Robot learning has seen remarkable advancements in recent years; however, achieving human-level performance in terms of accuracy, speed, and adaptability remains a significant challenge across various domains. One such domain is table tennis—a sport that demands years of rigorous training for human players to reach an advanced level of proficiency.
In a new paper Achieving Human Level Competitive Robot Table Tennis, a Google DeepMind research team introduces the first robot agent that attains amateur human-level performance in competitive table tennis.
The team employed a hierarchical and modular policy architecture, which includes multiple low-level skill policies managed by a high-level controller. Each low-level skill policy is tailored to a specific aspect of table tennis, such as executing a forehand topspin, targeting with a backhand, or performing a forehand serve. These skills are built upon a shared foundational policy, allowing for further specialization as each skill improves.
The robot table tennis agent operates with a two-tiered control system, comprising a high-level controller (HLC) and several low-level controllers (LLCs). The LLCs are responsible for executing different table tennis skills by generating joint velocity commands at a frequency of 50Hz. The HLC’s role is to determine which LLC should be activated during each ball episode. Within the HLC, six components work together to decide the optimal LLC for the situation.
Beyond developing the policy itself, the researchers also collect and store data both offline and online about each low-level skill’s strengths, weaknesses, and limitations. These skill descriptors provide the robot with crucial insights into its capabilities and areas for improvement. The HLC uses this information, along with current game statistics and the opponent’s profile, to select the most appropriate skill at any given moment.
In their empirical study, the researchers evaluated the robot’s performance through 29 matches against human opponents. The robot won 45% of the matches (13 out of 29), including 100% of the games against beginners and 55% against intermediate players, showcasing a solid amateur human-level performance.
The paper Achieving Human Level Competitive Robot Table Tennis is on arXiv.
Author: Hecate He | Editor: Chain Zhang