pith. sign in

arxiv: 2606.26643 · v1 · pith:VUUQOM5Unew · submitted 2026-06-25 · 💻 cs.RO

Hardware Design for Table Tennis Robot Capable of Beating Professional Players

Nobuhiko Mukai , Pavel Adodin , Stefan Heusser , Alexander Sigrist , Divij Grover , Guillem Torrente , Farshad Khadivar , Takekazu Kakinuma
show 1 more author
Peter D\"urr
This is my paper

Pith reviewed 2026-06-26 05:22 UTC · model grok-4.3

classification 💻 cs.RO
keywords table tennis robothardware designmotion analysistopology optimizationreinforcement learningrobotic armdynamics modeling
0
0 comments X

The pith

An 8-DoF robot built from elite player motion specs defeats multiple 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 derives concrete hardware targets for workspace, payload, force resistance, speed, serve, and accuracy by studying the motions of professional table tennis players. It then describes construction of an 8-degree-of-freedom robot named Ace whose structure was lightened through topology optimization and whose motors were sized with an inverse-dynamics model. Low-order joint dynamics with delay compensation were identified to support a reinforcement-learning control policy. Experiments showed repeated swings at 0.8-second cycle time and 22 m/s racket speed. The robot won matches against several professional opponents.

Core claim

The paper claims that hardware specifications taken directly from elite human motion analysis, when realized in an 8-DoF arm with optimized mechanics and dynamics-aware reinforcement learning control, produce a robot capable of defeating professional table tennis players.

What carries the argument

The 8-DoF robot Ace, whose mechanical links were refined by topology optimization, motors selected via inverse-dynamics torque modeling, and joints modeled with low-order dynamics plus delay compensation to enable reinforcement learning control.

If this is right

  • Robots can now reach the physical performance envelope of professional table tennis players.
  • The same motion-analysis-to-specs workflow can be repeated for other fast-action sports.
  • Reinforcement learning becomes practical for high-speed manipulation once per-joint dynamics models with delay compensation are available.
  • Hardware performance limits, rather than control algorithms, had been the main barrier to competitive robotic table tennis.

Where Pith is reading between the lines

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

  • Earlier table tennis robots were likely limited by hardware that fell short of professional motion demands.
  • The same robot could function as a consistent training partner for human players seeking to improve.
  • Extending the method to other racket sports requires fresh motion capture for each sport's distinct movement patterns.

Load-bearing premise

The target specifications taken from analysis of elite players' motions are sufficient to produce a robot that wins real matches against professionals.

What would settle it

A sequence of matches in which the robot loses to the same professional players or fails to maintain the reported 0.8 s cycle time and 22 m/s peak velocity.

Figures

Figures reproduced from arXiv: 2606.26643 by Alexander Sigrist, Divij Grover, Farshad Khadivar, Guillem Torrente, Nobuhiko Mukai, Pavel Adodin, Peter D\"urr, Stefan Heusser, Takekazu Kakinuma.

