Pith. sign in

REVIEW 4 major objections 4 minor 4 cited by

HITTER: A HumanoId Table TEnnis Robot via Hierarchical Planning and Learning

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A hierarchical planner-plus-learning system lets a general-purpose humanoid return table tennis balls at 92.3% and rally 106 shots.

desk verdict A genuine first: a general-purpose humanoid with agile locomotion returning table tennis balls in real time, but the headline rally is a single cooperative trial and the printed planner equation has a sign error that must be fixed. read the letter →

arxiv 2508.21043 v2 pith:AMF6ALF3 submitted 2025-08-28 cs.RO

Zhi Su , Bike Zhang , Nima Rahmanian , Yuman Gao , Qiayuan Liao , Caitlin Regan , Koushil Sreenath , S. Shankar Sastry This is my paper · ORCID
classification cs.RO
keywords humanoidroboticstabletenniswhole-bodycontrolreinforcementlearningmodel-basedplanningballtrajectorypredictionagilemanipulationsim-to-real
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that real-time humanoid table tennis is achievable with a two-level control architecture rather than end-to-end learning or specialized hardware. A model-based planner predicts the ball's flight, bounce, and the racket's needed position, velocity, and strike time; a reinforcement-learning whole-body controller then turns those targets into coordinated arm and leg motions. On a general-purpose humanoid, the system returns 24 of 26 test balls, plays rallies of up to 106 consecutive shots against a human, and can answer a smash with about 0.42 s of reaction time. The case matters because it shows agile manipulation of fast-moving objects, not just walking or static reaching, on a humanoid body.

What carries the argument

The load-bearing object is the virtual hit plane at x = -1.37 m, combined with Eq. (6), a closed-form racket-velocity rule: given the predicted incoming ball velocity and a desired outgoing velocity, it solves the required racket velocity under a coefficient-of-restitution contact model. This reduces the whole problem to a target-reaching task that the learned whole-body policy can track, while the base-position command lets the robot step to the ball. The hybrid flight/bounce model (Eq. 1) supplies the predictions that make the targets computable.

What would settle it

Run the same 26-ball protocol with balls aimed to land short of the fixed virtual hit plane (drop shots) or with visible topspin; if the return rate falls far below 92.3%, the fixed-plane and no-spin assumptions, not the controller, are carrying the result.

Watch

Extended reading notes

Core claim

The paper's central claim is that fast humanoid table tennis can be decomposed cleanly: high-level prediction and planning that is model-based, and low-level whole-body execution that is learned. The planner uses a hybrid flight-and-bounce ball model with drag and restitution coefficients, a fixed virtual hit plane, and a closed-form racket-velocity rule that converts the desired landing point into the racket motion at impact. The controller, trained with two human video-derived reference swings, tracks the racket target relative to the base and steps to reach balls across the table. Deployed fully autonomously, the system achieves a 96.2% hit rate and 92.3% return rate on a 26-ball test, up

Load-bearing premise

The headline results assume an opponent who keeps the ball within a fixed virtual hit plane with negligible spin; shots that land short, deep, or with heavy spin have no planned strike solution, so the rally numbers do not transfer to uncooperative play.

