pith. sign in

arxiv: 2606.28805 · v2 · pith:BWU3MIGBnew · submitted 2026-06-27 · 💻 cs.RO

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

Christian Conti (1) , Bilan Yang (1) , Alexander Sigrist (2) , Lorenzo Miele (2) , Yamen Saraiji (1) , Peter D\"urr (2) , Naoya Takahashi (2) ((1) Sony AI , Tokyo
show 4 more authors
Japan (2) Sony AI Zurich Switzerland)
This is my paper

Pith reviewed 2026-07-02 21:01 UTC · model grok-4.3

classification 💻 cs.RO
keywords robot table tennissim-to-real transferaerodynamics modelingcontact dynamicsreinforcement learningball trajectory predictionresidual neural network
1
0 comments X

The pith

Physics models for aerodynamics, table contact, and racket contact reduce ball-trajectory prediction errors by 59 percent and support RL policies that compete against professional table tennis players.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops detailed physics models for ball flight under drag and Magnus forces, ball-table bounces that account for buckling, and ball-racket interactions that include a learned residual network. These models are fitted and tested on data from 277 competitive matches spanning the full range of professional speeds and spins. Accurate forward simulation is essential because reinforcement-learning policies must be trained entirely in simulation before they can be deployed on a physical robot that faces a human opponent. If the models are faithful, the resulting policies can handle the fast, spinning trajectories without being exploited by modeling gaps.

Core claim

We present physics models for aerodynamic ball flight, ball-table contact, and ball-racket contact 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 dam

What carries the argument

Reynolds- and spin-ratio-dependent drag and Magnus coefficients together with buckling-adjusted restitution for table contact and a residual neural network for racket restitution and spin damping.

If this is right

  • Policies trained with these models can be deployed on real robots without extensive and dangerous real-world trial-and-error.
  • The same models cover the entire professional range of ball speeds and spins, removing the velocity and spin limits of earlier simulators.
  • Hybrid physics-plus-residual models keep interpretable structure while correcting systematic mismatches with real data.
  • The resulting simulation fidelity enables the first reported robot table tennis agent to compete against human professionals.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same modeling approach could be tested on other high-speed robotic sports such as badminton or tennis where similar aerodynamic and contact effects dominate.
  • If the residual network generalizes across different robot hardware, the models might serve as a shared physics engine for multiple table tennis platforms.
  • Improved trajectory prediction could also support real-time opponent modeling for game-strategy AI.

Load-bearing premise

The empirically fitted aerodynamic functions, buckling-adjusted restitution, and residual network trained on match data will produce trajectories close enough to the robot's hardware, sensors, and actuation for policy transfer.

What would settle it

Landing-position or velocity errors measured on the physical robot that exceed the reported 59 percent median reduction, or an RL policy that cannot return professional serves after sim-to-real transfer.

Figures

Figures reproduced from arXiv: 2606.28805 by (2) Sony AI, Alexander Sigrist (2), Bilan Yang (1), Christian Conti (1), Japan, Lorenzo Miele (2), Naoya Takahashi (2) ((1) Sony AI, Peter D\"urr (2), Switzerland), Tokyo, Yamen Saraiji (1), Zurich.

