Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

A discriminative feature embedding that separates trajectories by flight dynamics sharply improves early impact-point prediction for both seen and never-seen objects.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 16:19 UTC pith:HI6LIIO2

load-bearing objection Worth serious refereeing for the new 20-object trajectory dataset and the early-stage IE gains, but the catching-success claim is much weaker than the prediction claim. the 3 major comments →

arxiv 2509.15254 v3 pith:HI6LIIO2 submitted 2025-09-18 cs.RO

OIPP: Object-Adaptive Impact Point Predictor for Catching Diverse In-Flight Objects

Ngoc Huy Nguyen , Kazuki Shibata , Takamitsu Matsubara This is my paper
classification cs.RO
keywords impact point predictionin-flight object catchingdiscriminative feature embeddingLSTMaerodynamicsquadruped robotgeneralization to unseen objectsreal-world trajectory dataset
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tackles the problem of a quadruped robot catching diverse in-flight objects by predicting the object's impact point from a short early segment of its trajectory. The authors construct a real-world dataset of 8,000 trajectories from 20 objects that exhibit complex aerodynamic effects, more diverse than existing near-parabolic datasets. They propose DIPP, which uses a Discriminative Feature Embedding (DFE) to map motion histories into a feature space where trajectories with similar dynamics cluster together, enabling early discrimination and generalization to unseen objects. Combined with an Impact Point Enhanced (IPE) loss that explicitly penalizes impact-point error, DIPP outperforms baselines on both seen and unseen objects, and improved early predictions translate into higher catching success in simulation and real-robot demonstrations. If the reported results hold, the combination of a diverse real-world dataset, a dynamics-discriminative embedding, and target-aligned loss is a reliable recipe for landing-point prediction across diverse, aerodynamically complex objects.

Core claim

The central claim is that separating trajectories by their underlying dynamics in a learned feature space, rather than treating all objects as a single motion model, is the key to accurate early-stage impact point prediction for diverse in-flight objects. The DFE, an LSTM-based encoder, embeds historical states (position, velocity, acceleration) so that trajectories with similar aerodynamic behavior are mapped close together. This clustering enables the Impact Point Predictor (IPP) to relate unseen objects to dynamically similar training objects, substantially reducing impact point error (IE) at large time steps to impact compared to baselines. Two IPP variants are evaluated: an NAE-based on

What carries the argument

The Discriminative Feature Embedding (DFE) is an LSTM encoder that maps a window of historical states (position, velocity, acceleration in 3D) into a feature vector. The key property enforced by training with the IPE loss is that trajectories with similar dynamics are placed close to each other in feature space, while dissimilar ones are separated. This clustering is what enables early-stage discrimination and generalization to unseen objects. The Impact Point Predictor (IPP) then consumes these features in two variants: the Neural Acceleration Estimator (NAE) autoregressively generates future states and computes the intersection with the catching plane, while the Direct Point Estimator (DPE

Load-bearing premise

The simulation that links prediction accuracy to catching success assumes a single-integrator robot with PID control faithfully represents the quadruped's ability to reach the predicted impact point, and that starting within 0.3 m of the true impact point is representative of real-world conditions.

What would settle it

A direct test would be to evaluate DIPP-NAE on a new set of objects whose aerodynamic responses are genuinely novel (e.g., fluttering paper or objects with unpredictable vortex shedding) and compare early-stage IE (e.g., at 70 time steps to impact) against the NAE baseline; if the IE gap disappears or reverses, the generalization claim of the discriminative embedding would be refuted. Additionally, a real-robot catching experiment with an unseen object, logged with quantitative success rates rather than qualitative snapshots, would settle whether the simulation-reported improvements transfer t

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Improved early-stage impact point prediction leads to higher catching success rates in simulation, with DIPP-NAE achieving the highest success rates for both seen and unseen objects across basket radii.
  • The DFE's discriminative clustering enables generalization to unseen objects by associating them with dynamically similar training trajectories, reducing prediction error without additional retraining.
  • The DPE variant offers a computationally efficient alternative for fixed-height catching that nearly matches the NAE variant's accuracy, potentially enabling real-time onboard prediction.
  • The constructed dataset of 8,000 trajectories from 20 diverse objects with complex aerodynamics provides a new benchmark for evaluating trajectory and impact point prediction methods.