Editorial extensions

If this is right

  • Replacing the fixed hit plane with a reachable set over the whole table is the direct next step; the same controller should work with multi-surface targets.
  • Because the low-level policy is only given racket and base targets, the same controller can be reused for other fast interception sports by changing the planner's output.
  • Two-humanoid rallies show the policy is stable without opponent-specific training, so multi-agent co-play can be layered on later.
  • The small reference set, two swings, plus reinforcement learning yields human-like waist rotation and stepping, suggesting motion references need not be exhaustive.
  • The sub-second reactions measured in real play quantify the control-loop budget needed for humanoid ball sports.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if the fixed-plane constraint is relaxed to a learned reachability field, the same planner-controller split could cover short and wide balls without retraining the low-level policy.
  • Editorial inference: the 106-shot rally is a cooperative-opponent ceiling; a skilled human varying spin and depth would likely break the rally quickly, providing a sharper test of dexterity than rally length.
  • Editorial inference: replacing external motion capture with vision-based ball tracking would test whether millimeter-level 360 Hz perception is essential or whether the framework transfers beyond controlled environments.
  • Editorial inference: the flat-push stroke repertoire and spin-neglect approximation imply the current system competes at amateur pace, not expert pace; adding spin perception is the largest untested extension.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper presents HITTER, a hierarchical system for humanoid table tennis on a Unitree G1 robot. A model-based planner estimates the ball trajectory from motion capture and computes the racket striking position, velocity, and time using a virtual hit plane, while a reinforcement-learning whole-body controller trained in Isaac Lab tracks those targets and incorporates human motion references. Real-world experiments report a 96.2% hit rate and a 92.3% return rate over 26 balls, rallies of up to 106 consecutive shots against a human opponent, sustained humanoid-humanoid exchanges, and simulation-based agility metrics. The central claim is that this architecture enables sub-second reactive whole-body control for dynamic manipulation on a general-purpose humanoid.

Significance. If the reported results hold and the method description is corrected, this is a valuable demonstration of real-time reactive whole-body manipulation on a commodity humanoid. The paper's strengths include the modular separation of planning and control, the use of only two human motion references, zero-shot sim-to-real deployment, and a clear limitations section. The system-level demonstration of humanoid-humanoid rallies is also noteworthy. However, the printed planning law in Eq. (5) is internally inconsistent with the coordinate convention, and the real-world evaluation is presented with limited statistical context. These issues are fixable but must be addressed before the claims can be fully assessed.

major comments (4)
  1. [§IV-C, Eq. (5)] There is a sign error in the gravity term under the paper's coordinate convention. The coordinate frame has z upward (§III) and g is the gravity vector in Eq. (1a), so g = (0,0,-9.81). The standard projectile solution is v_o = (p_l - p_racket)/Δt - (1/2)gΔt, but Eq. (5) prints a plus sign. For a representative hit from z_racket ≈ 0.3 m to the table at z_l = 0 with Δt ≈ 0.45 s, the printed equation gives v_o,z ≈ -2.9 m/s, so the ball would hit the table within about 0.1 s and could not clear the net. Since Eq. (6) computes the racket velocity from v_o, this is the only bridge from trajectory prediction to the racket command. Please correct the sign and state explicitly whether the deployed software uses Eq. (5) as printed or the corrected version; if the code uses the corrected sign, the equation is a typo that must be fixed for reproducibility.
  2. [§VI-C and §VII-B] The headline real-world results are reported as a single 26-ball evaluation and one 106-shot rally, with no number of attempts, selection criteria, or variance. As the limitations in §VII-B state, the human opponent must avoid hitting overly short balls because of the fixed virtual hit plane; the rally therefore operates under a cooperative reachability constraint. To support the central empirical claim, please report the number of rally attempts and their distribution, the instructions given to the human opponent, and a quantitative account of the humanoid-humanoid exchanges (which are described only as 'sustained'). This is not a demand for a full tournament, but the current reporting makes it difficult to distinguish a robust capability from a favorable best-of trial.
  3. [§VI-A, Fig. 3] The prediction-error curves reaching exactly zero at the moment of contact are partly definitional: the state estimate is fit to the same position measurements used to define the strike point, so the error is expected to vanish at the current time regardless of predictive skill. Please report prediction errors at fixed horizons before contact (e.g., 0.2, 0.3, 0.5 s) with confidence intervals, and clarify how the ground-truth strike position and time are measured independently of the fitted trajectory. As presented, the figure does not quantify how far ahead the planner remains reliable, which is the actual question for sub-second reactive control.
  4. [§VI-B, Fig. 4] The agility evaluation is purely in simulation. The real-world evidence for agile reaching is a single qualitative observation in Fig. 5. Since the abstract claims 'sub-second reactive control' and the contributions emphasize agility, please provide quantitative real-world measures—for example, base-position convergence time or strike success as a function of lateral reach distance—or explicitly mark the simulation result as a forward indicator that has not yet been validated on hardware. This would prevent overstating the hardware-level agility claim.
minor comments (4)
  1. [Title/Abstract] The title contains the typos 'HumanoId' and 'TEnnis'; please correct to 'Humanoid' and 'Tennis'. Also, 'anthromorphically' in Sec. I should be 'anthropomorphically'.
  2. [§IV-B, Eq. (4)] The horizontal restitution equation has a misplaced brace: it should read {||v_x^+||, ||v_y^+||} = C_h {||v_x^-||, ||v_y^-||}. The current formatting is confusing.
  3. [§V-A] The phrase 'the 43nd frame' should be 'the 43rd frame'.
  4. [§VI-C] For the 26-ball test, please state how the incoming balls were generated (hand-thrown or machine-launched), whether they included spin, and whether the robot's choice of forehand/backhand was predetermined or autonomous. This would help interpret the 96.2% hit rate.

