Google DeepMind's Table Tennis Robot Competes at Human-Level

Researchers from Google DeepMind have developed a robot capable of playing table tennis at an amateur human level. This robotic system, featuring a 6 DoF ABB 1100 arm mounted on linear gantries, has been tested against human players of varying skill levels, winning 45% of the matches overall. 

The robot's design utilises a hierarchical and modular policy architecture, which includes low-level controllers for specific skills and a high-level controller for decision-making based on match statistics.

The robot employs advanced techniques to bridge the simulation-to-real-world gap, enabling it to adapt to unseen opponents and improve its decision-making process. It uses a combination of reinforcement learning and imitation learning to train its skills in simulation before deploying them in real-world matches. The system's adaptability and strategic capabilities are enhanced by real-time tracking of match statistics and opponent performance.

In a user study involving 29 participants, the robot demonstrated solid amateur-level performance, winning all matches against beginners and 55% against intermediate players. However, it struggled against advanced players. Participants found the experience engaging and enjoyable, with 26 out of 29 expressing interest in playing with the robot again.

Despite its success, the robot faces challenges in handling fast and low balls, as well as accurately detecting spin. Future research aims to address these limitations by improving control algorithms and hardware optimisations, enhancing collision detection, and refining the robot's strategic capabilities. This development marks a significant step towards achieving human-level performance in robotics, with potential applications beyond table tennis in various real-world tasks.

Google Deepmind’s robotics wing has constantly been advancing its research, as a few months back, they introduced ALOHA 2, a robotics technology with more dexterity to tasks with low-cost robots and AI. This development is likely an extension of the same research using ALOHA 2.