acceptodds
Under review as a conference paper at ICLR 2027

DynaBeam: From Motion-Driven Parameter Identification to Real-Time Physical Simulation of 3D Vegetation

Abstract

Our 3D world is filled with vegetation that is often in motion due to interactions with wind and other physical forces. While recent representations excel at capturing and generating these dynamic scenes, extracting their underlying physics traditionally requires computationally heavy inverse simulations or appearance-based learning that struggle with speed and generalization. In this work, we introduce DynaBeam, a framework that models dynamic 3D vegetation, ranging from houseplants to trees, as hierarchical graphs of connected beam elements. We show that the geometric structure and motion of these beams provide a strong enough signal to efficiently extract both the intrinsic physical parameters like stiffness and the external forces driving the movement. By explicitly separating geometry, motion, and force, DynaBeam bypasses expensive simulation solvers to achieve accurate dynamic scene reconstruction in mere seconds. Furthermore, because we recover physical parameters rather than fitting visual patterns, our representation natively supports real-time physical simulation under entirely new and unobserved forces. Extensive evaluations show that DynaBeam delivers faster and more accurate physical inversion and forward simulation than prior methods, enabling interactive animation and physical reconstruction from real-world videos.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Physical Simulation, Dynamic Reconstruction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_mQcYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DynaBeam models vegetation as connected Euler–Bernoulli beam structures, builds reference modes from geometry, and then estimates spectral scaling, a shared damping ratio, and density-normalized external forces from observed motion. Its key design is to separate the structure-determined modal space from the observation-determined dynamics parameters, avoiding repeated forward simulations during identification and enabling modes that are not sufficiently excited in the input to respond to new external forces. The paper evaluates parameter estimation, resimulation, and new external force prediction using synthetic FEM data, and demonstrates a qualitative interactive application with models reconstructed from real videos.

Critical strengths
  1. The beam-based representation is well suited to branching vegetation and makes physical parameter recovery much more efficient by avoiding repeated forward simulations. The reported recovery time of around 2 seconds per tree is a meaningful practical improvement.

  2. The generalization experiments under unseen forces are convincing. By deriving vibration modes from geometry rather than observed motion, DynaBeam can reproduce responses involving previously unexcited modes. The results against FEM ground truth show a substantial improvement over the motion-based baseline.

Critical weaknesses
  1. The contribution of motion-based material estimation is not sufficiently isolated. The five tree species have relatively similar baseline stiffness-to-density ratios, but the paper does not include a simple constant-material baseline. It is therefore unclear how much of the parameter recovery accuracy comes from observed motion rather than the material prior.

  2. The validation of recovered physical parameters is incomplete. The paper does not directly report damping or force estimation errors. Moreover, Section 3.5 and Appendix D.6 appear to use different force formulations for re-simulation. If the recovered modal residuals are used directly, accurate training reconstruction provides limited independent evidence of physical parameter accuracy.

Questions for authors
  1. Could the authors report the ground-truth distribution of stiffness-to-density ratios across the 72 trees and compare against a constant-material baseline using the same beam model, both for parameter recovery and novel-force simulation?

  2. Does the reconstruction experiment in Table 1 use the spatially constrained force projected onto the modal basis, or the directly recovered modal forces? Are the held-out joints excluded from all parameter estimation steps?

  3. Could the authors report the estimation errors for damping and external forces? In particular, how accurately can the recovered model predict free-decay motion when the external force is set to zero?

5 comments
AuthorsOpenReview ↗

W1/Q1: Motion-based material recovery vs. a constant-material prior

We thank the reviewer for the precise questions. Each is answered below with a new experiment or a direct clarification.

Summary. DynaBeam recovers material properties directly from observed motion and generalizes across substantial material variations and novel force directions, as demonstrated by three additional experiments. We drew materials from real wood species so that VoMP is evaluated in its intended regime. On this narrow range (1.32×), a constant based on the nominal species mean achieves lower ALDE (0.046) than DynaBeam (0.152). However, this constant becomes substantially less accurate as the material range expands, while DynaBeam maintains accurate material recovery and predicts novel-force motion more reliably. These experiments demonstrate the advantages of motion-based recovery beyond the original benchmark.