Circularity Check

2 steps flagged · score 2.0 of 10

No load-bearing circularity: the 106-shot rally and 92.3% return rate are external real-world experiments. Two peripheral construction-equality steps (human-likeness via imitation reward, forehand/backhand via heuristic) do not force the headline result.

  1. self definitional [Sec. V-B2 (Eq. 7); Sec. VI-B (Fig. 6)]
    "r = wiri + wgrg + wrrr, (7) where ri encourages imitation of the upper body reference motion ... training with human motion references produces striking behaviors that closely resemble human motions, including waist rotation during the hit, as demonstrated in Fig. 6."

    The claimed 'human-like' striking behavior is the direct output of the imitation reward ri, which minimizes deviation from the recorded human reference clips (forehand/backhand). The optimizer's objective is exactly to reproduce the reference, so the observed resemblance is enforced by the loss function by construction, not derived from independent first principles. This step is peripheral: the headline 106-shot rally is a real-world measurement and does not reduce to this reward.

  2. self definitional [Sec. V-B3; Sec. VI-C]
    "Based on pbase,xy and the predicted racket position ˆpracket, we heuristically determine whether a forehand or backhand strike should be used. ... We also observe that the robot tends to use forehand strokes for balls incoming at y <0 and backhand strokes for balls incoming at y >0, which is consistent with human table tennis play."

    The forehand/backhand 'observation' is hard-wired by the heuristic described in Sec. V-B3, which selects the swing type from the predicted racket position (which depends on the ball's y-coordinate). Reporting this tendency as an empirical finding consistent with human play is a restatement of the controller's input logic, not an emergent behavior discovered from the experiments. This is peripheral to the central rally claim.

full rationale

The paper's central claims—92.3% return rate (24/26 balls), a 106-shot rally against a human opponent, and sustained humanoid-humanoid rallies—are real-world experimental measurements. They are not computed by any equation in the paper and therefore cannot reduce to their inputs by construction. The planner's strike position/time/velocity predictions are forward integrations of a fitted ball model (k, Ch, Cv estimated from 15 trajectories, Eq. 4); these fitted parameters are inputs, and the forecast errors of Fig. 3 are evaluated empirically, not presented as the fitted values themselves. Equations (5)-(6) convert a hand-set landing target (center of the opponent's side) into a racket command; the paper never claims to 'predict' the landing point as a validated output. The two genuine construction-equalities are peripheral: (i) 'human-like' striking motion is the minimizer of the imitation reward ri that contains the human reference, so the resemblance is a training target rather than an emergent discovery; and (ii) the observed forehand/backhand choice is produced by the explicit heuristic in Sec. V-B3, not by the policy alone. Both are disclosed in the text and neither bears on the rally claim. The only self-citation used in the method, BeyondMimic [2] (shared authors Liao and Sreenath), describes preprocessing details; the paper gives the processing steps and the headline result is external to that citation, so it is not load-bearing. The limitation that 'a human opponent must avoid hitting overly short balls' (Sec. VII-B) is honestly stated and does not affect circularity. Separately, reviewers have flagged a possible sign error in Eq. (5); this is an internal-consistency/correctness concern, not circularity, and does not change the circularity score.