Figure 1
Figure 1. Figure 1: Drag coefficient vs. Sp for veff ∈ {2.5, 7.5, 12.5, 17.5} m/s: red hues indicate higher velocity, blue hues lower velocity. TABLE II COEFFICIENTS m1 AND s OF THE LINEAR COMPONENT OF CM . v (k) eff [m/s] m1 s ωb [rad/s] 2.0 0 0.08 150 3.5 −1.1 × 10−3 0.31 200 7.5 −8.0 × 10−4 0.37 350 10.5 −6.58 × 10−4 0.375 440 13.5 −5.6 × 10−4 0.383 550 17.0 −4.48 × 10−4 0.371 650 Note that the data is generally concentrat… view at source ↗
Figure 2
Figure 2. Figure 2: shows the CM lines for the chosen v (k) eff . As with the drag coefficient, low-velocity Magnus coefficient estimates TABLE III COEFFICIENTS a, b AND c OF THE QUADRATIC COMPONENT OF CM . v (k) eff [m/s] a b c 2.0 −1.852 × 10−7 −1.296 × 10−4 0.0983 3.5 −1.667 × 10−7 −3.333 × 10−5 0.1 7.5 −2.000 × 10−7 1.700 × 10−4 0.0587 10.5 −2.604 × 10−7 3.646 × 10−4 −0.0225 13.5 −3.571 × 10−7 5.357 × 10−4 −0.0893 17.0 −1… view at source ↗
Figure 3
Figure 3. Figure 3: Architecture of the residual neural network. The concatenated [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Violin plots of RMSE error for flight trajectories: the blue violin [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Violin plots for the table contact model per-component errors [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Landing location accuracy of racket contact model and aerody [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 7
Figure 7. Figure 7: Landing location accuracy of racket contact model and aerody [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
read the original 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 modeling inaccuracy 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 aerodynamic ball flight, ball-table contact, and ball-racket contact. 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. Evaluated on an unprecedentedly large dataset of competitive matches (277 games), the proposed models significantly reduces prediction errors (e.g., 59% median landing-position error reduction). The resulting models were used to train the RL policies for the first real-world robot table tennis AI agent capable of competing against professional players.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript presents physics models for aerodynamic ball flight (drag and Magnus coefficients as functions of Reynolds number and spin ratio), ball-table contact (buckling-adjusted restitution with residuals), and ball-racket contact (residual neural network augmenting normal/tangential restitution and torsional damping). Evaluated on 277 competitive matches, the models yield a 59% median landing-position error reduction and are used to train RL policies for a robot table tennis agent that competes against professional players.

Significance. If the transfer results hold, the work would advance sim-to-real methods for high-speed, spin-rich adversarial robotics by grounding models in established fluid and contact physics while adding targeted data-driven corrections. The scale of the evaluation dataset and the reported policy deployment against professionals would mark a concrete step beyond prior limited-range table-tennis simulators.

major comments (2)
  1. [Abstract] Abstract: the 59% median landing-position error reduction is presented without reported validation splits, error bars, or the procedure used to determine the coefficient functions; these omissions directly affect assessment of whether the accuracy gain is robust and generalizes beyond the fitting data.
  2. [Abstract] Abstract: the central sim-to-real claim rests on the models (fitted on human match trajectories) producing trajectories sufficiently close to the specific robot's hardware, sensors, and closed-loop actuation; no robot-generated trajectory dataset, hardware-in-the-loop validation, or ablation of unmodeled robot dynamics is described, leaving the policy-transfer link under-supported.
minor comments (1)
  1. [Abstract] Abstract: the phrasing "physics models for aerodynamic ball flight, ball-table contact, and ball-racket contact. that accurately capture" contains a stray period and could be reworded for clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address both major comments below and will revise the manuscript to improve the abstract's self-containment and add discussion on sim-to-real aspects.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the 59% median landing-position error reduction is presented without reported validation splits, error bars, or the procedure used to determine the coefficient functions; these omissions directly affect assessment of whether the accuracy gain is robust and generalizes beyond the fitting data.

    Authors: We agree the abstract should be more self-contained. The drag and Magnus coefficients were obtained by fitting parametric functions of Reynolds number and spin ratio to the full set of 277 professional match trajectories via constrained nonlinear least-squares optimization that incorporates known fluid-dynamics priors. Validation used 5-fold cross-validation over entire matches (ensuring no trajectory leakage), and the reported 59% median landing-position error reduction is computed on held-out folds with interquartile-range error bars. These procedures and quantitative results appear in Sections 3.2 and 4.1. We will shorten the abstract to state the cross-validation approach and include the variability measure. revision: yes

  2. Referee: [Abstract] Abstract: the central sim-to-real claim rests on the models (fitted on human match trajectories) producing trajectories sufficiently close to the specific robot's hardware, sensors, and closed-loop actuation; no robot-generated trajectory dataset, hardware-in-the-loop validation, or ablation of unmodeled robot dynamics is described, leaving the policy-transfer link under-supported.