Where Pith is reading between the lines

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

  • If the DFE's clustering genuinely captures dynamics rather than object identity, the same learned embedding could be repurposed for object dynamics classification or anomaly detection, with minimal architectural changes.
  • The IPE loss could be applied to other robot tasks where the ultimate objective is a specific endpoint of a predicted trajectory (e.g., interception or landing), not just full trajectory fidelity.
  • The reported simulation success rates may overestimate real-world performance because the robot model is a simplified single integrator and initial positions are placed within 0.3 m of the true impact point; actual quadruped dynamics and sensor noise could erode the gains seen in simulation.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper addresses impact-point prediction for robotic catching of diverse in-flight objects with complex aerodynamics. It introduces a real-world dataset of 2,000 measured trajectories from 20 objects, augmented to 8,000, and proposes DIPP (also referred to as OIPP in the abstract/title), which combines a discriminative feature embedding (DFE) based on an LSTM with an impact point predictor (IPP). Two IPP variants are implemented: an NAE-based trajectory predictor and a DPE-based direct predictor. A new IPE loss explicitly penalizes impact-point error. The method is evaluated on 15 seen and 5 unseen objects against Newton, SVR, and NAE baselines, with lower impact-point error reported especially at early stages. The paper also presents a simulation-based catching study and real-robot demonstrations, and claims that improved early-stage prediction enhances catching success.

Significance. If the results hold, the dataset and DIPP framework would be a useful empirical contribution: the held-out evaluation on 5 truly unseen objects is a strength, and the DFE ablation (DPE vs. DIPP-DPE) supports the role of the encoder. The PDS analysis provides a quantitative basis for the claim that the dataset is more complex than existing ones. However, the catching-success claim is currently supported only by a simplified single-integrator simulation and qualitative real-robot demos, and the contribution of the IPE loss is never isolated. These gaps prevent the paper from fully establishing the causal link between improved early prediction and catching success.

major comments (3)
  1. [Section IV-G, Table I] The simulation evidence for RQ3 is not sufficient to support the claim that improved early-stage prediction enhances catching success. The robot is modeled as a single integrator with a PID controller and max velocity 2.5 m/s, and is initialized uniformly within a 0.3 m radius of the ground-truth impact point. This omits acceleration limits, gait dynamics, and basket-placement constraints, and the success rates in Table I are strongly influenced by the chosen initial offset and basket radius. No sensitivity analysis is provided, and the real-robot validation is qualitative. The sentence in Section IV-G, 'the improved early-stage prediction accuracy contributes significantly to the catching performance,' is therefore not established. I recommend either using a more realistic robot model, varying the initial offset and time-to-impact systematically, or collecting quantitative real-robot su
  2. [Section III-C and Section IV-F] The IPE loss is presented as a key component of the method, but no experiment ablates it. In the RQ2 comparison, DIPP-NAE vs. NAE changes both the DFE and the IPE loss, while DPE vs. DIPP-DPE isolates the DFE but keeps the IPE loss in both variants. Consequently, the marginal contribution of the IPE loss to the reported impact-point-error improvements is unknown. An ablation such as DIPP-NAE or DPE trained without the IPE loss terms (i.e., without the trajectory-alignment and impact-point penalties) is needed to separate the effects of the loss from the effects of the architecture.
  3. [Section IV-H] The real-robot validation is reported only qualitatively. The text states that two seen and two unseen objects were tested, but Figure 8 shows only one seen object (boomerang) and one unseen object (pinwheel), with no trial counts, no success/failure statistics, and no comparison across repeated runs. The claim in the conclusion that the method is 'effective in the real world' is not quantitatively supported. Please provide a table with per-object trial counts and success rates, or substantially soften the real-robot claim and frame the demonstrations as anecdotal.
minor comments (6)
  1. [Title and Abstract] The arXiv title and abstract refer to 'OIPP: Object-Adaptive Impact Point Predictor', while the full text uses 'DIPP: Discriminative Impact Point Predictor'. This inconsistency must be fixed before publication.
  2. [Section IV-C] Several implementation details are missing: the history length T and prediction horizon K are not specified, the data-augmentation ranges are not given, and the PID gains used in simulation and on the real robot are not reported. These details are needed for reproducibility.
  3. [Eq. (2)] The indexing in Eq. (2) is unclear: the initial hidden state h_{t-T} is not defined, and the recurrence for i > T appears to pass only the previous hidden state without an input. Please clarify the initial condition and the exact recurrence.
  4. [Section IV-F, Fig. 5] The statistical significance markers (***) are not accompanied by a description of the test used, the pairing structure, or whether multiple comparisons were corrected. Please provide this information.
  5. [Section III-A] The term 'discriminative' is used to describe the feature embedding, but the training loss in Eq. (5)/(6) contains no explicit separation or metric-learning term. The t-SNE visualization shows separation, but it is post hoc. Consider clarifying that the separation is an emergent property, or add an explicit objective that enforces it.
  6. [Section I and Section IV-A] The dataset is described as '8,000 trajectories from 20 objects', but only 2,000 are measured and the rest are generated by translational and rotational augmentation. Please state this distinction clearly in the abstract and contribution list to avoid overstating the amount of real measured data.

Circularity Check

0 steps flagged

No meaningful circularity: predictions are evaluated on held-out and unseen-object data, and the method's components are not defined in terms of the claimed outputs.

full rationale

The paper's central claims are (i) a new dataset, (ii) a DFE+IPP architecture, and (iii) improved impact-point prediction on seen and unseen objects. The IE metric used for evaluation is the Euclidean distance between predicted and ground-truth impact points, and the training objective includes an IPE loss that directly penalizes this error. This is standard supervised learning, not a disguised fit: the reported IE values are computed on held-out test trajectories from 15 training objects and on 5 objects never seen during training. The IPE loss does not make the evaluation circular, because baselines are trained without it and the held-out/unseen evaluation is independent of the training fit. No parameter is fitted to the test set or to the unseen objects. The paper cites prior work only as external baselines or related methods (e.g., NAE [5], Hu et al. [6]); there are no self-citations carrying a load-bearing uniqueness or ansatz argument. The DFE's 'discriminative' property is asserted and visualized with t-SNE, but it is not derived from the evaluation metric by construction. The RQ3 simulation initializes the robot within 0.3 m of the true impact point, which may weaken the external-validity claim that improved IE transfers to real catching, but that is a question of experimental realism, not circular reasoning: the success-rate numbers are outcomes of a simulator, not identities or fitted values renamed as predictions. Similarly, the real-robot demonstration is qualitative, but qualitative evidence is not circular. The discussion section explicitly lists limitations (motion-capture dependence, fixed-height catching plane), which further supports that the authors do not present assumptions as derived results. Overall, the derivation chain is self-contained with respect to the stated predictions, and the evaluation protocol does not reduce any claimed result to its own input.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim is an empirical ML result; it rests on dataset quality, the discriminative-feature assumption, and simulation-to-reality transfer. No physical free parameters are fit, but several experimental choices (hyperparameters, T/K, augmentation ranges, simulation parameters) are under-specified and affect reproducibility of the reported margins.

free parameters (4)
  • Hyperparameters (learning rates 1e-4 and 3e-5, batch size 512, 128 hidden units, up to 3.0e4 epochs) = Section IV-C
    Chosen by empirical tuning on validation performance; baselines and variants use different learning rates, so tuning choices can affect the margin of the reported gains.
  • History length T and prediction horizon K = Not specified numerically
    The encoder window and the number of steps to the catching plane are defined in Section III-B but never given values, so the early-stage regime (70 time steps before impact) cannot be reproduced.
  • Data augmentation ranges = Not specified
    Translation and rotation augmentation expanded 2,000 to 8,000 trajectories (Section IV-A), but the ranges are not given, making dataset reproduction impossible.
  • Simulation parameters (max velocity 2.5 m/s, initial offset 0.3 m, basket radii 0.05 to 0.20 m) = Section IV-G
    These choices set the success-rate magnitudes in Table I; the robot is initialized close to the intercept point, which inflates SR across all methods.
axioms (4)
  • domain assumption Parabola deviation (PDS) is a valid measure of trajectory complexity and aerodynamic diversity
    Section IV-E defines PDS as average deviation from a least-squares parabola; measurement noise and throwing inconsistency also inflate PDS, and no error bars or noise floor are given for the comparison against the NAE dataset.
  • domain assumption The learned feature space transfers to unseen objects via dynamics similarity
    Section III-A assumes the DFE maps dynamically similar trajectories close together; the only evidence is a qualitative t-SNE plot (Fig. 6) with no quantitative cluster metric.
  • domain assumption Motion-capture ground truth is accurate and the catching plane is at a fixed known height
    Section IV-A and IV-H rely on mocap at 120 Hz for training targets and evaluation; the fixed-height catching plane constrains the task (acknowledged in Section V).
  • domain assumption Augmentation by translation and rotation preserves physically plausible trajectories
    Section IV-A uses augmentation to expand the dataset 4x; whether rotated trajectories remain physically realizable for hand-thrown objects is not validated.

pith-pipeline@v1.3.0-alltime-deepseek · 10435 in / 16721 out tokens · 162522 ms · 2026-08-04T16:19:03.772391+00:00 · methodology

0 comments
read the original abstract

In this study, we address the problem of in-flight object catching using a quadruped robot with a basket. Our objective is to accurately predict the impact point, defined as the object's landing position. This task poses two key challenges: the absence of public datasets capturing diverse objects under unsteady aerodynamics, which are essential for training reliable predictors; and the difficulty of accurate early-stage impact point prediction when trajectories appear similar across objects. To overcome these issues, we construct a real-world dataset of 8,000 trajectories from 20 objects, providing a foundation for advancing in-flight object catching under complex aerodynamics. We then propose the Object-Adaptive Impact Point Predictor (OIPP), consisting of two modules: (i) an Object-Adaptive Encoder (OAE) that extracts object-dependent representations from motion histories, and (ii) an Impact Point Predictor (IPP) that estimates the impact point from these representations. Two IPP variants are implemented: a Neural Acceleration Estimator (NAE)-based method that predicts trajectories and derives the impact point, and a Direct Point Estimator (DPE)-based method that directly outputs it. Experimental results show that our dataset is more diverse and complex than existing datasets, and that our method outperforms baselines on both 15 seen and 5 unseen objects. Furthermore, we show that improved early-stage prediction enhances catching success in simulation and demonstrate the effectiveness of our approach through real-robot experiments. The demonstration is available at https://sites.google.com/view/robot-catching-2025.

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. Pixel2Catch: Multi-Agent Sim-to-Real Transfer for Agile Manipulation with a Single RGB Camera

    cs.RO 2026-02 conditional novelty 6.0

    Pixel2Catch shows that pixel-level bounding-box cues from one RGB camera, with separate arm and hand reinforcement-learning policies, are enough to catch thrown objects in the real world.

Reference graph

Works this paper leans on

21 extracted references · cited by 1 Pith paper

  1. [1]

    Estimating the non-linear dynamics of free- flying objects,

    S. Kim and A. Billard, “Estimating the non-linear dynamics of free- flying objects,”Robotics and Autonomous Systems, vol. 60, no. 9, pp. 1108–1122, 2012

    2012

  2. [2]

    Catching objects in flight,

    S. Kim, A. Shukla, and A. Billard, “Catching objects in flight,”IEEE Transactions on Robotics, vol. 30, no. 5, pp. 1049–1065, 2014

    2014

  3. [3]

    Trajectory planning for optimal robot catching in real-time,

    R. Lampariello, D. Nguyen-Tuong, C. Castellini, G. Hirzinger, and J. Peters, “Trajectory planning for optimal robot catching in real-time,” in2011 IEEE International Conference on Robotics and Automation, 2011, pp. 3719–3726

    2011

  4. [4]

    A dynamical system approach for softly catching a flying object: Theory and experiment,

    S. S. M. Salehian, M. Khoramshahi, and A. Billard, “A dynamical system approach for softly catching a flying object: Theory and experiment,”IEEE Transactions on Robotics, vol. 32, no. 2, pp. 462– 471, 2016

    2016

  5. [5]

    Neural motion prediction for in-flight uneven object catching,

    H. Yu, D. Guo, H. Yin, A. Chen, K. Xu, Z. Chen, M. Wang, Q. Tan, Y . Wang, and R. Xiong, “Neural motion prediction for in-flight uneven object catching,” in2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021, pp. 4662–4669

    2021

  6. [6]

    Modular neural network policies for learning in-flight object catching with a robot hand-arm system,

    W. Hu, F. Acero, E. Triantafyllidis, Z. Liu, and Z. Li, “Modular neural network policies for learning in-flight object catching with a robot hand-arm system,” in2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2023, pp. 944–951

    2023

  7. [7]

    Event- based agile object catching with a quadrupedal robot,

    B. Forrai, T. Miki, D. Gehrig, M. Hutter, and D. Scaramuzza, “Event- based agile object catching with a quadrupedal robot,” in2023 IEEE International Conference on Robotics and Automation (ICRA), 2023, pp. 12 177–12 183

    2023

  8. [8]

    Cooperative quadrocopter ball throwing and catching,

    R. Ritz, M. W. M ¨uller, M. Hehn, and R. D’Andrea, “Cooperative quadrocopter ball throwing and catching,” in2012 IEEE/RSJ Interna- tional Conference on Intelligent Robots and Systems, 2012, pp. 4972– 4978

    2012

  9. [9]

    Playing catch and juggling with a humanoid robot,

    J. Kober, M. Glisson, and M. Mistry, “Playing catch and juggling with a humanoid robot,” in2012 12th IEEE-RAS International Conference on Humanoid Robots (Humanoids 2012), 2012, pp. 875–881

    2012

  10. [10]

    3D monocular robotic ball catching with an iterative trajectory estimation refinement,

    V . Lippiello and F. Ruggiero, “3D monocular robotic ball catching with an iterative trajectory estimation refinement,” in2012 IEEE International Conference on Robotics and Automation, 2012, pp. 3950–3955

    2012

  11. [11]

    EV-Catcher: High-speed object catching using low-latency event-based neural networks,

    Z. Wang, F. Cladera, A. Bisulco, D. Lee, C. J. Taylor, K. Daniilidis, M. A. Hsieh, D. D. Lee, and V . Isler, “EV-Catcher: High-speed object catching using low-latency event-based neural networks,”IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 8737–8744, 2022

    2022

  12. [12]

    Agile catching with whole-body MPC and blackbox policy learning,

    S. Abeyruwan, A. Bewley, N. M. Boffi, K. M. Choromanski, D. B. D’Ambrosio, D. Jain, P. R. Sanketi, A. Shankar, V . Sindhwani, S. Singh,et al., “Agile catching with whole-body MPC and blackbox policy learning,” inLearning for Dynamics and Control Conference, 2023, pp. 851–863

    2023

  13. [13]

    Run and catch: Dynamic object-catching of quadrupedal robots,

    Y . You, T. Liu, X. Liang, Z. Xu, M. Zhou, Z. Li, and S. Zhang, “Run and catch: Dynamic object-catching of quadrupedal robots,” in2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2023, pp. 743–750

    2023

  14. [14]

    Impact-friendly object catching at non-zero velocity based on combined optimization and learning,

    J. Zhao, G. J. G. Lahr, F. Tassi, A. Santopaolo, E. De Momi, and A. Ajoudani, “Impact-friendly object catching at non-zero velocity based on combined optimization and learning,” in2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2023, pp. 4428–4435

    2023

  15. [15]

    Dynamic object catching with quadruped robot front legs,

    A. Schakkal, G. Bellegarda, and A. Ijspeert, “Dynamic object catching with quadruped robot front legs,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024, pp. 6848– 6855

    2024

  16. [16]

    Impact-aware bimanual catching of large-momentum objects,

    L. Yan, T. Stouraitis, J. Moura, W. Xu, M. Gienger, and S. Vijayaku- mar, “Impact-aware bimanual catching of large-momentum objects,” IEEE Transactions on Robotics, vol. 40, pp. 2543–2563, 2024

    2024

  17. [17]

    IMA-catcher: An impact-aware nonprehensile catching framework based on combined optimization and learning,

    F. Tassi, J. Zhao, G. J. Lahr, L. Gava, M. Monforte, A. Glover, C. Bartolozzi, and A. Ajoudani, “IMA-catcher: An impact-aware nonprehensile catching framework based on combined optimization and learning,”The International Journal of Robotics Research, p. 02783649251345851, 2025

    2025

  18. [18]

    Offset-free model predictive control: A ball catching application with a spherical soft robotic arm,

    Y . Huang, M. Hofer, and R. D’Andrea, “Offset-free model predictive control: A ball catching application with a spherical soft robotic arm,” in2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021, pp. 563–570

    2021

  19. [19]

    Catch planner: Catching high-speed targets in the flight,

    H. Yu, P. Wang, J. Wang, J. Ji, Z. Zheng, J. Tu, G. Lu, J. Meng, M. Zhu, S. Shen, and F. Gao, “Catch planner: Catching high-speed targets in the flight,”IEEE/ASME Transactions on Mechatronics, vol. 28, no. 4, pp. 2387–2398, 2023

    2023

  20. [20]

    Quadrocopter ball jug- gling,

    M. M ¨uller, S. Lupashin, and R. D’Andrea, “Quadrocopter ball jug- gling,” in2011 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2011, pp. 5113–5120

    2011

  21. [21]

    Catch the ball: Accurate high-speed motions for mobile manipulators via inverse dynamics learning,

    K. Dong, K. Pereida, F. Shkurti, and A. P. Schoellig, “Catch the ball: Accurate high-speed motions for mobile manipulators via inverse dynamics learning,” in2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020, pp. 6718–6725

    2020