Figure 1
Figure 1. Figure 1: Matches against professional table tennis players. In the spring of 2026, Ace competed against 9 professional players and defeated 8 players, including Miu Hirano (Olympic silver medalist at Tokyo 2020 and Paris 2024, highest world ranking of No.5) (see the supplementary video) [6]. Y X J1 J2 J3 J4 J5 J6 Cup Pedestal Cantilever Arm Link 1 Link 2 Link 3 Link 4 XY-stage Racket holder [PITH_FULL_IMAGE:figure… view at source ↗
Figure 2
Figure 2. Figure 2: Table tennis robot Ace. It is an 8-DoF robot that mounts a 6-DoF arm on an XY-stage. It has an end effector equipped with a racket and a cup to hold the ball, facilitating one-handed serves in conjunction with an external ball feeder (see the supplementary video for the movement of each axis and the one-handed serve). II. SETTING TARGET SPECIFICATIONS We recorded two advanced table tennis players with uni￾… view at source ↗
Figure 3
Figure 3. Figure 3: Determined workspace based on measured racket trajectories. Travel range of arm rotation axis, arm length, and the workspace were determined based on the movement of the racket centers for various strokes during serves and rallies by advanced players. Wind velocity 20 [m/s] 5.8 [N] Velocity[m/s] Racket Force 20.8 0.0 5.2 10.4 15.6 [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Airflow simulation of a racket. The force acting on the end-effector holding a racket was obtained as 5.8 N at a wind velocity of 20 m/s. to insufficient torque, depending on the racket angle. Our simulation estimated external force of approximately 5.8 N on the racket with 20 m/s wind velocity ( [PITH_FULL_IMAGE:figures/full_fig_p002_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: YR-stage configuration of Ace. By replacing the XY-stage with a YR-stage, the game can be played even in a small laboratory that does not meet ITTF regulations. robot arm able to be positioned above the tabletop. This allowed the racket to reach the net without the need to further increase the length of moving links, thus keeping the robot as lightweight as possible. The arm is moved by a 2-axis linear sta… view at source ↗
Figure 6
Figure 6. Figure 6: Topology-optimization design progression for the main structural components. The columns show the initial design, the optimization setup with design space shown in red, blue, and yellow and non-design space shown in gray, the raw topology-optimization result, and the final AM-ready design. TABLE II MASS REDUCTION SUMMARY FOR TOPOLOGY-OPTIMIZED COMPONENTS Component Baseline (kg) Optimized (kg) Reduction (%)… view at source ↗
Figure 8
Figure 8. Figure 8: Histogram of J2 motor and J2 gearbox torque margins over the course of a match with an elite player before the motor upgrade. Y axes are logarithmic, torque margin data is sampled at 1 kHz. Top-left: Cumulative histogram of motor margins from the measured motor torque. Top-right: Histogram of motor torque margin errors. Bottom-left: Cumulative histogram of predicted motor torque margins from the inverse dy… view at source ↗
Figure 9
Figure 9. Figure 9: Histogram of J2 motor and J2 gearbox torque margins over the course of a match with an elite player after the motor upgrade. Y axes are logarithmic, torque margin data is sampled at 1 kHz. Top-left: Cumulative histogram of motor margins from the measured motor torque. Top-right: Histogram of motor torque margin errors. Bottom-left: Cumulative histogram of predicted motor torque margins from the inverse dyn… view at source ↗
Figure 10
Figure 10. Figure 10: Hardware configuration of the control system. The Control PC generates target position trajectories, drives actuators by sending commands to the servo amplifiers acting as EtherCAT slaves via an EtherCAT master board, synchronizes the camera system, and controls the ball feeder and signage tablet. The GCS PC measures ball spin with the gaze control system. The Perception PC measures ball position with the… view at source ↗
Figure 11
Figure 11. Figure 11: Full-stroke swing experiment at 0.8 s intervals. (a) Linear velocity and tracking error of the center of the racket during four round-trip swings. (b) 3D visualization of the target and actual trajectories of the center of the racket and the posture of the arm at maximum velocity. actuators by transmitting target positions to each Yaskawa Sigma-X servo amplifier [16] at 1 ms intervals using CSP (Cyclic Sy… view at source ↗
Figure 12
Figure 12. Figure 12: and [PITH_FULL_IMAGE:figures/full_fig_p007_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Tracking error statistics per joint. For each joint, two side-by￾side violin–box plots compare the tracking error (act − ref, red) and the model residual (act − sim, green) over the full trajectory. Violin shapes indicate the empirical distribution; box plots show the interquartile range and median. Linear axes (X, Y) and rotary joints (J1–J6) are shown in separate panels. models with sufficient accuracy … view at source ↗
read the original abstract

This paper focuses on the hardware specifications required for a table tennis robot to beat professional players. After analyzing the motions of elite players, we defined target specifications for the workspace, payload, external-force resistance, physical performance, serve capability, and end-effector accuracy. Based on these specifications, we developed "Ace", a custom 8-DoF robot. The mechanical structure was improved through topology optimization to minimize mass while preserving stiffness. Motor and gearbox selection was optimized using an inverse-dynamics torque model. Low-order per-joint dynamics models with delay compensation were identified and integrated into simulation to enable the use of an RL control policy. Experiments demonstrated repeated full-stroke swings with a cycle time of 0.8 s and a peak racket-center velocity of 22 m/s. The robot successfully defeated multiple 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 / 0 minor

Summary. The paper derives target hardware specifications for a table tennis robot from analysis of elite players' motions (workspace, payload, external-force resistance, physical performance, serve capability, end-effector accuracy), presents the design of the resulting 8-DoF robot 'Ace' with topology-optimized structure, inverse-dynamics motor selection, and low-order per-joint dynamics models with delay compensation for RL control, and reports experimental results of 0.8 s full-stroke cycles at 22 m/s racket velocity together with the claim that the robot defeated multiple professional players.