Ground-truth distribution. Across the 72 trees, E/ρE/\rho ranges from 7.13×1067.13\times10^6 to 9.42×1069.42\times10^6 m²/s². The five-species constant is their nominal mean, 8.531×1068.531\times10^6 m²/s².

1. Generalization to a novel force direction on the original 72 trees. All methods use the same beam model with ground-truth damping and force amplitude, so only E/ρ differs. The trees are driven along +Z, whereas training used +X, and no method is fitted to these trajectories.

E/ρ used in the beam modelMedian novel-force Rel-L2 [Q1, Q3] ↓
DynaBeam (Table 1 estimates)0.330 [0.205, 0.442]
Five-species constant0.570 [0.358, 0.674]

DynaBeam achieves lower error than the five-species constant on 56/72 trees. ALDE measures agreement with the FEM's average material, while the beam model needs the effective E/ρ that reproduces the dynamics in its reduced representation. DynaBeam recovers the latter from motion, which explains its better dynamic predictions even when its ALDE is higher.

2. Material scaling (360 cases). We scaled E by α∈{0.25,0.5,1,2,4}\alpha\in\{0.25,0.5,1,2,4\} on all 72 geometries, widening the tested E/ρE/\rho range by approximately 21×; the scaling factor is never given to the method. We compare DynaBeam against a fixed five-species prior E/ρ=8.531×106E/\rho =8.531\times10^6, using the same constant across all geometries and scales, without access to the ground-truth material values. The median within-tree log-log slope between recovered and true E/ρE/\rho is 0.997, indicating that DynaBeam closely tracks the imposed material changes. Across all 360 cases, DynaBeam achieves a median ALDE of 0.0745 (mean 0.096), compared with 0.3345 (mean 0.371) for the fixed prior, outperforming it in 304/360 cases.

Under the novel +Z force, DynaBeam achieves a median held-out Rel-L2 of 0.268 [0.161, 0.417], compared with 0.913 [0.786, 1.801] for the fixed prior, outperforming it in 347/360 cases. Its median cosine similarity is 0.970 (vs. 0.652), and its dominant-frequency error is only 1.9% (vs. 43.2%).

3. Six additional wood species (120 cases). We assigned balsam poplar, water tupelo, Port-Orford cedar, noble fir, black spruce and coast Douglas-fir to 20 existing geometries, following App. F, and regenerated the FEM ground truth. The nominal E/ρE/\rho range becomes 6.60–14.71×10⁶ m²/s².

Six new species, 120 casesMean ALDE ↓Median ALDE ↓
DynaBeam0.0670.065
Eleven-species constant (includes the test species' values)0.1510.147
Five-species constant0.1750.198

DynaBeam beats the eleven-species constant in 116/120 cases and on all 20 geometries (paired mean difference −0.083, 95% CI [−0.089, −0.077], resampling geometries). Its error barely changes from the original materials to the new ones on the same geometries (mean ALDE 0.070 → 0.067), while the five-species constant's error grows 4.1× (0.043 → 0.175). The recovered values also preserve material differences: log-log slope 0.965, within-tree Spearman correlation at least 0.94, and a high-to-low ratio of 2.04 versus 2.07 in the ground truth.

Scope. This experiment isolates material-magnitude variation across species while keeping the spatial noise pattern fixed within each geometry.

AuthorsOpenReview ↗

W2/Q2: Direct validation of force recovery and re-simulation

Summary. We directly evaluate damping and force recovery over 72 trees, obtaining median relative errors of 2.0% and 8.5%, respectively, without ground-truth amplitude calibration. Unlike the baselines, which receive ground-truth force information for re-simulation, DynaBeam estimates the force amplitude and damping from observed motion. Table 1 uses a spatially constrained, mass-proportional force along X with a prescribed pulse profile; only its amplitude is unknown and recovered by least squares, while force-direction recovery is demonstrated separately (JdPW, W2). Held-out joints are excluded from all estimation steps and used only for evaluation, with their static geometry retained for beam construction.

Direct errors. Over 72 trees, the median relative damping error is 2.0% and the median relative force-amplitude error is 8.5%, with no ground-truth amplitude calibration.

