Physics Models for Sim-to-Real Transfer in Professional-Level Robot Table Tennis

2026-06-27

AI Summary

Existing robotic table tennis systems struggle to achieve realistic simulation and effective policy transfer under high-speed, high-spin conditions, limiting their ability to compete against professional human players. This work proposes a high-fidelity physical model that, for the first time, incorporates Reynolds number– and spin ratio–dependent drag and Magnus force coefficients, as well as the effect of ball buckling on the coefficient of restitution. Furthermore, a residual network is employed to model racket-ball contact dynamics, significantly enhancing simulation realism. Leveraging this model, reinforcement learning policies trained entirely in simulation are successfully transferred to the real world, yielding the first robotic table tennis agent capable of competing against professional-level human players in physical environments. The results demonstrate the method’s effectiveness and strong sim-to-real transfer capability in complex, dynamic scenarios.

Abstract

At competitive speeds and spins, a table tennis ball follows complex, counterintuitive trajectories that a robot must track and precisely counter within fractions of a second. Training a reinforcement learning policy capable of these skills is prohibitively expensive and dangerous in the real world, making high-fidelity simulation essential. Transferability of such policies, however, critically depends on how faithfully the simulation captures real-world dynamics--a requirement made even more stringent by the adversarial nature of the game, where any regime in which a model fails to approximate reality becomes an exploitable weakness for the opponent. Prior state-of-the-art in robot table tennis generally focuses on a limited range of velocities and spins and fails to capture the richness of ball behaviors encountered in professional-level play. In this work, we present physics models for the aerodynamic ball flight, for the contact dynamics between the ball and the table, as well as between the ball and the racket that accurately capture the ball behavior over a vast range of speeds and spins relevant to the game. Specifically, we model drag and Magnus force coefficients as functions of Reynolds number and spin ratio in the aerodynamics equations. For the table contact model we model effects of ball buckling on the coefficient of restitution and incorporate residuals into the instantaneous point-contact models. For the racket contact model we introduce a residual neural network component to complement coefficients related to normal and tangential coefficients of restitution as well as torsional spin damping. The resulting models were used for the first real-world robot table tennis AI agent capable of competing against professional players, to train reinforcement learning policies.

Problem

• sim-to-real transfer
• robot table tennis
• aerodynamic modeling
• contact dynamics
• professional-level play

Innovation

• sim-to-real transfer
• aerodynamic modeling
• contact dynamics
• Magnus effect
• residual neural network