Over the past four decades, humans have maintained an advantage over the table tennis robot. However, recent studies by Google DeepMind may be changing the landscape. According to a preliminary paper published Aug. 7, researchers have developed a robotic system that has achieved a level of performance comparable to that of an “amateur ping pong player.”
Google DeepMind studies
Traditionally, games such as chess and Go have been the games of choice for evaluating the strategic capabilities of artificial intelligence. However, table tennis has become a standard in the robotics industry due to its unique combination of real-time strategy and physicality. The sport requires rapid adaptation to dynamic variables, complex movements and visual coordination, which has led engineers to engage in countless rounds of ping pong with machines for more than 40 years.
Meet our AI-powered robot that’s ready to play table tennis. 🤖🏓
– Google DeepMind (@GoogleDeepMind) August 8, 2024
It’s the first agent to achieve amateur human level performance in this sport. Here’s how it works. 🧵 pic.twitter.com/AxwbRQwYiB
In an announcement on the X platform, Google DeepMind explained that the robot had to master basic skills, such as ball reception, as well as more advanced skills, including strategizing and long-term planning to achieve a goal. To achieve this, engineers began by collecting an extensive dataset of “initial states of table tennis balls,” which contained detailed information about the position, spin and velocity of the balls.
With this data, the artificial intelligence practiced in highly accurate virtual simulations, learning techniques such as returning serves, aiming the backhand and applyingtopspin on the forehand. They then integrated the AI with a robotic arm, capable of executing complex and fast movements, and pitted it against human players.
Machine learning of the table tennis robot
This training process did not end there. The visual information captured by the robot’s cameras was analyzed again in simulations, creating a continuous cycle of feedback and machine learning. This cycle allowed the system to adapt and constantly improve its performance.
Finally, it was time to put the robot to the test in a tournament. The project’s researchers called 29 human players, classified into four skill levels: beginner, intermediate, advanced and “advanced+”. The robotic arm faced each of them, winning 13 of the matches, representing 45% of the time. According to the researchers, these results demonstrate that the robot has achieved solid performance at the amateur level.
Some highlights, among the human level of competition in table tennis with robots. Source: Atil Iscen
Those passionate about table tennis who fear losing their superiority to robots can rest easy(at least for now). Although the machine beat all beginner-level players, its win rate dropped to 55% against intermediate players, and it failed to win against advanced players. Participants in the study described the experience as “fun” and “entertaining,” regardless of the outcome, and showed a strong interest in having rematches with the robot.
Follow us on social networks and don’t miss any of our publications!
YouTube LinkedIn Facebook Instagram X
Source and photo: Google DeepMind
Loading...