Which force drives re-simulation. Force recovery has two steps (App. D.4). First, the modal force needed to explain the fitted motion, Fmodal(t)=q¨(t)+2ζΩq˙(t)+Ω2q(t),F_{\mathrm{modal}}(t)=\ddot q(t)+2\zeta\Omega\dot q(t)+\Omega^2q(t), is computed as an intermediate quantity. Second, a physical force F(t)=Va^(t)F(t)=V\hat a(t) is fitted to it, with a^(t)\hat a(t) minimizing ∣ΦTVa−Fmodal(t)∣2.|\Phi^T V a-F_{\mathrm{modal}}(t)|^2. For the Table 1 experiments, forward simulation is driven by ΦTVa^(t)\Phi^T V\hat a(t), the projection of this constrained physical force. Since the force profile is known, the entire re-simulated trajectory must emerge from the recovered E/ρE/\rho, ζ\zeta, and the single unknown amplitude. In App. D.6, we now explicitly distinguish this projection from the intermediate modal force.

Force formulation. The general formulation (App. D.4) supports combinations of spatial force patterns with time-varying coefficients, recovered through least squares. When no low-dimensional force prior is available, as for unknown forces on real plants, we instead use unconstrained, time-varying modal forces, recovered directly from the observed modal coordinates. This alternative is used for the real plants and the continuous-excitation case study (JdPW, W1). This distinction between the two formulations is now explicitly described in Sec. 3.4 and App. D.4.

Held-out joints. The 80/20 split is applied before any estimation. Only training joints are used for spectral analysis and material recovery, modal-coordinate fitting, damping and force-amplitude estimation. Held-out trajectories are used exclusively for evaluation. Thus, the held-out metrics measure generalization to unseen joints under the observed excitation; prediction over time is evaluated in Q3, and prediction under novel forces in Table 2.

AuthorsOpenReview ↗

Q3: Free decay with zero external force

Summary. After the pulse, the force is set to exactly zero in both the FEM reference and our simulation, so the free decay is predicted from the recovered stiffness and damping alone. DynaBeam's held-out free-decay error is 7.6× lower than VoMP's and 33× lower than PhysFlow's, and nearly identical on training and held-out joints. Only the oracle PhysFlow†, which is given the ground-truth density, force and damping, is lower.

MethodTreesMean training Rel-L2 ↓Mean held-out Rel-L2 ↓
DynaBeam720.5250.540
VoMP723.854.10
PhysFlow4517.417.9
PhysFlow† (oracle)450.1240.126

Errors are higher than over the full trajectory for every method. During decay the displacement amplitude shrinks, so small frequency mismatches accumulate into phase differences that relative error penalizes heavily.

Reviewer_AicHOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work propose a method that can recover physical parameters and driving force of a 3D vegetation dynamics. Without using RGB supervision to optimize the target parameters differentiablely, authors directly infer the physical conditions based on classical dynamic formula and Euler–Bernoulli Beam Structures. On their own benchmark, Dynabeam can recover the dynamic motion and physical parameters with higher accuracy. Meanwhile, this method is capable of predicting motions under unseen external forces.

Critical strengths
  1. The inference of physical properties and driving force directly from the dynamic modality is accurate and efficient.
Critical weaknesses
  1. The paper seems to combine established Euler–Bernoulli beam modeling, modal analysis, spectral matching, and least-squares force estimation. I doubt its methodological advances. Authors may comparie with classical modal identification approaches to clarify the work's contribution.
  2. There is no qualitative comparison with the baseline methods. Although the authors provide 3 results in the supp materials, clearly the reconstructed geometry in the 1st and 3rd demo are terrible -- either has obvious artifacts or the shape of the vase is wrong. These make me concern the reliability of the following prediction method.
Questions for authors
  1. See the weakness.
  2. Section4.2 and Appendix D.5 use the known training force to estimate a global calibration factor. How does novel-force prediction perform without this calibration, and how to obtained this in real-world scenarios where the original force is unknown?
  3. What does the 2.1s in table 1 cover? What's the running time of the reconstruction including video -to-3d and skeleton extraction? What's the reconstruction error and its effect on following recovery of physical properties?
5 comments
AuthorsOpenReview ↗

W1: Relation to classical modal identification

We thank the reviewer for the detailed questions. Each is answered below, with a new experiment where relevant.

Summary. DynaBeam deliberately builds on established structural-dynamics tools [1–7]. Our contribution is making them work from 3D vision data, where no hand-built structural model exists, and deriving a modal basis from geometry that responds to modes never excited in the input. New comparisons support both points: a classical ERA pipeline recovers material less accurately, and classical model updating (C3) matches our accuracy only through iterative optimization.