Assumptions & free parameters 7 free parameters · 7 assumptions · 0 invented entities

The central claim depends on several fitted or chosen constants (ball model parameters, restitution coefficients, hit plane, flight time) and on the cooperative-opponent assumption. These are inputs, not results; no new physical entities are invented.

free parameters (7)
  • k, aerodynamic drag coefficient = not reported
    Estimated from 15 recorded ball trajectories via Eq. (4) and used in Eq. (1a) for trajectory prediction.
  • C_h, horizontal restitution coefficient = not reported
    Estimated via Eq. (4), used in Eq. (1b)/(2) for the bounce model.
  • C_v, vertical restitution coefficient = not reported
    Estimated via Eq. (4), used in Eq. (1b)/(2).
  • Δ t, predefined flight time from hit to landing = not reported
    Chosen constant in Eq. (5) to compute the desired outgoing ball velocity.
  • C_r, racket-ball restitution coefficient = not reported
    Assumed constant in Eq. (6) for computing the desired racket velocity.
  • virtual hit plane x = -1.37 m = -1.37 m
    Chosen plane at which the robot must strike; constrains reachable balls (Sec. IV-B, VII-B).
  • reward weights w_i, w_g, w_r = not reported
    Weights in Eq. (7) adopted from [35]; shape policy behavior but values are not given.
assumptions (7)
  • domain assumption Hybrid ball dynamics: drag = -k||v||v during flight; velocity jump v+ = C v- on bounce (Eq. 1).
    Standard model taken from [16]; accuracy is verified indirectly via prediction errors (Fig. 3).
  • domain assumption Ball spin is negligible, so Magnus force is ignored.
    Stated in Sec. IV-B; acknowledged limitation against spinning balls (Sec. VII-B).
  • domain assumption At impact the racket plane is perpendicular to the racket velocity vector.
    Assumed in Sec. IV-C to derive Eq. (6).
  • domain assumption After racket impact the ball flies under gravity only; tangential friction is neglected.
    Assumed in Sec. IV-C for the return model.
  • domain assumption External motion capture provides millimeter-accurate ball and base pose at 360 Hz.
    Sec. III; restricts deployment to controlled environments, acknowledged.
  • domain assumption The RL policy trained in simulation transfers zero-shot to the real robot.
    Sec. V; supported by deployment results but not formally guaranteed.
  • domain assumption The human opponent restricts shots to the reachable envelope, not too short or too deep.
    Sec. VII-B limitation: "a human opponent must avoid hitting overly short balls"; load-bearing for the 106-shot rally.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HITTER: A HumanoId Table TEnnis Robot via Hierarchical Planning and Learning." pith.science (2026). https://pith.science/paper/AMF6ALF3

@misc{pith2026250821043,
  author       = {Pith},
  title        = {Pith review of: HITTER: A HumanoId Table TEnnis Robot via Hierarchical Planning and Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AMF6ALF3}},
  note         = {Machine review of arXiv:2508.21043}
}
read the original abstract

Humanoid robots have recently achieved impressive progress in locomotion and whole-body control, yet they remain constrained in tasks that demand rapid interaction with dynamic environments through manipulation. Table tennis exemplifies such a challenge: with ball speeds exceeding 5 m/s, players must perceive, predict, and act within sub-second reaction times, requiring both agility and precision. To address this, we present a hierarchical framework for humanoid table tennis that integrates a model-based planner for ball trajectory prediction and racket target planning with a reinforcement learning-based whole-body controller. The planner determines striking position, velocity and timing, while the controller generates coordinated arm and leg motions that mimic human strikes and maintain stability and agility across consecutive rallies. Moreover, to encourage natural movements, human motion references are incorporated during training. We validate our system on a general-purpose humanoid robot, achieving up to 106 consecutive shots with a human opponent and sustained exchanges against another humanoid. These results demonstrate real-world humanoid table tennis with sub-second reactive control, marking a step toward agile and interactive humanoid behaviors.

