- Rating
- 3
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
The paper investigates the problems and failure modes associated with physical parameter identification from videos. It focuses specifically on a class of nonlinear scalar dynamical systems, as opposed to previous work which focused on LTIs. The paper establishes a framework for determining parameter identifiability and diagnoses which ODE families present uniquely identifiable parameters, and which parameter identification questions may be ill-posed. As a solution, the authors present the coordinate ambiguity, invariants and necessary parameter calibrations for different ODE families. The theoretical predictions are tested on synthetic data spanning 6 ODE families and 3 real-world datasets.
Critical strengths
I rank the paper highly in terms of significance and timeliness. Physical parameter estimation is an important problem in many fields of science and engineering, and while there are many emerging methods to solve this problem, as described by the authors, many of the existing methods need to implement manual tricks or specific normalization schemes in order to deal with the intrinsic invariances, symmetries, coordinate ambiguities in the specific dynamical system/governing equation they are considering. This both wastes time and resources for researchers and often makes it difficult to compare different methods (as they may employ different tricks or normalizations). This paper establishes a generalized framework for identifying a priori what parameters can be uniquely identified from the data, or what calibrations/initial conditions are necessary, to start from a well-posed inverse problem. I believe Table I can already be very useful for many natural scientists as well as AI researchers who are tackling these inverse problems, by clarifying what ground truths, dataset calibrations and initial conditions are necessary. It can also make new methods more standardized and comparable.
The analysis of different ODE families in the paper seems to follow sufficient rigor. The theoretical contributions are meaningful and are confirmed by the experiments on real and synthetic data.
Critical weaknesses
Real-world datasets: I found the recording of yet new datasets for this problem slightly confusing. There are already multiple established datasets for physical parameter estimation in the literature: for example many pendulum datasets are available which could provide more baseline comparisons, and I believe the Delfys dataset also contains free-fall videos. Tests on existing datasets would strengthen the results by providing existing baselines.
Lack of baselines: The empirical methods are explained in a manner which lacks crucial details about the parameter estimation method, the specifics of the shared encoder, loss function and optimizer used, etc. in the main paper. However, my understanding is that the same encoder is used for all empirical results. The accuracy/stability/reliability of this encoder and the results gained from it are difficult to interpret or trust in the absence of baseline comparisons. Would it be possible to show how the theoretical understanding gained from this paper empirically improves upon previous baseline methods?
Presentation: The clarity of the language is good, however, I found the methods/implementation details described in Section 5 extremely difficult to follow. I also found the explanation of the empirical results and the captions in Section 5 vague and confusing. What is shown in the tables/figures (what are shown on the axes, the legend, what settings are being compared, etc.) need to be clarified in the text.
Interpretation of the results: Section 5 text lists the empirical results. However, it is a bit unclear what the take-home message is from these results. What were the hypotheses for these experiments, what did the authors expect to find? What are the conclusions, important observations? It would be nice to detail if/how/why the experimental observations confirm the theoretical predictions.
Reproducibility: Will the code for the empirical experiments be made available? Sharing the code would make the methods more clear and ease any reproducibility concerns. The reproducibility statement mentions that the code for the synthetic data will be available, but what about the real-world experiments? Will the datasets and the evaluation code be made publicly available?
Generalizability: What about other types of dynamical systems, families of ODEs? How well can the presented framework generalize to non-scalar/coupled systems? Is it possible to say a few words about it in the main text and not just Appendix K?
Questions for authors
-
Why not use existing datasets instead of collecting yet new datasets with no baseline comparisons?
-
What is the parameter estimation model which was used? Is the encoder the same one as used in the IRIS paper? If yes, can the authors state this explicitly?
-
Why are other baseline parameter estimation methods not used/compared against? Would it be possible to show how the theoretical understanding gained from this paper empirically improves upon previous baseline methods?
-
Is it possible to better explain/clarify the methods used in Section 5 in the main text (i.e. not just in the appendix, as it is currently almost impossible to understand the basics of the methods only from the main text)? The tables/figures (what are shown on the axes, the legend, what settings are being compared, etc.) also need to be clarified in the text.
-
What are the take-home messages of Section 5? What were the hypotheses for these experiments, what are the conclusions, important observations? It would be nice to detail if/how/why the experimental observations confirm the theoretical predictions.
-
Will the datasets and the evaluation code be made publicly available for the real-world experiments?
-
Is it possible to say a few words about the generalizability/possible extensions of the framework in the main text and not just Appendix K?
5 comments
Author Response (Part 2/3)
W2. Ablation study with IRIS baselines: what enables accurate recovery?
A. Thank you for this question. To understand which factors contribute to the recovery differences between our method and the IRIS baselines, we conduct an ablation study along three dimensions: encoder architecture, training objective, and data conditions. The complete study is reported in Appendix P.2 (Page 66).