What differs from the classical setting.

  1. The structural model is built automatically from a 3D reconstruction, including topology and branch radii, with manual refinement in 19 of 72 trees (about 1–2 min each). Classical model updating starts from a hand-built FE model.
  2. Output-only identification (ERA, SSI, FDD) recovers only the modes excited in the measurement, and yields no predictive model for the others. This is the limitation we demonstrate for ModalNeRF in Table 2 and Figure 3; our geometry-derived basis retains those modes.
  3. The beam structure lets Proposition 2 reduce model updating to one closed-form spectral scale fit, with no iterative updates or repeated simulation.

Comparisons with classical methods.

ComparisonClassical methodDynaBeam
ERA identification + frequency-based updating (C1), 72 trees, median ALDE ↓0.0870.070
FDD vs. our FFT detector, full sequences (10 trees, 19 excited modes), precision / recall ↑0.12 / 0.161.00 / 0.53
FDD vs. our FFT detector, post-excitation only, precision / recall ↑0.77 / 0.531.00 / 0.58
FE model updating (C3) on our beam model, 10 trees, median ALDE ↓0.1130.098
C3 total runtime per tree ↓3.5–4.9 s2.1 s

Frequency matching uses a 3% tolerance. Material recovery needs only a few reliably matched frequencies, so the FFT detector's perfect precision matters more than its recall. Table 1 reports DynaBeam's mean ALDE (0.152); these comparisons use medians (C3 details in our response to JdPW, W2). We have added these comparisons and a discussion of classical structural dynamics to the revised paper.

AuthorsOpenReview ↗

W2: Qualitative comparison and reconstruction artifacts

Summary. The artifacts come from the video-to-3D reconstruction, not from DynaBeam. We use a pipeline built using state-of-the-art methods, and this is an active research area that continues to improve. Our beam model is built from the annotated skeleton and radii, and the pot enters only as the clamped root. We have added the requested comparison with baselines. In a 57-participant user study, DynaBeam's novel-force responses were preferred in all three comparisons, each involving a choice among three methods (DynaBeam, ModalNeRF, and VoMP), with preference rates of 61–88% (p < 0.001). Its real-world trajectory error is also about 3× lower than that of ModalNeRF and VoMP.

Why the artifacts do not affect the physics. The video-to-3D front end is upstream of DynaBeam and replaceable. The reconstructed surface only sets the depth of the annotated skeleton nodes; it does not define the beam geometry, and the pot is not part of the deformable model. Surface artifacts therefore affect appearance, not the mechanical system.

The pot does set the metric scale. A relative scale error ε changes E/ρE/\rho by about 2ε (App. J), but leaves the reconstructed dynamics unchanged: in our sensitivity analysis, 0.8× and 1.2× scale errors leave the held-out Rel-L2 at 0.267 (Q3 below).

User study. In a user study, 57 participants each evaluated three plant–force configurations. For each, they viewed three anonymized simulations (DynaBeam, ModalNeRF, VoMP) under the same applied force, indicated by an arrow, and selected the response that best matched the expected physical behavior.

MethodComparison 1Comparison 2Comparison 3
DynaBeam61.4%87.7%70.2%
ModalNeRF33.3%7.0%24.6%
VoMP5.3%5.3%5.3%

One-sided binomial tests against chance (33.3%) give p < 0.001 in all three. Method order varied between comparisons but remained fixed across participants within each comparison. Given the rebuttal timeline, this study is small; we will extend it in the final version with more plants and force configurations.

Visual comparisons. Figure 4 already compares DynaBeam and ModalNeRF against the FEM ground truth. We will add side-by-side videos of DynaBeam, ModalNeRF and VoMP under identical novel forces to the supplementary material, together with the quantitative real-world comparison (JdPW, W3).

AuthorsOpenReview ↗

Q2: Novel-force prediction without calibration

Summary. Importantly, DynaBeam outperforms ModalNeRF in novel-force prediction even without cFc_F: for distributed forces, it achieves Rel-L2 of 0.312 vs. 0.999, and for point forces, 0.610 vs. 1.014 (retaining only the geometry-based mass correction). The factor cFc_F compensates for compliance differences between the volumetric FEM ground truth and our reduced beam model, arising from their different discretizations. It is specific to our synthetic evaluation, rather than an inherent requirement of DynaBeam. Without cFc_F, cosine similarity and MDE remain unchanged; only the amplitude error increases. In real use, novel forces are applied directly to the beam model, so cF=1c_F = 1 and no conversion or access to the original force is needed.