Figures

Figures reproduced from arXiv: 2508.21043 by the authors.

Figure 1
Figure 1. Humanoid table tennis rallies. Our system enables both humanoid-humanoid (left) and humanoid-human (right) matches, achieving rallies of up to 106 consecutive shots against a human opponent. Project website: humanoid-table-tennis.github.io. Abstract— Humanoid robots have recently achieved impres￾sive progress in locomotion and whole-body control, yet they remain constrained in tasks that demand rapid interaction wit… view at source ↗
Figure 2
Figure 2. System overview. (a) The racket is mounted on the robot’s right wrist using a 3D-printed connector, and the ball is covered with reflective tape for motion capture. (b) The motion capture system tracks the ball position pball, the robot base position pbase, and the base forward vector ebase,x. (c) The model-based planner uses pball and the desired landing point pˆl to predict the racket’s striking position pˆracket,… view at source ↗
Figure 3
Figure 3. Prediction errors of the model-based planner. Striking position error (top) and striking time error (bottom) are evaluated over 20 ball trajectories. The shaded regions indicate the standard deviation, and the red dashed line marks the critical position error of 7.5 cm, corresponding to the racket radius. At 0.5 s before the strike, the position error falls below this threshold [PITH_FULL_IMAGE:figures/full_fig_p00… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Agility evaluation of the WBC policy. Based on 943 successful simulated trials (94.3% success rate), when the initial distance is within 0.75 m, the target can be reached in under 0.8 s on average, which is faster than the strike time of 0.86 s. Error bars denote stand…
Figure 6
Figure 6. Figure 6: Real-world human-like striking motion. The whole-body control policy generates human-like striking motions, including coordinated waist rotation during hits, which mimics the way humans play table tennis. 2) Reward Functions: To generate motions that both re￾semble the…
Figure 7
Figure 7. Figure 7: Return performance evaluation. The results are displayed on a virtual hit plane, with each block side representing 0.2 m. Of the 26 incoming balls distributed across the plane, the robot successfully returned 24 (blue: forehand strokes, green: backhand strokes), while …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 4 Pith papers

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

  1. VLASH: Real-Time VLAs via Future-State-Aware Asynchronous Inference

    cs.RO 2025-11 conditional novelty 7.0 of 10

    Conditioning a VLA on a forward-rolled future robot state, trained with temporal-offset augmentation, removes the prediction-execution delay in asynchronous control.

  2. VOFA: Visual Object Goal Pushing with Force-Adaptive Control for Humanoids

    cs.RO 2026-05 unverdicted novelty 6.0 of 10

    VOFA combines a high-level visuomotor policy with a low-level force-adaptive controller to let humanoids push objects up to 17 kg to arbitrary goals using only noisy onboard vision, achieving over 80% real-world success.

  3. CBF-RL: Safety Filtering Reinforcement Learning in Training with Control Barrier Functions

    cs.RO 2025-10 conditional novelty 6.0 of 10

    Training RL policies with a closed-form CBF safety filter plus CBF reward lets a Unitree G1 humanoid avoid obstacles and climb stairs without a runtime safety filter.

  4. Multi-Embodiment Locomotion at Scale with extreme Embodiment Randomization

    cs.RO 2025-09 conditional novelty 5.0 of 10

    A single locomotion policy trained on 50 simulated robots transfers without fine-tuning to real quadrupeds and to one real humanoid, though a second humanoid needed in-distribution training.

Reference graph

Works this paper leans on

37 extracted references · 30 canonical work pages · cited by 4 Pith papers

  1. [1]

    Learning humanoid locomotion with perceptive internal model,

    J. Long, J. Ren, M. Shi, Z. Wang, T. Huang, P. Luo, and J. Pang, “Learning humanoid locomotion with perceptive internal model,” in 2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025

    work page 2025

  2. [2]

    Beyondmimic: From motion tracking to versatile humanoid control via guided diffusion,

    Q. Liao, T. E. Truong, X. Huang, G. Tevet, K. Sreenath, and C. K. Liu, “Beyondmimic: From motion tracking to versatile humanoid control via guided diffusion,” arXiv preprint arXiv:2508.08241 , 2025

    arXiv 2025

  3. [3]

