of a table tennis coach who gave the team insights into the types of skills a person (or technology) should train for.
When DeepMind put its new contraption to the test against real players, it beat 100% of matches against beginners. It won 55% of matches with intermediate opponents but wasn’t able to win a single game against more skilled players.
Some of the people who went head-to-head with the technology described it as being intimidating and loud. While the robot isn’t yet up to scratch with the best of the best, its skills could be improved upon in the future.
Featured Image: Via Google DeepMind
