Sports are a great test for robots. For example, the annual RoboCup football tournament has been held since the mid-1990s. And table tennis has been used for comparative testing of robotic arms since the 1980s. The robot requires high mobility, quick reaction and strategic thinking. The robot presented by Google DeepMind showed the level of an average amateur, but could not cope with professionals.
Image source: Google DeepMind
In a recently published paper titled “Achieving Human Level Competitive Table Tennis Robot,” Google’s DeepMind Robotics team unveiled a robot “capable of human-level competitive sports and represents a milestone in robot learning and control.”
During testing, the robot was able to defeat all entry-level players it faced. With average players, the robot won 55% of matches. However, the robot lost all the fights to the professionals. Overall, the system won 45% of the 29 matches played. The biggest drawback of the system is the delayed reaction to fast balls. The robot also has difficulty playing the backhand, receiving high and low balls, and assessing the spin of the ball.
«To address the latency limitations that hinder the robot’s reaction time on fastballs, we propose to explore advanced control algorithms and hardware optimizations, the developers believe. “This could include exploring predictive models to predict ball trajectories or implementing faster communication protocols between the robot’s sensors and actuators.”
DeepMind is confident in the promise of its development: “This is just a small step towards a long-standing goal in robotics – achieving human-level performance in many useful real-world skills. There is still much work to be done to consistently achieve human-level performance on individual tasks and then beyond, creating versatile robots that can perform a variety of useful tasks while interacting skillfully and safely with people in the real world.”