    Kungfubot: Physics-based humanoid whole-body control for learning highly-dynamic skills,

    W. Xie, J. Han, J. Zheng, H. Li, X. Liu, J. Shi, W. Zhang, C. Bai, and X. Li, “Kungfubot: Physics-based humanoid whole-body control for learning highly-dynamic skills,” arXiv preprint arXiv:2506.12851, 2025

    arXiv 2025

  4. [4]

    Hub: Learning extreme humanoid balance,

    T. Zhang, B. Zheng, R. Nai, Y . Hu, Y .-J. Wang, G. Chen, F. Lin, J. Li, C. Hong, K. Sreenath, et al., “Hub: Learning extreme humanoid balance,” arXiv preprint arXiv:2505.07294 , 2025

    arXiv 2025

  5. [5]

    Learning coor- dinated badminton skills for legged manipulators,

    Y . Ma, A. Cramariuc, F. Farshidian, and M. Hutter, “Learning coor- dinated badminton skills for legged manipulators,” Science Robotics, vol. 10, no. 102, p. eadu3922, 2025

    work page 2025

  6. [6]

    Athletic mobile manipulator system for robotic wheelchair tennis,

    Z. Zaidi, D. Martin, N. Belles, V . Zakharov, A. Krishna, K. M. Lee, P. Wagstaff, S. Naik, M. Sklar, S. Choi, et al. , “Athletic mobile manipulator system for robotic wheelchair tennis,” IEEE Robotics and Automation Letters, vol. 8, no. 4, pp. 2245–2252, 2023

    work page 2023

  7. [7]

    A biomimetic approach to robot table tennis,

    K. M ¨ulling, J. Kober, and J. Peters, “A biomimetic approach to robot table tennis,” in 2010 IEEE/RSJ International Conference on Intelligent Robots and Systems . IEEE, 2010, pp. 1921–1926

    work page 2010

  8. [8]

    A biomimetic approach to robot table tennis,

    ——, “A biomimetic approach to robot table tennis,” Adaptive Behav- ior, vol. 19, no. 5, pp. 359–376, 2011

    work page 2011

Show all 37 references
  1. [9]

    Learning to select and generalize striking movements in robot table tennis,

    K. M ¨ulling, J. Kober, O. Kroemer, and J. Peters, “Learning to select and generalize striking movements in robot table tennis,” The International Journal of Robotics Research , vol. 32, pp. 263–279, 03 2013

    2013

  2. [10]

    Modeling opponent actions for table-tennis playing robot,

    Z. Wang, A. Boularias, K. M ¨ulling, and J. Peters, “Modeling opponent actions for table-tennis playing robot,” Proceedings of the AAAI Conference on Artificial Intelligence , vol. 25, no. 1, pp. 1828–1829, Aug. 2011. [Online]. Available: https: //ojs.aaai.org/index.php/AAAI/a...

    2011

  3. [11]

    Learning anticipation policies for robot table tennis,

    Z. Wang, C. H. Lampert, K. M ¨ulling, B. Sch ¨olkopf, and J. Pe- ters, “Learning anticipation policies for robot table tennis,” in 2011 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2011, pp. 332–337

    2011

  4. [12]

    Probabilistic movement modeling for intention inference in human-robot interaction,

    Z. Wang, K. Muelling, M. P. Deisenroth, H. Ben Amor, D. V ogt, B. Schoelkopf, and J. Peters, “Probabilistic movement modeling for intention inference in human-robot interaction,” International Journal of Robotics Research , vol. 32, no. 7, pp. 841 – 858, June 2013

    2013

  5. [13]

    Anticipatory action selection for human–robot table tennis,

    Z. Wang, A. Boularias, K. M ¨ulling, B. Sch ¨olkopf, and J. Peters, “Anticipatory action selection for human–robot table tennis,” Artificial Intelligence, vol. 247, pp. 399–414, 2017, special Issue on AI and Robotics. [Online]. Available: https://www.sciencedirect.com/science/...