(1) Encoder architecture. We first examine whether recovery depends on our particular CNN. Replacing it with the IRIS MLP encoder, while retaining our nonlinear objective, preprocessing, and multi-angle fitting, yields a length estimate of m and an RMSE of m for the m IRIS pendulum. Our CNN yields m and an RMSE of m. These results demonstrate compatibility with the IRIS encoder.
(2) Training objective. We retain our CNN, preprocessing, and multi-angle data, and adopt the IRIS finite-difference prediction losses and optimizer settings. With KL regularization, the Corrected and Multi-step variants yield m and m, respectively, for the m pendulum. Replacing KL with our variance-floor regularizer gives m and m. Our original finite-difference ODE-residual loss with a variance floor yields m (Appendix P.2, Table 29, Page 67). Considered alongside the original IRIS baseline results in Table 27 (Page 66), these results demonstrate that useful recovery is compatible with different dynamical losses and regularizers, although their choice still affects estimation accuracy.
(3) Data conditions. Finally, we examine the role of trajectory coverage. On individual IRIS recordings, the tested first-ten-frame segments do not provide three-slope coverage, and our method also fails under this restriction. Using 100 frames (10 s) provides richer state–velocity coverage and reduces our RMSE to , , and m at , , and , respectively (Table 30, Page 68). The adapted IRIS objectives also improve at and , while retaining large errors at . One possible explanation is that weaker nonlinear excitation at the smaller angle increases sensitivity to the KL regularizer's preference for unit latent variance; this mechanism has not been independently verified.
Taken together, these ablations show that successful recovery is not specific to our encoder or loss, and support sufficient trajectory coverage as an important contributing factor. This is consistent with the theory's guidance on data collection, while the remaining performance differences show that coverage alone does not ensure state consistency or successful optimization.
Q3.2. How does the theoretical understanding guide practical parameter estimation?
A. We thank the reviewer for this question. Our theory provides practical guidance on what parameter information is identifiable and what data conditions support its recovery. As illustrated by the ablations in our W2 response, improving trajectory coverage can substantially improve adapted IRIS baseline estimates without changing their encoder architecture or training objective. Under the stated assumptions, the theory also distinguishes parameters identifiable directly from passive video from those requiring additional physical calibration. For the squared-velocity setting, Theorem 4.2 additionally provides a law-derived normalization that permits affine comparison in canonical coordinates. For further discussion, we respectfully refer the reviewer to our response to Reviewer raqQ, W2 & Q1. Together, these results guide the design and interpretation of practical estimators by specifying what data are needed, which quantities are identifiable, and when normalization or physical calibration is required.
W4. Could Section 5 explain the implementation details, experimental settings, and figures and tables more clearly?
A. We thank the reviewer for this helpful suggestion. We have expanded Section 5 to clarify the experimental settings and the interpretation of the figures and tables. The opening now describes joint encoder–dynamics fitting, explicitly references the training objective in Eq. (3), and identifies the encoder architecture. It also directs readers to Appendices M and N for architectures, data preparation, optimizers, and training settings, and to Appendix L for metric definitions. Throughout the section, we explain the comparisons and findings more explicitly and connect them to the corresponding theoretical predictions. These additions are highlighted in red in the revised manuscript.
Author Response (Part 3/3)
Q5. What are the take-home messages of Section 5, and how do the experiments support the theoretical predictions?
A. We thank the reviewer for this suggestion. The revised Section 5 makes the purpose and main finding of each experiment explicit:
(1) Coordinate ambiguity and parameter information (Section 5.1). Figure 1 tests the family-specific predictions of Section 3. The learned coordinates follow the predicted transformations; invariant quantities remain approximately stable, while scale-dependent coefficients follow their parameter orbits. These observations support the coordinate relations and identifiable parameter information summarized in Table 1.
(2) Velocity coverage (Section 5.2). Table 2 tests the rank condition in Section 4.1. Coordinate errors are similar across the tested designs, but parameter recovery improves substantially under full-rank velocity coverage. Accurate coordinate fitting alone therefore does not ensure accurate parameter recovery: the full-rank design separates the declared velocity features from coordinate curvature.
(3) Canonical normalization (Section 5.2). Table 3 tests Section 4.2 using the same fitted model before and after law-derived normalization. Normalization reduces coordinate error and improves recovery of the reference restoring force, supporting the predicted affine relation in canonical coordinates when raw affine alignment is insufficient.
(4) Real-video recovery (Section 5.3). Figure 2 and Table 4 examine these predictions on real pendulum and free-fall recordings. The observed invariants, parameter orbits, and normalized coordinate maps provide empirical support under real imaging conditions. The calibrated length and gravity estimates further illustrate how external anchors resolve the remaining physical ambiguity.