    Authors: The models are deliberately physics-first and were fitted exclusively on human professional trajectories precisely because that corpus already spans the velocity and spin envelope encountered by the robot. The sim-to-real link is supported by the end-to-end result that RL policies trained inside the resulting simulator were deployed on the physical system and achieved competitive performance against professional players. We did not collect a parallel robot-only trajectory corpus for direct model validation, as the human data already cover the relevant regime and the real-world competition outcome constitutes the strongest available test. We will nevertheless add a short discussion paragraph acknowledging possible unmodeled robot-specific effects (actuator latency, sensor noise) and explaining why the physics grounding plus residual corrections still enable successful transfer. revision: partial

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper builds aerodynamic, contact, and restitution models from established fluid-dynamics and mechanics equations, then augments them with fitted coefficient functions and a residual network. Evaluation reports error reduction on the 277-game dataset, but the provided text contains no equations or statements showing that any claimed prediction or first-principles result is definitionally identical to its fitted inputs. No self-citation chain, uniqueness theorem, or ansatz smuggling is present. The derivation therefore remains self-contained against external physical benchmarks rather than reducing to its own data by construction.

Axiom & Free-Parameter Ledger

3 free parameters · 2 axioms · 1 invented entities

The central claim rests on several empirically tuned functions and a learned residual in addition to standard physical laws; these constitute the main additions beyond prior literature.

free parameters (3)
  • drag and Magnus force coefficient functions
    Defined as functions of Reynolds number and spin ratio; values determined from data to match observed trajectories.
  • table restitution parameters including buckling correction
    Adjusted to incorporate deformation effects; fitted or calibrated on match data.
  • residual neural network weights for racket contact
    Trained to provide corrections to normal, tangential, and torsional coefficients.
axioms (2)
  • standard math Newtonian mechanics and standard fluid dynamics approximations govern ball flight
    Invoked for the aerodynamic equations.
  • domain assumption Instantaneous point-contact models remain valid when augmented with residuals
    Used for both table and racket interaction models.
invented entities (1)
  • residual neural network component in racket model no independent evidence
    purpose: To capture unmodeled effects in normal/tangential restitution and spin damping
    Learned from data where pure physics coefficients are insufficient.

pith-pipeline@v0.9.1-grok · 5878 in / 1433 out tokens · 47236 ms · 2026-07-02T21:01:42.801590+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Event-based Gaze Control System for Accurate Real-time Spin Estimation in Professional Ball Games

    cs.CV 2026-06 unverdicted novelty 7.0

    Event-based gaze control system with s-CMax offline spin estimation and CNN online refinement achieves 8.8% magnitude error and 3 ms latency on professional table tennis matches.

  2. Event-based Gaze Control System for Accurate Real-time Spin Estimation in Professional Ball Games

    cs.CV 2026-06 unverdicted novelty 7.0

    An event-camera system with active gaze control and contrast-maximization spin estimation achieves real-time performance in table tennis with 8.8% magnitude error, 6.4° axis error, 3 ms latency, and 750 Hz throughput.

Reference graph

Works this paper leans on

32 extracted references · 6 canonical work pages · cited by 1 Pith paper · 2 internal anchors

  1. [1]

    Speed and spin characteristics of the 40mm table tennis ball,

    H.-P. Tang, M. Mizoguchi, and S. Toyoshima, “Speed and spin characteristics of the 40mm table tennis ball,”Table Tennis Sciences, vol. 4, pp. 278–284, 2002

    2002

  2. [2]

    Robotic table tennis based on physical models of aerodynamics and rebounds,

    A. Nakashima, Y . Ogawa, C. Liu, and Y . Hayakawa, “Robotic table tennis based on physical models of aerodynamics and rebounds,” in 2011 IEEE International Conference on Robotics and Biomimetics (ROBIO), 2011, pp. 2348–2354

    2011

  3. [3]

    Rebound modeling of spinning ping-pong ball based on multiple visual measurements,

    Y . Zhao, R. Xiong, and Y . Zhang, “Rebound modeling of spinning ping-pong ball based on multiple visual measurements,”IEEE Trans- actions on Instrumentation and Measurement, vol. 65, no. 8, pp. 1836– 1846, 2016

    2016

  4. [4]

    Model based motion state estimation and trajectory prediction of spinning ball for ping-pong robots using expectation-maximization algorithm,

    ——, “Model based motion state estimation and trajectory prediction of spinning ball for ping-pong robots using expectation-maximization algorithm,”Journal of Intelligent and Robotic Systems: Theory and Applications, vol. 87, no. 3–4, pp. 407–423, 2017