    2017

  6. [14]

    Learning strategies in table tennis using inverse reinforcement learn- ing,

    K. M ¨ulling, A. Boularias, B. Mohler, B. Sch ¨olkopf, and J. Peters, “Learning strategies in table tennis using inverse reinforcement learn- ing,” Biological Cybernetics, vol. 108, no. 5, pp. 603–619, 2014

    2014

  7. [15]

    Latte-mv: Learning to anticipate table tennis hits from monocular videos,

    D. Etaat, D. Kalaria, N. Rahmanian, and S. S. Sastry, “Latte-mv: Learning to anticipate table tennis hits from monocular videos,” in Proceedings of the Computer Vision and Pattern Recognition Confer- ence, 2025, pp. 7115–7124

    2025

  8. [16]

    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,” in 2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025

    2025

  9. [17]

    Model- based trajectory prediction and hitting velocity control for a new table tennis robot,

    Y . Ji, X. Hu, Y . Chen, Y . Mao, G. Wang, Q. Li, and J. Zhang, “Model- based trajectory prediction and hitting velocity control for a new table tennis robot,” in 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2021, pp. 2728–2734

    2021

  10. [18]

    Robotic Table Tennis: A Case Study into a High Speed Learning System,

    D. B. D’Ambrosio, N. Jaitly, V . Sindhwani, K. Oslund, P. Xu, N. Lazic, A. Shankar, T. Ding, J. Abelian, E. Coumans, et al. , “Robotic Table Tennis: A Case Study into a High Speed Learning System,” in Proceedings of Robotics: Science and Systems , Daegu, Republic of Korea, July 2023

    2023

  11. [19]

    Sas- prompt: Large language models as numerical optimizers for robot self- improvement,

    H. B. Amor, L. Graesser, A. Iscen, D. D’Ambrosio, S. Abeyruwan, A. Bewley, Y . Zhou, K. Kalirathinam, S. Mishra, and P. Sanketi, “Sas- prompt: Large language models as numerical optimizers for robot self- improvement,” arXiv preprint arXiv:2504.20459 , 2025

    2025 arXiv

  12. [20]

    Achieving human level competitive robot table tennis,

    D. B. D’Ambrosio, S. Abeyruwan, L. Graesser, A. Iscen, H. B. Amor, A. Bewley, B. J. Reed, K. Reymann, L. Takayama, Y . Tassa, et al. , “Achieving human level competitive robot table tennis,” in 2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025

    2025

  13. [21]

    Sample-efficient reinforce- ment learning in robotic table tennis,

    J. Tebbe, L. Krauch, Y . Gao, and A. Zell, “Sample-efficient reinforce- ment learning in robotic table tennis,” in 2021 IEEE international conference on robotics and automation (ICRA) . IEEE, 2021, pp. 4171–4178

    2021

  14. [22]

    Impedance control and its effects on a humanoid robot playing table tennis,

    R. Xiong, Y . Sun, Q. Zhu, J. Wu, and J. Chu, “Impedance control and its effects on a humanoid robot playing table tennis,” International Journal of Advanced Robotic Systems , vol. 9, no. 5, p. 178, 2012

    2012

  15. [23]

    Expressive whole-body control for humanoid robots,

    X. Cheng, Y . Ji, J. Chen, R. Yang, G. Yang, and X. Wang, “Expressive whole-body control for humanoid robots,” in 20th Robotics: Science and Systems, RSS 2024 . MIT Press Journals, 2024

    2024

  16. [24]

    Omnih2o: Universal and dexterous human- to-humanoid whole-body teleoperation and learning,

    T. He, Z. Luo, X. He, W. Xiao, C. Zhang, W. Zhang, K. M. Kitani, C. Liu, and G. Shi, “Omnih2o: Universal and dexterous human- to-humanoid whole-body teleoperation and learning,” in 8th Annual Conference on Robot Learning , 2024

    2024

  17. [25]

