This Robot Can Beat You at Ping Pong

Robots are able to greatly surpass human-level speed and performance when it comes to well-defined and repeatable tasks, like many of those that are needed on an assembly line. But in most real-world tasks, like cooking, cleaning, playing sports, and so on, the performance of a robot is laughable at best. This is because working in unstructured environments under variable conditions is much more difficult than following a hard-coded algorithm. And the challenges associated with these situations are preventing the development of practical general-purpose robotics systems, such as domestic robots that could do our chores for us.

Researchers at Google DeepMind took on a fairly challenging problem as a way to start working through some of the issues standing in the way of general-purpose robotics. They have created a robot that can play table tennis, and it is actually pretty good. It may not be able to take on a professional player (at least not yet), but it can give most people a run for their money. The work is not all fun and games, however. The team believes that their efforts to tackle high-speed motion, precise control, and real-time decision-making will pay dividends for other applications as well.

The hardware consists of a robotic arm (a 6 DoF ABB IRB 1100) mounted on a pair of Festo linear gantries and a 3D-printed paddle coated with rubber. Dual Ximea MQ013CG-ON cameras capture images of the ball at 125 frames per second to feed into the perception system. Additionally, a PhaseSpace motion capture system — with 20 cameras around the area of play — is utilized to track the opponent’s paddle position.

The control system was divided into two segments — a high-level control system that makes strategic decisions about the moves to make, and a low-level system that executes the physical skills to carry out the moves. Strategic decisions are made by a neural-perception system that is supplied with images from the robot’s cameras and the motion capture system. Based on the robot’s position, the opponent’s actions, and the position and spin of the ball, it makes a decision about which action to take. The physical action is then carried out by the low-level system, which was trained with moves like a forehand topspin, backhand targeting, and a forehand serve.

Beginning table tennis players, beware! The robot beat beginners in 100 percent of matches. If you have a bit of experience under your belt, however, you will be in much better shape. Only 55 percent of intermediate players suffered a loss to the machine, while advanced players were able to beat the robot every time.

How the robot fared had a lot to do with the pace of the game. The faster the ball moved, the more trouble it had keeping up. This was attributed in large part to latency, so if the researchers can speed up the system, either with more efficient algorithms or beefier hardware, it might eventually be able to hold its own with the Chinese Olympic table tennis team. But bragging rights aside, those advances might also lead to the development of domestic service robots, which would be a much bigger win for us all.