Significance. If the defeat claim is substantiated with match data, the work would establish concrete, experimentally validated hardware benchmarks for professional-level table-tennis performance and demonstrate a viable path from motion analysis through optimized mechanics and identified dynamics to RL-based control in a high-speed interactive task.

major comments (2)
  1. [Abstract / Experiments] Abstract and experimental results: the central claim that the robot 'successfully defeated multiple professional players' is unsupported by any quantitative match data (win rates, rally counts, opponent rankings or ratings, rules followed, or sensing/control performance during actual play), rendering the leap from isolated swing metrics to interactive match outcomes unverifiable.
  2. [Specification derivation] Target-specification derivation (early sections): the claim that the listed workspace, payload, force-resistance, and accuracy targets are sufficient to beat professionals rests on the untested assumption that matching elite motion statistics produces winning hardware; no sensitivity analysis or comparison against lower-spec baselines is provided to support this sufficiency.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the two major comments point by point below, indicating the revisions that will be made to the manuscript.

read point-by-point responses

  1. Referee: [Abstract / Experiments] Abstract and experimental results: the central claim that the robot 'successfully defeated multiple professional players' is unsupported by any quantitative match data (win rates, rally counts, opponent rankings or ratings, rules followed, or sensing/control performance during actual play), rendering the leap from isolated swing metrics to interactive match outcomes unverifiable.

    Authors: We agree that the claim is unsupported by quantitative match data. The experimental results section reports only isolated swing metrics (cycle time and velocity), with no recorded match statistics or opponent details. We will revise the abstract, introduction, and conclusion to remove the claim of defeating professional players. The revised text will focus exclusively on the achieved hardware performance benchmarks. revision: yes

  2. Referee: [Specification derivation] Target-specification derivation (early sections): the claim that the listed workspace, payload, force-resistance, and accuracy targets are sufficient to beat professionals rests on the untested assumption that matching elite motion statistics produces winning hardware; no sensitivity analysis or comparison against lower-spec baselines is provided to support this sufficiency.

    Authors: The specifications were derived by extracting workspace, velocity, force, and accuracy requirements directly from motion capture of elite players. The manuscript demonstrates that these targets can be met through the described design process. We acknowledge that no sensitivity analysis or baseline comparisons are included to prove sufficiency for winning matches. We will add a short limitations paragraph in the discussion section noting that the specs represent necessary conditions based on player analysis but that sufficiency for consistent match outcomes would require additional validation. revision: yes

Circularity Check

0 steps flagged

No circularity: specs from external motion analysis, performance from physical experiments

full rationale

The paper analyzes elite players' motions to set target specifications for workspace, payload, velocity, etc., then builds an 8-DoF robot to those specs and reports experimental swing metrics (0.8 s cycle, 22 m/s) plus match outcomes. No equations, fitted parameters, or predictions are shown to reduce by construction to the inputs. No self-citations, uniqueness theorems, or ansatzes are invoked in a load-bearing way. The derivation chain is self-contained against external benchmarks (player motion data and hardware tests), so the central claim does not collapse into a renaming or fit.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The work is an engineering design report; the abstract mentions no mathematical axioms, free parameters, or invented entities beyond standard robotics components.

pith-pipeline@v0.9.1-grok · 5699 in / 989 out tokens · 29710 ms · 2026-06-26T05:22:04.054696+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

18 extracted references · 2 canonical work pages

  1. [1]

    Japan’s 1000 players (table tennis statistics),

    I. Ogimura, “Japan’s 1000 players (table tennis statistics),” 1994. [Online]. Available: https://ogi-museum.com/1994-nihonnosennin-2

    1994

  2. [2]

    The ping pong robot to return a ball precisely

    K. Asai, M. Nakayama, and S. Yase, “The ping pong robot to return a ball precisely.”Omron Technics, vol. 51, pp. 1–6, 2020

    2020

  3. [3]

    Achieving human level competitive robot table tennis,

    D. B. D’Ambrosio, S. W. Abeyruwan, L. Graesser, A. Iscen, H. B. Amor, A. Bewley, B. Reed, K. Reymann, L. Takayama, Y . Tassa,et al., “Achieving human level competitive robot table tennis,” in7th Robot Learning Workshop: Towards Robots with Human-Level Abilities, 2025

    2025

  4. [4]

    Outplaying elite table tennis players with an autonomous robot,

    P. D. et al., “Outplaying elite table tennis players with an autonomous robot,”Nature, vol. 652, pp. 886–891, 2026

    2026

  5. [5]

