Google DeepMind Table Tennis Robot

DeepMind's 2024 table-tennis robot, the first learning-based robot to reach amateur human level in real matches.

This work was released by Google DeepMind in August 2024, and the paper claims it's the first learning-based robot to reach amateur human level in competitive table tennis. The hardware is a 6-degree-of-freedom ABB IRB 1100 arm mounted on two linear rails that let it move forward, back, left, and right, with two 125Hz cameras tracking the ball. Control has two layers: low-level specialized skills (such as forehand and backhand), each carrying a “skill descriptor” recording what it's good and bad at; and a high-level controller that picks a skill based on the incoming ball and statistics about the opponent, adapting to that opponent in real time during a match. The skills are trained with reinforcement learning in MuJoCo simulation and deployed zero-shot to the real robot; the distribution of incoming balls observed in real matches is then fed back into simulation, forming an iterative training curriculum. Playing 29 human opponents it had never faced before, the robot won 45% of matches overall: it won every match against beginners, 55% against intermediate players, and lost every match against advanced players.

ExampleAgainst an opponent it has never played before, the high-level controller keeps tallying each low-level skill's return success rate during the match, and increasingly picks whichever skills are working against that particular opponent.

Also called

Achieving Human Level Competitive Robot Table Tennis, Robot Table Tennis

Related

Sim-to-Real Transfer · Hierarchical Architecture · Reinforcement Learning · Dynamic Manipulation · MuJoCo (Multi-Joint dynamics with Contact) · Google DeepMind

Sources

Achieving Human Level Competitive Robot Table Tennis (arXiv 2408.03906)
Competitive Robot Table Tennis 项目页 (Chinese)

As of

2024-08