- 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?
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.