- The paper introduces a dataset of 8,194 bounces across 10 racket configurations and shows that rebound parameters vary with incident normal and tangential velocities rather than remaining constant.
- The Gaussian Process model achieves the best average accuracy, reducing post-impact velocity and spin errors to 19 cm/s and 9 rad/s, with especially large gains for nonlinear long-pip rubbers.
- The Bayesian adaptation method improves velocity prediction after roughly 30 observations, reducing error by about 6 cm/s on average and up to 17 cm/s for out-of-distribution rackets.
Motivation and problem statement
Accurate models of ball–racket contact are a prerequisite for reliable trajectory prediction and stroke planning in robotic table tennis, yet the dominant racket models in the literature rely on constant-parameter linear impulse formulations and have been validated almost exclusively on inverted rubbers (2604.11349). This paper addresses both deficiencies simultaneously: it introduces a dataset of 8,194 recorded bounces across 10 racket configurations spanning inverted, anti-spin, short-pip, medium-pip, and long-pip rubbers, and it proposes a Gaussian Process (GP) framework that learns state-dependent bounce parameters while retaining the physical structure of an impulse-based contact model.
The paper's central empirical claim is that the parameters governing rebound — the Coefficient of Restitution (COR) and the tangential stiffness kp​ — are not constants but vary systematically with the incident normal velocity vz​ and tangential surface velocity vs​, and that these variations differ substantially across rubber types. This directly contradicts the assumption underlying the widely used linear bounce model in which the matrices A,B,C,D mapping pre-impact to post-impact velocity and spin are fixed per racket.
Dataset acquisition
Balls were launched by a throwing machine toward rackets rigidly mounted on an industrial robot arm, constraining motion to a plane so that a single high-speed camera at 370 fps sufficed for trajectory recording. Velocities were obtained by least-squares fitting of gravity-compensated positions over a 50 ms window, neglecting drag and Magnus effects over that interval; spin was measured with dotted balls using SpinDOE. The launcher limited incident conditions to 12 m/s and 125 rps — a constraint the author acknowledges as restricting coverage of professional-level impact states. The dataset comprises 8,194 bounce events evenly distributed across the 10 rackets, split 80/20 into training and test sets.
Rubber-specific dynamics
The analysis yields several quantitative findings with direct modeling consequences:
- Inverted rubbers: the COR decreases approximately linearly with ∣vz​∣, with differences of up to 0.15 between low (~2 m/s) and high (~12 m/s) impacts. Stiffer blades yield higher COR; offensive rubbers show steeper COR slopes (average 0.021) than allround rubbers (0.017). The tangential parameter kp​ averages around 0.019 across blades but depends on both vs​ and vz​, with the sign of the vz​ dependence differing between offensive and allround rubbers.
- Anti-spin rubber: modeled via Coulomb friction, the sliding/rolling coefficient α initially exceeds the theoretical rolling plateau of 0.4 before converging to it, which the paper attributes to the abrupt sliding-to-rolling transition assumed by the standard model. A smoothed parametric variant using a normal CDF transition fits better, yielding vz​0 versus 0.207 for the standard fit. This racket exhibits the lowest COR (vz​1).
- Long pips: the most nonlinear case. COR depends primarily on vz​2 (pips bending fully reduces energy restitution), and vz​3 saturates beyond vz​4 m/s for the sponge configuration, indicating pip deformation limits. The paper concludes explicitly that a simple linear model is insufficient for long pips.
These observations motivate reformulating the bounce as a state-dependent linear map whose coefficients are functions of vz​5, estimated non-parametrically with GPs using RBF kernels and hyperparameters tuned by marginal likelihood maximization.
Benchmark results
The evaluation compares constant-parameter, linear state-dependent, MLP, residual MLP, and GP variants of the same physical model, reporting MAE on post-impact velocity and spin. Mean results across all ten rackets:
| Model |
MAE vz​6 (cm/s) |
MAE vz​7 (rad/s) |
| Constant |
33 ± 26 |
15 ± 17 |
| Linear |
21 ± 16 |
11 ± 9 |
| MLP |
20 ± 15 |
10 ± 9 |
| Residual MLP |
25 ± 20 |
12 ± 13 |
| GP |
19 ± 15 |
9 ± 8 |
The gains from state dependence are largest precisely where the dynamics are most nonlinear: for long-pip rackets 5 and 6, the GP reduces velocity error from 48 and 43 cm/s (constant) to 19 and 20 cm/s respectively, whereas the linear model only reaches 27 and 25 cm/s. Notably, the residual MLP performs worse than the plain linear model; the author attributes this to its lack of explicit physical structure, forcing it to learn both the base behavior and the correction from limited data. This supports the paper's broader argument that physically structured parameter estimation outperforms generic function approximation in this data regime.
Online adaptation
Because an opponent's racket properties are unknown at rally start, the paper formulates adaptation as Bayesian inference: a prior over bounce parameters is built offline from training data of same-type rackets, then conditioned on observed interactions during play. To keep inference tractable in real time, the exact RBF prior is replaced by a linear prior on features vz​8, justified by the finding that inverted-rubber parameters are already well approximated by linear functions. After roughly 30 observations, the posterior reduces mean velocity prediction error by about 6 cm/s, and by up to 17 cm/s in the most out-of-distribution cases. The conservative prior variance vz​9 accounts for inter-racket variability within a rubber type. This result implies that a robot can identify an unfamiliar opponent's racket within a single rally without discarding prior knowledge — a capability neither naive system identification nor online MLP fine-tuning provides economically.
Limitations and open questions
The paper concedes three substantive restrictions. First, planar motion recording implicitly assumes isotropic tangential bounce behavior, which may not hold for full 3D impacts. Second, the launcher's velocity and spin ceilings limit extrapolation to professional-level play. Third, the impulse-based formulation assumes instantaneous contact and therefore cannot capture prolonged-contact effects relevant to serves, where extended dwell time maximizes spin transfer. An additional open question concerns whether the GP posterior's uncertainty estimates can be exploited effectively inside uncertainty-aware MPC or risk-sensitive stroke selection during live play; the paper motivates this use case but does not demonstrate it on a physical robot.
Conclusion
This work provides a systematic empirical characterization of racket-ball bounce dynamics across ten blade–rubber combinations, demonstrating quantitatively that constant-parameter contact models are inadequate — particularly for pimpled and anti-spin surfaces — and that GP-estimated state-dependent parameters yield the best predictive accuracy among the tested approaches, with mean errors of 19 cm/s in velocity and 9 rad/s in spin. The Bayesian formulation further enables few-shot online identification of unseen rackets during gameplay. The main open problems left by the paper are extension to full 3D impacts, broader velocity/spin coverage, and closed-loop integration of the learned model and its uncertainties into a real-time control stack.