Google-powered robot beats humans at ping pong

An embedded robotic arm with technology developed by Google DeepMind reached the amateur level in table tennis, winning 45% of their matches against 29 human players of varying skill. Against beginners, the machine won every match.

The news was announced in a scientific article. A video showing excerpts from the match was also released. Check it out below:

To develop the system, researchers created a large dataset of the ball’s initial states, including information about its position, spin and speed. The AI ​​was then trained through virtual simulations to learn various playing techniques. The robotic arm, an ABB IRB 1100 model, has 20 cameras that track the opponent’s racket.

The result is considered a major advance in this type of research, since ping-pong is considered a very important activity in the evolution and training of perception and results that robots can deliver, since it requires a wide combination of strategic skills, visual coordination and complex physical movements in real time.

The progress around this ping-pong example represents an advance in other contexts in which the robot could be used, performing tasks that involve skills in environments such as homes and warehouses.

Mapping the variables

Development will continue to refine the machine’s perception of the large set of variables that are present in a real match, such as wind speed or even the change in trajectory due to dust on the table, for example.

Google DeepMind believes that these limitations can be addressed in a number of ways. These include developing predictive AI models designed to anticipate the ball’s trajectory and also introducing even better collision detection algorithms.

Human players who participated in the test described the experience as “fun” e “engaging”regardless of the outcome of the match. Many also expressed great interest in playing against the bot again in the future.