Google DeepMind taught a robot to play table tennis at a human level
Posted: Thu Dec 12, 2024 4:26 am
Google DeepMind’s AI robot has reached the level of an amateur at table tennis. It won all of its matches against beginners and more than half of its matches against intermediate players, but has yet to beat a pro. Table tennis requires hand-eye coordination, strategic thinking, speed, and adaptability, which makes it possible to train these skills in machines.
The Google subsidiary published a study on Arxiv describing how the robot functions and also showing footage of it competing with humans.
According to DeepMind, the robot had to be good at low-level skills, such hungary phone number resource as passing a ball, as well as more complex tasks, such as long-term planning and strategy development. Its opponents also used different styles, and it refined and adapted its approach based on vast amounts of data.
A robotic arm with a 3D-printed racket won 13 out of 29 games against humans of varying skill levels. The robot won 100% of its matches against novices and 55% against intermediate players. However, it lost all of its matches against advanced opponents.
According to DeepMind, the results of this project are a step forward towards achieving human-level speed and performance in real-world tasks, which is what the robotics community is striving for.
The scientists used four approaches that could make the results useful for other fields:
hierarchical and modular policy architecture;
methods for transferring a model from simulation to the real world without additional tuning, including an iterative approach to determining the distribution of learning tasks tied to the real world;
adaptation to unknown opponents in real time;
research with human participation – testing the model in real matches against unknown opponents in a physical environment.
The company also added that this approach allowed for a competitive game at a human level and made it so that people enjoyed playing with the robot agent.
It's not just DeepMind that's training its systems on table tennis. The sport requires hand-eye coordination, strategic thinking, speed, and adaptability, making it a good fit for training and testing these skills in AI robots.
The world's first "table tennis robot trainer" from Japanese developer OMRON made it into the Guinness Book of World Records in 2017. The company claims that its robot "embodies the relationship that will exist between people and technology in the future."
MORE INTERESTING:
The Google subsidiary published a study on Arxiv describing how the robot functions and also showing footage of it competing with humans.
According to DeepMind, the robot had to be good at low-level skills, such hungary phone number resource as passing a ball, as well as more complex tasks, such as long-term planning and strategy development. Its opponents also used different styles, and it refined and adapted its approach based on vast amounts of data.
A robotic arm with a 3D-printed racket won 13 out of 29 games against humans of varying skill levels. The robot won 100% of its matches against novices and 55% against intermediate players. However, it lost all of its matches against advanced opponents.
According to DeepMind, the results of this project are a step forward towards achieving human-level speed and performance in real-world tasks, which is what the robotics community is striving for.
The scientists used four approaches that could make the results useful for other fields:
hierarchical and modular policy architecture;
methods for transferring a model from simulation to the real world without additional tuning, including an iterative approach to determining the distribution of learning tasks tied to the real world;
adaptation to unknown opponents in real time;
research with human participation – testing the model in real matches against unknown opponents in a physical environment.
The company also added that this approach allowed for a competitive game at a human level and made it so that people enjoyed playing with the robot agent.
It's not just DeepMind that's training its systems on table tennis. The sport requires hand-eye coordination, strategic thinking, speed, and adaptability, making it a good fit for training and testing these skills in AI robots.
The world's first "table tennis robot trainer" from Japanese developer OMRON made it into the Guinness Book of World Records in 2017. The company claims that its robot "embodies the relationship that will exist between people and technology in the future."
MORE INTERESTING: