acceptodds
Under review as a conference paper at ICLR 2027

HydroWrench: A Dataset and Benchmark for Dynamic Load Reconstruction in Bioinspired Aquatic Mechanisms

Abstract

Dynamic loads shape the propulsive performance and structural response of bioinspired aquatic mechanisms, making their reconstruction important for surrogate modeling and motion optimization. We introduce HydroWrench, a dataset and benchmark of 116,756 command–load trajectories from caudal-fin, webbed-foot, and flipper mechanisms representative of added-mass, drag-based, and lift-based propulsion. The dataset provides recorded joint-angle commands and derived kinematics as inputs, with measured six-axis interface loads as targets. With executed motion and flow unobserved, the benchmark compares nonlinear representation and additional command history under a shared evaluation protocol, with models fitted separately for each mechanism. A multilayer perceptron reduces normalized mean squared error by 25.5–81.0% relative to linear regression with matched features. Additional history improves time-varying reconstruction for the webbed foot and flipper, while caudal-fin model rankings depend on the temporal scoring grid. In the webbed-foot high-frequency holdout, a long-history predictor reduces error by 15.1% relative to its matched local-input counterpart, while its error remains 32.3% higher than under in-distribution conditions. These results connect reconstruction accuracy to temporal error structure and motion-condition coverage. HydroWrench provides a shared experimental basis for load-relevant representations, dynamic surrogate modeling, and input–output system identification in partially observed fluid–mechanism systems.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
datasets and benchmarks, physical systems, dynamic load reconstruction, bioinspired aquatic mechanisms, system identification

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_HS45OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

HydroWrench contains 116,756 measured command-to-load trajectories from a caudal fin, a webbed foot, and a flipper. Models reconstruct six interface force/moment channels from commanded angles and their derivatives, with separate training for each mechanism. The experiments compare local nonlinear predictors with models using finite history, then look at waveform error, high-frequency holdouts and training-data quantity. Nonlinear prediction helps in all three domains. Extra history helps most on the webbed foot, while the caudal-fin ranking changes when scoring densely instead of at the prescribed endpoints. I consider this mainly a dataset and benchmark contribution.

Critical strengths

The measured force/moment data at this scale could be useful for studying partially observed physical systems. The authors define the observations clearly: commands are the inputs, and the loads contain both mechanical and fluid effects. They keep acquisition groups together when splitting the data and evaluate W1 and longer windows on the same valid interval. The standard-model comparisons in Table 3 suit the dataset contribution. Webbed-foot LSTM error, for example, falls from 0.3910 to 0.3390 with history.

I found the waveform analysis interesting. Nearly all of the webbed-foot history gain is in the varying response, as shown by separating trial-mean error from centered-waveform error. History still reduces OOD error by 15.1% in the high-frequency experiment, although the long-window model has a 32.3% ID-to-OOD gap. The authors also report the caudal-fin ranking reversal openly and keep seed variability separate from acquisition-group uncertainty.

Critical weaknesses

I am less convinced about the temporal preprocessing. Commands have only nominal 20-Hz timing. Appendix A also reports gaps up to 0.174 s in the audited webbed-foot sensor records, which are filled by interpolation. I could not tell how common these gaps are or how much of the evaluated signal falls inside them. Since most of the history benefit concerns waveform reconstruction, I would want to know whether it survives different preprocessing choices. The offline setting with two-sided derivatives is reasonable, but the sensitivity of the temporal findings remains unclear.

It is hard to check this from the submitted benchmark. Section 3.2 provides only example trajectories and defers the full data, split identities and code until acceptance. Appendix A says the amplitude-screening rule was developed post hoc using records from all three splits. That alone does not establish bias. Still, for a dataset paper I would expect counts of the removed records and a check using thresholds derived from training data only. My recommendation is weak reject: the resource looks useful, but its data-quality effects need closer examination, and the complete benchmark is not yet available to inspect.

Questions for authors

What fraction of trials and evaluated points have timestamp gaps, by mechanism and split? Please check whether the main history gains remain after masking long gaps and the affected input windows. I would also like to see how many trials the amplitude screen removes from each split, and whether using training-only thresholds changes the main results.

Can you make the complete anonymized dataset, split manifests and evaluation code available during review? Also, why keep sparse endpoint error as the primary score when dense scoring reverses the model ranking?

1 comment
AuthorsOpenReview ↗

Looking forward to your feedback

We sincerely appreciate your time and consideration of our manuscript. We have carefully addressed your comments and provided detailed responses and clarifications to each point. We would be grateful if you could review our revised responses and reconsider your assessment of the manuscript in light of the additional explanations and revisions. Please let us know if you have any further questions or concerns.

Reviewer_pTxzOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

HydroWrench introduces a dataset and benchmark consisting of 116,756 command-to-load trajectories derived from three bioinspired aquatic mechanisms: a 1-DOF caudal fin, a 2-DOF webbed foot, and a 3-DOF flipper. The primary task is to reconstruct measured six-axis interface loads (forces and moments) using recorded joint-angle commands and numerically derived kinematics as inputs. The authors evaluate various linear models and non-linear sequence models (such as MLP, LSTM, TCN, and Mamba-2) under a unified evaluation protocol. The study investigates the impact of historical motion context on prediction accuracy, introduces an error decomposition method to analyze waveform fidelity, and assesses model generalization under out-of-distribution (high-frequency) motion conditions.

Critical strengths
  1. Valuable Real-World Experimental Data: Unlike many existing scientific machine learning (SciML) benchmarks that rely heavily on simulated PDE data, HydroWrench provides real physical measurements from a controlled experimental platform. This is highly consequential as it provides a grounded, shared experimental basis for studying partially-observed fluid-mechanism systems.
  2. Rigorous Protocol and Insightful Error Decomposition: The benchmark is thoughtfully constructed with fixed temporal grids and acquisition-group splits. Furthermore, decomposing the Mean Squared Error (MSE) into trial-mean error and centered-waveform error is highly informative. This specific analysis successfully isolates how providing historical motion context primarily improves the time-varying shape of the predictions rather than just the average load levels.
Critical weaknesses
  1. Non-Causal Offline Preprocessing Limits Applicability: The input velocity and acceleration features are derived using a five-point, third-degree Savitzky-Golay (SG5) polynomial, which requires neighboring command samples on both sides of the evaluation point. Because this relies on future states, the benchmark strictly evaluates offline reconstruction. This limits the immediate utility of the trained models for real-time, online robotic control applications, which is a stated motivation for learning these dynamic loads.
  2. Limited Environmental Flow Variation: All experimental trials in the dataset were conducted at a single fixed towing speed of 0.167 m/s. While the dataset thoroughly explores variations in joint kinematics and includes high-frequency out-of-distribution tests, the lack of variation in the background flow conditions restricts the benchmark's ability to assess how well these surrogate models generalize across different fluid environments.
Questions for authors
  1. Could you provide a baseline evaluation using only causal filtering for the kinematic derivatives? Understanding the exact performance penalty when future states are strictly inaccessible is critical for determining whether these models can be used in online control loops.
  2. The out-of-distribution (OOD) evaluation shows a significant performance gap for high-frequency motions, particularly in the webbed-foot experiments. Based on your results, do you believe purely data-driven sequence models can overcome this extrapolation gap, or does this indicate a fundamental necessity to incorporate physics-informed inductive biases (such as rigid-body dynamic equations) into the models?
  3. Are there future plans to expand this dataset to include varying towing speeds to test generalization against changes in the baseline flow state?
1 comment
AuthorsOpenReview ↗

Looking forward to your feedback

We sincerely appreciate your time and consideration of our manuscript. We have carefully addressed your comments and provided detailed responses and clarifications to each point. We would be grateful if you could review our revised responses and reconsider your assessment of the manuscript in light of the additional explanations and revisions. Please let us know if you have any further questions or concerns.

Reviewer_FfnQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces HydroWrench as a dataset and benchmark with a large amount of experimentally measured command-to-six-axis-interface-load trajectories. It covers three representative bioinspired aquatic mechanisms: a 1-DOF caudal fin, a 2-DOF webbed foot, and a 3-DOF flipper, comprising a total of 116,756 trajectories.

Based on this dataset, the authors establish a benchmark for evaluating nonlinear representations and motion-history modeling, with models trained separately for each mechanism. The experiments compare various linear and nonlinear models, investigate the effects of additional motion history, and evaluate waveform reconstruction, temporal scoring sensitivity, and generalization to held-out motion conditions. These results may provide empirical guidance for developing and evaluating dynamic surrogate models.

Critical strengths

This paper shows two critical strengths for me:

  1. This dataset covers representative bioinspired aquatic mechanisms with command-to-six-axis-interface-load trajectories. Different from dataset generated from simulators, it provide real responses from a coupled, partially observed system with actuator dynamics, mechanical transmission, and evolving flow connect the recorded command to the measured interface load. With a large amount of empirically measured trajectories, it may provide useful base for some related training and system identification tasks.

  2. The authors also demonstrate how this dataset can facilitate the analysis of nonlinear representations and motion-history modeling and provide related evaluations.

Critical weaknesses

Despite the strengths mentioned above, I still have two main concerns:

  1. The coverage of motion and physical operating conditions is not sufficiently clear to me. I appreciate that the authors acknowledge the limitation of using a fixed towing speed. However, based on my background and hands-on research experience in fluid dynamics, I think both flow conditions and motion settings can strongly affect the resulting loads and system responses. I would like to see a clearer picture of how the 116,756 trajectories are distributed across different motion settings, such as amplitudes, frequencies, and phase differences. It would also be helpful to report the relevant dimensionless flow parameters, such as Reynolds and Strouhal numbers. Without this information, it is difficult for me to judge how representative the dataset is and what range of physical conditions it actually covers, which I think is important when evaluating a dataset like this.

  2. I am not yet fully convinced of the unique contribution of this benchmark. I think the related work could include a more direct comparison with existing experimental motion-to-load prediction studies (e.g., Bai et al., 2022; Han et al., 2025). Although these works are not exactly the same, I would like to understand more clearly what HydroWrench adds and what new research questions it enables. Also, some of the findings, such as nonlinear models performing better than linear baselines, are somewhat expected. While the paper provides some interesting observations, I am still unsure about its broader contribution to the machine learning community.

I would be happy to reconsider my assessment if the authors could clarify these points, especially the dataset coverage and its contribution compared with existing work.

Questions for authors
  1. How many distinct motion settings are represented in the 116,756 trajectories? Could the authors show the distributions of relevant parameters (e.g., amplitude, frequency, and phase)?

  2. What ranges of Reynolds and Strouhal numbers does the dataset cover, and how are they defined?

  3. What does HydroWrench offer beyond existing datasets and benchmarks, and what new research questions does it enable given its current coverage?

1 comment
AuthorsOpenReview ↗

Looking forward to your feedback

We sincerely appreciate your time and consideration of our manuscript. We have carefully addressed your comments and provided detailed responses and clarifications to each point. We would be grateful if you could review our revised responses and reconsider your assessment of the manuscript in light of the additional explanations and revisions. Please let us know if you have any further questions or concerns.

Reviewer_Q3bEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces HydroWrench, an experimentally collected dataset and benchmark for reconstructing dynamic loads in bioinspired aquatic mechanisms. The dataset contains 116,756 command–load trajectories collected from three physical mechanisms with one, two, and three degrees of freedom, respectively. The inputs consist of commanded joint angles and their derived velocities and accelerations, while the outputs are measured six-axis interface loads. The authors benchmark linear regression, MLPs, recurrent and convolutional models, and several modern sequence architectures. They investigate the effects of nonlinear representations, temporal history, model architecture, training-data quantity, and high-frequency holdouts. Their results show that nonlinear predictors outperform matched-feature linear baselines across the three mechanisms, while the benefits of longer temporal histories vary by mechanism. The paper also provides analyses of dense-grid waveform errors and frequency-based generalization.

Critical strengths
  1. Valuable experimental dataset. The dataset covers three physically distinct aquatic mechanisms and provides measured six-axis loads rather than relying exclusively on simulation. It could be useful for research on data-driven modeling of bioinspired mechanisms.

  2. Broad benchmark and diagnostic evaluation. The authors compare conventional and modern sequence models and investigate nonlinear representations, temporal history, and waveform-level errors. The matched-feature Ridge–MLP comparison and fixed-architecture LSTM history ablation are particularly useful.

Critical weaknesses
  1. Unclear target audience, use cases, and contribution relative to prior work.

    The paper does not sufficiently establish who the primary users of HydroWrench are or what concrete research questions they can investigate with it. Potential audiences include hydrodynamics researchers studying load prediction, aquatic-robotics researchers developing dynamic models, and machine-learning researchers studying time-series reconstruction. However, the paper does not clearly explain what each group gains from this dataset or why it would be useful beyond existing experimental datasets and learned surrogates.

    Closely related work already exists. For example, Hamamatsu et al. (2025) use a learned force surrogate for soft-fin robot control; Lee et al. (AAAI 2023) use learned thrust prediction for flapping-fin motion optimization and control; and Bai et al. (2022) predict experimental six-axis hydrodynamic loads for a manta-ray robot. These studies differ in scope and do not provide the same multi-mechanism benchmark, but they demonstrate that experimental force prediction and downstream use of learned hydrodynamic models are not new. HydroWrench briefly mentions Hamamatsu et al. but does not adequately compare its contribution with this broader line of research.

    The authors should clarify whether the main contribution is dataset scale, mechanism diversity, a unified evaluation protocol, or a new modeling capability. More importantly, they should explain who is expected to use the benchmark, for what tasks, and what it enables that existing resources do not.

  2. Limited generalization and downstream validation.

    Although three mechanisms are included, models are trained and evaluated separately, and all experiments use a fixed towing speed of 0.167 m/s. The high-frequency holdout is useful but is limited to the webbed-foot mechanism. The results therefore do not establish cross-mechanism or broader operating-condition generalization.

    Furthermore, the paper motivates the dataset through applications such as motion optimization and control, but evaluates only offline command-to-load reconstruction. It does not demonstrate how the benchmark improves a downstream robotics task or provides a practical advantage over existing learned surrogates.

  3. Limitations in physical interpretation and online applicability.

    The inputs are commanded joint trajectories, not independently measured executed motion, and the flow state is not observed. Thus, the learned mapping captures command-history-to-load relationships but does not isolate the contributions of actuator execution, mechanism dynamics, and fluid interactions. In addition, velocity and acceleration are computed using a centered Savitzky–Golay estimator that uses future samples. This is reasonable for offline reconstruction but does not directly support causal online inference.

  4. Data curation and benchmark protocol need further validation.

    The appendix reports timestamp gaps of up to 0.174 seconds in webbed-foot records, bridged through linear interpolation. Given the emphasis on waveform reconstruction, a sensitivity analysis excluding records with large gaps would strengthen confidence in the results. The authors should also clarify whether test records influenced the development of the amplitude-screening thresholds. Finally, equal epoch budgets do not necessarily provide comparable optimization budgets across architectures, while the fixed-epoch data-scaling experiment gives larger datasets more optimizer updates. These factors limit the strength of the corresponding comparisons.

  5. Insufficient description of the physical experimental setup and data acquisition.

    Although the paper states that the dataset comes from physical experiments on the REEF platform, the description of the actual experimental setup is limited. Section 3.1 provides the towing speed and identifies the use of parallel mechanism lanes and six-axis sensing, while Figure 1 is adapted from the REEF paper rather than showing the apparatus used for the reported acquisitions. Appendix A provides useful mechanism dimensions, but important experimental details remain unclear, including the towing-tank and lane dimensions, water depth, mechanism mounting and positioning, load-cell specifications and calibration, and how the flow conditions are established and verified. These details are important for assessing measurement reliability, understanding the physical conditions represented by the dataset, and reproducing the experiments. The authors should provide a clearer description and, ideally, photographs of the actual experimental apparatus.

References

  1. Y. Hamamatsu et al. "Underwater Soft Fin Flapping Motion with Deep Neural Network Based Surrogate Model." IEEE International Conference on Soft Robotics (RoboSoft), 2025. https://doi.org/10.1109/ROBOSOFT63089.2025.11020964

  2. J. Lee et al. "Data-Driven Machine Learning Models for a Multi-Objective Flapping Fin Unmanned Underwater Vehicle Control System." Proceedings of AAAI, 37(13), 15703–15709, 2023. https://doi.org/10.1609/aaai.v37i13.26863

  3. J. Bai et al. "Data-Driven Prediction of Experimental Hydrodynamic Data of the Manta Ray Robot Using Deep Learning Method." Journal of Marine Science and Engineering, 10(9), 1285, 2022. https://doi.org/10.3390/jmse10091285

Questions for authors

Questions for Authors

  1. Who are the intended primary users of HydroWrench, and what concrete research tasks or questions does it enable? How does it differ in practical value from existing experimental hydrodynamic datasets and learned surrogate models?
  2. How does the benchmark advance beyond Hamamatsu et al. (2025), Lee et al. (AAAI 2023), and Bai et al. (2022), particularly regarding dataset diversity, evaluation capabilities, and downstream robotics applications?
  3. Can the authors provide evidence of generalization beyond the fixed towing speed and separately trained mechanisms, or clarify the intended scope of the benchmark?
  4. How sensitive are the waveform-level results to excluding records with large timestamp gaps? Can the authors also clarify whether test-set records influenced the amplitude-screening procedure?
  5. How would the architecture rankings and data-scaling conclusions change under compute-matched training budgets?
  6. Could the authors provide photographs and a more detailed description of the actual experimental setup, including the towing facility and lane dimensions, water depth, mechanism mounting, six-axis sensor specifications and calibration, and flow-condition verification? Which of these details are available in the cited REEF platform paper, and which are specific to the acquisitions reported here?
1 comment
AuthorsOpenReview ↗

Looking forward to your feedback

We sincerely appreciate your time and consideration of our manuscript. We have carefully addressed your comments and provided detailed responses and clarifications to each point. We would be grateful if you could review our revised responses and reconsider your assessment of the manuscript in light of the additional explanations and revisions. Please let us know if you have any further questions or concerns.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.