W5 & Q6. Will the code and real-world datasets be publicly available?
A. We thank the reviewer for this question. The supplementary material already includes data-generation and training code for all synthetic experiments, together with the identifiability tool described in our response to S1. Upon acceptance, we will publicly release the complete codebase and all self-recorded video datasets.
Q7. Could the main text discuss generalizability and possible extensions beyond the scalar setting?
A. Thank you for this suggestion. We have expanded the discussion in Section 6 to clarify the scope and possible extensions of our framework. Our results cover the stated scalar autonomous ODE families. In higher dimensions, finitely many smooth trajectories leave coordinate freedom outside the observed trajectories, so the scalar argument does not extend directly. We outline two possible directions: restricting the admissible coordinate transformations and considering trajectory ensembles with richer position–velocity coverage. Appendix K.2 provides the corresponding conditional vector extension and discusses the remaining challenges.
References.
- Wang, Y., Wang, W., Zhang, K. and Gong, M. (2026). Physics from video: Identifiability of time-invariant second-order ODEs under minimal trajectory conditions. ICML.
Author Response (Part 1/3)
We sincerely thank the reviewer for the positive assessment of our paper and for recognizing its significance, theoretical rigor, and practical value. Below, we address the concerns and questions raised in turn.
A. We appreciate the recognition of Table 1's potential value to researchers working on physical parameter estimation across different fields. This reflects our motivation: to clarify what the data can identify and what additional calibration is needed. Table 1 summarizes representative dynamical families; our analysis covers broader classes of nonlinear scalar second-order ODEs under the stated assumptions. To support applications beyond these examples, we provided a lightweight tool in the supplementary material (Appendix O, Pages 63–65). For a user-declared ODE family within its supported scope, the tool reports the required data conditions, identifiable parameter invariants, and calibration requirements.
A. Thank you for the question. Among our three real-world experiments, the pendulum experiment directly uses the public IRIS dataset, while the side-view and overhead free-fall experiments use our own recordings. As explained at the beginning of Appendix N (Page 50), although IRIS provides similar free-fall videos, these datasets do not satisfy the three-slope divergence and state-consistency conditions required by our theory. For the side-view setting, IRIS uses different balls released from the same height. This provides insufficient variation in release conditions for the three-slope divergence condition, while differences in ball appearance mean that the same physical position may correspond to different visual observations. For the overhead setting, although IRIS uses the same ball, the camera configuration varies across recordings, as illustrated in Fig. 12 (Page 51). Consequently, the same physical state may produce different visual observations, violating the shared state-to-image mapping assumption. We therefore recorded our own free-fall datasets, keeping the ball and camera configuration fixed within each view while varying the release height. This provides a controlled setting for evaluating our theoretical assumptions and complements our experiments on the public IRIS pendulum dataset.
A. We agree that the encoder architecture and its provenance should be stated more explicitly. For the synthetic and real-pendulum experiments, we adopt the encoder architecture from LTI (Wang et al., ICML 2026), rather than the IRIS encoder, retaining the same network configuration except for replacing BatchNorm with GroupNorm. GroupNorm computes normalization statistics independently for each frame, avoiding dependence on other frames in the training batch and consistent with the framewise encoder assumed by the theory. For the side-view and overhead real free-fall experiments, we use similar CNN architectures with configurations adapted to their respective input resolutions. To improve reproducibility, we have added detailed encoder architecture tables in the appendices (Page 42, 52 and 56). We have also added the following clarification at the beginning of Section 5 (Page 7) of the revised manuscript:
The synthetic experiments and real-pendulum experiments use the CNN architecture of the LTI method, replacing BatchNorm with GroupNorm to avoid dependence on other frames in the training batch. Side-view and overhead free fall use similar CNNs with configurations adapted to their inputs. Implementation details are in Appendices M and N.
A. Thank you for this suggestion. We have added pendulum-length comparisons on Delfys75 and IRIS in Appendix P.1 (Page 65). The tables directly include published results from LTI (Wang et al, ICML 2026) and statistics computed from released IRIS result, retaining their respective evaluation protocols. We additionally compare LTI and nonlinear models using matched training configuration. The nonlinear model reduces RMSE from 0.0583, 0.0762, and 0.1034 m to 0.0495, 0.0565 and 0.0757 m on Delfys75, and from 0.0646 to 0.0150 m on IRIS. These matched comparisons support the benefit of retaining the nonlinear pendulum restoring term. We also clarify the reference-length convention. Under the point-mass approximation, g/κ represents the pivot-to-centre-of-mass distance, which may differ from the reported string length. The measurement endpoints are not explicit in the documentation examined, so this possible source of systematic offset remains unverified. We retain the published reference lengths and apply no post-hoc correction.