Ace was created by a team at Sony AI, and the results of the experiment were described in the prestigious journal Nature<\/em>. In the study, the robot did not face amateurs. 5 excellent players stood against it, including competitors with more than 10 years of training experience and 2 professionals. They played 5 matches. The result? The machine won 3 of them.<\/strong><\/p>\n\n\n\n It might seem that this is nothing new in table tennis. Nothing could be further from the truth. This marks another step in the development of AI in sport. In practice, table tennis is an exceptionally brutal test for robotics. At a high level, the ball can travel at more than 20 metres per second. The time between shots often falls below half a second. Then there is spin, which radically changes the ball\u2019s flight path and bounce. A robot cannot simply swing a paddle. AI has to predict, and the machine has to react almost instantly. Until now, this posed a huge problem in robotics. Artificial intelligence has overcome those difficulties.<\/strong><\/p>\n\n\n\n This research breakthrough shows the potential of physical AI agents to perform interactive real-time tasks and represents an important step toward creating robots with broader applications in fast, precise and interactive activities with humans,<\/p>\n<\/blockquote>\n\n\n\n said Peter D\u00fcrr, director of Sony AI in Zurich and head of the Ace project, quoted by ScienceAlert<\/em>.<\/p>\n\n\n\n Table-tennis-playing robots are not new. Similar devices have been built for decades, and researchers note that the first attempts date back to the 1980s. The problem is that many earlier machines worked under simplified conditions, with a limited playing area. Sony AI\u2019s Ace is the first to cope in a real robot-versus-human contest.<\/strong><\/p>\n\n\n\n The secret of this achievement lies in its design. Ace rests on 3 main pillars: perception, learning and hardware. The robot uses a camera system that allows it to locate the ball in space and estimate its spin. It makes decisions using deep reinforcement learning methods, previously trained in simulations<\/a>. Supporting all this is an 8-axis robotic arm.<\/p>\n\n\n\nMore than a technological demonstration<\/h3>\n\n\n\n
\n
One of the hardest tests for AI<\/h2>\n\n\n\n
How a robot that can beat a human works<\/h3>\n\n\n\n
Robot versus human: it does not always win<\/h2>\n\n\n\n