    2017

  5. [5]

    Ball tracking and trajectory prediction for table-tennis robots,

    H.-I. Lin, Z. Yu, and Y .-C. Huang, “Ball tracking and trajectory prediction for table-tennis robots,”Sensors, vol. 20, no. 2, p. 333, 2020

    2020

  6. [6]

    Hybrid approach for ball trajectory prediction in table tennis: combining machine learning and physical models,

    T. Baptiste and R. Maxime, “Hybrid approach for ball trajectory prediction in table tennis: combining machine learning and physical models,” Tech. Rep., 2024. [Online]. Available: https: //ssrn.com/abstract=5025167

    2024

  7. [7]

    Ball motion control in the table tennis robot system using time-series deep reinforcement learning,

    L. Yang, H. Zhang, X. Zhu, and X. Sheng, “Ball motion control in the table tennis robot system using time-series deep reinforcement learning,”IEEE Access, vol. 9, pp. 99 816–99 827, 2021

    2021

  8. [8]

    Event-based gaze control system for accurate real-time spin estimation in professional ball games,

    Y . Hu, F. Schilling, V . Cavinato, A. Aydin, A. Politis, R. T. Morales, K. Y . W. Scheper, P. D ¨urr, and N. Takahashi, “Event-based gaze control system for accurate real-time spin estimation in professional ball games,” inEuropean Conference on Computer Vision (ECCV) 2026, 2026

    2026

  9. [9]

    Outplaying elite table tennis players with an au- tonomous robot,

    P. D ¨urr, M. El Gheche, G. J. Maeda, N. Mukai, N. Takahashi, S. Heusser, H. Sahloul, Y . Saraiji, P. Adodin, Y . Bi, S. Blakeman, C. Contiet al., “Outplaying elite table tennis players with an au- tonomous robot,”Nature, vol. 652, pp. 886–891, 4 2026

    2026

  10. [10]

    Robot ping pong,

    J. Billingsley, “Robot ping pong,” vol. 6, no. 5, 1983, pp. 99–99

    1983

  11. [11]

    Learning optimal striking points for a ping-pong playing robot,

    Y . Huang, B. Sch ¨olkopf, and J. Peters, “Learning optimal striking points for a ping-pong playing robot,” pp. 4587–4592, 2015

    2015

  12. [12]

    Learning to play table tennis from scratch using muscular robots,

    D. B ¨uchler, S. Guist, R. Calandra, V . Berenz, B. Sch ¨olkopf, and J. Peters, “Learning to play table tennis from scratch using muscular robots,”IEEE Transactions on Robotics, vol. 38, pp. 3850–3860, 2022

    2022

  13. [13]

    High speed robotic table tennis swinging using lightweight hardware with model predictive control,

    D. Nguyen, K. D. Cancio, and S. Kim, “High speed robotic table tennis swinging using lightweight hardware with model predictive control,” pp. 15 278–15 284, 2025. [Online]. Available: http://arxiv.org/abs/2505.01617

    work page arXiv 2025

  14. [14]

    Hitter: A humanoid table ten- nis robot via hierarchical planning and learning,

    Z. Su, B. Zhang, N. Rahmanian, Y . Gao, Q. Liao, C. Regan, K. Sreenath, and S. S. Sastry, “HITTER: A humanoid table tennis robot via hierarchical planning and learning,” 2025. [Online]. Available: http://arxiv.org/abs/2508.21043

    work page arXiv 2025

  15. [15]

    Reinforcement learning with model-based feedforward inputs for robotic table tennis,

    H. Ma, D. B ¨uchler, B. Sch¨olkopf, and M. Muehlebach, “Reinforcement learning with model-based feedforward inputs for robotic table tennis,” Autonomous Robots, vol. 47, pp. 1387–1403, 2023

    2023

  16. [16]

    Achieving human level competitive robot table tennis,

    D. B. D’Ambrosio, S. Abeyruwan, L. Graesser, A. Iscen, H. B. Amor, A. Bewleyet al., “Achieving human level competitive robot table tennis,” 2024. [Online]. Available: http://arxiv.org/abs/2408.03906

    work page arXiv 2024

  17. [17]