    Falcon: Learning force-adaptive humanoid loco-manipulation,

    Y . Zhang, Y . Yuan, P. Gurunath, T. He, S. Omidshafiei, A.-a. Agha- mohammadi, M. Vazquez-Chanlatte, L. Pedersen, and G. Shi, “Falcon: Learning force-adaptive humanoid loco-manipulation,” arXiv preprint arXiv:2505.06776, 2025

    2025

  18. [26]

    Twist: Teleoperated whole-body imitation system,

    Y . Ze, Z. Chen, J. P. Ara ´ujo, Z. ang Cao, X. B. Peng, J. Wu, and C. K. Liu, “Twist: Teleoperated whole-body imitation system,” arXiv preprint arXiv:2505.02833, 2025

    2025 arXiv

  19. [27]

    Gmt: General motion tracking for humanoid whole-body control,

    Z. Chen, M. Ji, X. Cheng, X. Peng, X. B. Peng, and X. Wang, “Gmt: General motion tracking for humanoid whole-body control,” arXiv preprint arXiv:2506.14770, 2025

    2025 arXiv

  20. [28]

    Unitree g1 humanoid robot,

    Unitree Robotics, “Unitree g1 humanoid robot,” https://www.unitree. com/g1, 2025, accessed: 2025-08-14

    2025

  21. [29]

    Online optimal trajectory generation for robot table tennis,

    O. Koc ¸, G. Maeda, and J. Peters, “Online optimal trajectory generation for robot table tennis,” Robotics and Autonomous Systems , vol. 105, pp. 121–137, 2018

    2018

  22. [30]

    Orbit: A unified simulation framework for interactive robot learning environments,

    M. Mittal, C. Yu, Q. Yu, J. Liu, N. Rudin, D. Hoeller, J. L. Yuan, R. Singh, Y . Guo, H. Mazhar, A. Mandlekar, B. Babich, G. State, M. Hutter, and A. Garg, “Orbit: A unified simulation framework for interactive robot learning environments,” IEEE Robotics and Automa- tion Lette...

    2023

  23. [31]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  24. [32]

    Smpl: A skinned multi-person linear model,

    M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black, “Smpl: A skinned multi-person linear model,” in Seminal Graphics Papers: Pushing the Boundaries, Volume 2 , 2023, pp. 851–866

    2023

  25. [33]

    World-grounded human motion recovery via gravity-view coordinates,

    Z. Shen, H. Pi, Y . Xia, Z. Cen, S. Peng, Z. Hu, H. Bao, R. Hu, and X. Zhou, “World-grounded human motion recovery via gravity-view coordinates,” in SIGGRAPH Asia 2024 Conference Papers , 2024, pp. 1–11

    2024

  26. [34]

    Gmr: General motion retargeting,

    Y . Ze, J. P. Ara ´ujo, J. Wu, and C. K. Liu, “Gmr: General motion retargeting,” 2025, gitHub repository. [Online]. Available: https://github.com/YanjieZe/GMR

    2025

  27. [35]

    Deepmimic: Example-guided deep reinforcement learning of physics-based char- acter skills,

    X. B. Peng, P. Abbeel, S. Levine, and M. Van de Panne, “Deepmimic: Example-guided deep reinforcement learning of physics-based char- acter skills,” ACM Transactions On Graphics (TOG) , vol. 37, no. 4, pp. 1–14, 2018

    2018

  28. [36]

    Asymmetric actor critic for image-based robot learning,

    L. Pinto, M. Andrychowicz, P. Welinder, W. Zaremba, and P. Abbeel, “Asymmetric actor critic for image-based robot learning,” in 14th Robotics: Science and Systems, RSS 2018 . MIT Press Journals, 2018

    2018

  29. [37]

    RACER: Real time game theoretic motion planning and control in autonomous racing using near potential functions,

    D. Kalaria, C. Maheshwari, and S. Sastry, “RACER: Real time game theoretic motion planning and control in autonomous racing using near potential functions,” in Proceedings of Machine Learning Research, Conference on Learning and Control L4DC, vol. 283, 2025, pp. 1–18

    2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.