    International table tennis feder- ation (ittf),

    International Table Tennis Federation, “International table tennis feder- ation (ittf),” At: https://www.ittf.com, 2025

    2025

  6. [6]

    What happened after the nature paper: Ace vs. professional players,

    SonyAI, “What happened after the nature paper: Ace vs. professional players,” 2026. [Online]. Available: https://ai.sony/blog/ what-happened-after-the-nature-paper-ace-vs.-professional-players

    2026

  7. [7]

    Analysis of a movement of a top table tennis player using a direct linear transformation (dlt) method,

    K. Yoshida, Y . Iimoto, Y . Ushiyama, M. Kaga, and K. Suzuki, “Analysis of a movement of a top table tennis player using a direct linear transformation (dlt) method,”Japanese Journal of Sport Education Studies, vol. 11, no. 2, pp. 91–102, 1991

    1991

  8. [8]

    Topology optimization for multi-component robotic arms under time-varying loads,

    C. Wu, Y . Xu, J. Fang, G. Sun, G. P. Steven, and Q. Li, “Topology optimization for multi-component robotic arms under time-varying loads,”Structural and Multidisciplinary Optimization, vol. 68, no. 9, p. 188, 2025. [Online]. Available: https://doi.org/10. 1007/s00158-025-04129-1

    2025

  9. [9]

    Enhancing the reliability of a robotic arm through lightweighting and vibration control with modal analysis and topology optimization,

    M. Alshihabi, M. Ozkahraman, and M. Y . Kayacan, “Enhancing the reliability of a robotic arm through lightweighting and vibration control with modal analysis and topology optimization,”Mechanics Based Design of Structures and Machines, vol. 53, no. 3, pp. 1950–1974, 2025. [Online]. Available: https://doi.org/10.1080/15397734.2024.2400207

    work page doi:10.1080/15397734.2024.2400207 1950

  10. [10]

    A topology optimization method of robot lightweight design based on the finite element model of assembly and its applications,

    L. Sha, A. Lin, X. Zhao, and S. Kuang, “A topology optimization method of robot lightweight design based on the finite element model of assembly and its applications,”Science Progress, vol. 103, no. 3, p. 0036850420936482, 2020, pMID: 32609583. [Online]. Available: https://doi.org/10.1177/0036850420936482

    work page doi:10.1177/0036850420936482 2020

  11. [11]

    [Online]

    Altair Engineering Inc.,Altair OptiStruct, Altair Engineering Inc., Troy, MI, USA, 2025, version 2025, accessed April 22, 2026. [Online]. Available: https://altair.com/optistruct

    2025

  12. [12]

    The pinocchio c++ library – a fast and flexible implementation of rigid body dynamics algorithms and their analytical derivatives,

    J. Carpentier, G. Saurel, G. Buondonno, J. Mirabel, F. Lamiraux, O. Stasse, and N. Mansard, “The pinocchio c++ library – a fast and flexible implementation of rigid body dynamics algorithms and their analytical derivatives,” inIEEE International Symposium on System Integrations (SII), 2019

    2019

  13. [13]

    Pinocchio: fast forward and inverse dynamics for poly-articulated systems,

    J. Carpentier, F. Valenza, N. Mansard,et al., “Pinocchio: fast forward and inverse dynamics for poly-articulated systems,” https://stack-of- tasks.github.io/pinocchio, 2015–2021

    2015

  14. [14]

    [Online]

    The MathWorks Inc.,System Identification Toolbox version: 25.1 (R2025a), Natick, MA, USA, 2025. [Online]. Available: https: //www.mathworks.com

    2025

  15. [15]

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

    Y . Hu, F. Schiling, V . Cavinato, A. Aydin, A. Politis, R. T. Morales, K. Yannick, 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

  16. [16]

    Sigma-X Servopacks

    Y ASKAW A Electoric Corp., “Sigma-X Servopacks.” [On- line]. Available: https://www.yaskawa.eu.com/motion-control/sigma-x/ sx-servopacks

  17. [17]

    EtherCAT Specification

    EtherCAT Technology Group, “EtherCAT Specification.” [Online]. Available: https://ethercat.org

  18. [18]

    AdEXP1572 PCI Express EtherCAT master board

    Advanet Inc., “AdEXP1572 PCI Express EtherCAT master board.” [Online]. Available: https://www.advanet.co.jp/products/ industrial-network/ethercat/adexp1572/