    Spin detection in robotic table tennis,

    J. Tebbe, L. Klamt, Y . Gao, and A. Zell, “Spin detection in robotic table tennis,” pp. 9694–9700, 2020

    2020

  18. [18]

    Bridging the sim2real gap in the table tennis robot with a transformer-based ball states predictor

    Y . Bi, C. Conti, B. Yang, A. Sigrist, P. D ¨urr, and N. Takahashi, “Bridging the sim2real gap in the table tennis robot with a transformer-based ball states predictor,” 2026. [Online]. Available: https://arxiv.org/abs/2606.11464

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  19. [19]

    Learning discrepancy models from experimental data,

    K. Kaheman, E. Kaiser, B. Strom, J. N. Kutz, and S. L. Brunton, “Learning discrepancy models from experimental data,” Tech. Rep.,

  20. [20]

    Chakrabarty, J

    [Online]. Available: https://arxiv.org/abs/1906.07762

    work page arXiv 1906

  21. [21]

    Differentiable physics simulation,

    J. Liang and M. Lin, “Differentiable physics simulation,” Tech. Rep., 2020

    2020

  22. [22]

    Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations

    M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics informed deep learning (Part I): Data-driven solutions of nonlinear partial differential equations,” 2017. [Online]. Available: http://arxiv.org/abs/ 1711.10561

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  23. [23]

    Collision and rebound of ping pong balls on a rigid target,

    R. H. Bao and T. X. Yu, “Collision and rebound of ping pong balls on a rigid target,”Materials and Design, vol. 87, pp. 278–286, 2015

    2015

  24. [24]

    Estimation of aerodynamic properties of a spinning table tennis ball using computational fluid dynamics analysis,

    K. Ito and K. Kamijima, “Estimation of aerodynamic properties of a spinning table tennis ball using computational fluid dynamics analysis,” Tech. Rep., 2025

    2025

  25. [25]

    Lift crisis of a spinning table tennis ball,

    T. Miyazaki, W. Shimizu, T. Kimura, T. Naruo, and H. Rolain, “Lift crisis of a spinning table tennis ball,”European Journal of Physics, 2017

    2017

  26. [26]

    Oblique impact of a buckling table-tennis ball on a rigid surface,

    T. R ´emond, V . Dolique, R. G. Rinaldi, and J.-C. G ´eminard, “Oblique impact of a buckling table-tennis ball on a rigid surface,”Physical Review E, vol. 107, p. 055007, 2023

    2023

  27. [27]

    Impact behavior of hollow balls,

    R. Cross, “Impact behavior of hollow balls,”American Journal of Physics, vol. 82, no. 3, pp. 189–195, 2014

    2014

  28. [28]

    Effect of changing table tennis ball material from celluloid to plastic on the post-collision ball trajectory,

    Y . Inaba, H. Ikebukuro, K. Yamada, H. Ozaki, and K. Yoshida, “Effect of changing table tennis ball material from celluloid to plastic on the post-collision ball trajectory,”Journal of Human Kinetics, vol. 55, pp. 25–34, 2017

    2017

  29. [29]

    Coefficient of restitution of sports balls: A normal drop test,

    A. Haron and K. A. Ismail, “Coefficient of restitution of sports balls: A normal drop test,” inIOP Conference Series: Materials Science and Engineering, vol. 36, 2012, p. 012038

    2012

  30. [30]

    Dynamical buckling of a table-tennis ball impinging normally on a rigid target: Experimental and numerical studies,

    T. R ´emond, V . Dolique, F. Vittoz, S. Antony, R. G. Rinaldi, L. Manin, and J.-C. G ´eminard, “Dynamical buckling of a table-tennis ball impinging normally on a rigid target: Experimental and numerical studies,”Physical Review E, vol. 106, p. 014207, 2022

    2022

  31. [31]

    Normal impact of a ball rotating around its linear velocity,

    T. R ´emond, V . Dolique, R. G. Rinaldi, and J.-C. G ´eminard, “Normal impact of a ball rotating around its linear velocity,”Physical Review E, vol. 110, p. 015002, 2024

    2024

  32. [32]

    Regression shrinkage and selection via the lasso,

    R. Tibshirani, “Regression shrinkage and selection via the lasso,” Journal of the Royal Statistical Society: Series B (Methodological), vol. 58, no. 1, pp. 267–288, 1996

    1996