Why a conversion is needed at all. The two discretizations respond to the same load with different compliance, and their total masses differ. Such discrepancies between approximate physical models and reference dynamics are well documented [8]. Two corrections handle this, both used only in the synthetic novel-force experiment (Table 2); Table 1 uses no calibration.

  • cFc_F: a single scalar fitted once per tree on the training sequence and frozen for all novel forces (App. D.5).
  • Mass correction for point forces: MFEM/Mbeam=VFEM/VbeamM_{\mathrm{FEM}}/M_{\mathrm{beam}} = V_{\mathrm{FEM}}/V_{\mathrm{beam}}, computed from the two geometries with no material information. Distributed forces are body accelerations, so the mass ratio cancels; a point force is an absolute force, so it does not. Table 2 already includes this correction, which is now explicitly described in App. D.5.

ModalNeRF needs no conversion only because we gave it the ground-truth FEM mass matrix for force projection (App. K.3), information DynaBeam does not receive.

Ablation (22 trees).

Novel distributed forceCosine ↑MDE ↓Rel-L2 ↓
DynaBeam, with cFc_F (Table 2)0.9770.0430.231
DynaBeam, without cFc_F0.9770.0430.312
ModalNeRF (Table 2)0.0560.7840.999
Novel point forceCosine ↑MDE ↓Rel-L2 ↓
DynaBeam, cFc_F + mass correction (Table 2)0.9050.0990.450
DynaBeam, mass correction only0.9050.0990.610
DynaBeam, cFc_F only0.9050.0992.758
DynaBeam, neither0.9050.0993.832
ModalNeRF (Table 2)0.0440.8481.014

The spatial pattern and modal content of the response (cosine, MDE) come entirely from the recovered model; the corrections affect only its amplitude. Without cFc_F, keeping only the material-free mass correction, DynaBeam still has lower Rel-L2 than ModalNeRF for both forces. As for any method using kinematic observations, a force in newtons additionally requires the density (Proposition 1).

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

DynaBeam represents vegetation as connected beam elements and derives vibration modes from the physical structure. From observed geometry and motion, it estimates effective material and damping parameters along with external forces, avoiding repeated expensive forward simulation during inversion. The resulting model supports interactive simulation under new forces, with synthetic quantitative tests and qualitative real-video examples.

Critical strengths
  1. Physically grounded design. Beam mechanics provide an interpretable model of the bending behavior central to plant motion.
  2. Geometry-derived modes can respond to forces that excite motion absent from the observed sequence, unlike methods restricted to observed motion bases.
  3. The paper assesses both physical-parameter recovery and response to new applied forces.
  4. Promising efficiency. Fast parameter fitting and real-time forward simulation are attractive for interactive use.
Critical weaknesses
  1. Identification is evaluated under controlled excitation conditions. Short pulses followed by free decay make frequency and damping estimation easier. Real vegetation often experiences continuous, unknown, and changing wind loads. The robustness and identifiability of the method in such cases remain unclear.
  2. Baseline assumptions are not fully matched. The comparisons include methods that differ in their knowledge of force, damping, or other physical quantities, complicating interpretation of accuracy and runtime differences. A matched optimization-based system-identification baseline would be informative.
  3. Real-world validation is mostly qualitative. The examples show interactive simulations, but physical accuracy under novel measured forces is not quantitatively established on real plants.
Questions for authors

NA

4 comments
AuthorsOpenReview ↗

W1: Continuous, unknown excitation

We thank the reviewer for the positive assessment and the concrete suggestions; each one led to a new experiment.

Summary. DynaBeam does not need a pulse or a free-decay interval. Under continuous random excitation, applied throughout and never given to the method, E/ρE/\rho recovery shows no detectable difference from the pulse protocol, and an output-only damping estimator recovers ζ\zeta within 20% in 59 of 60 runs.

Why we used a pulse. Pulse excitation followed by free decay mirrors standard impact testing in experimental modal analysis, and gives a controlled protocol shared by all baselines. Propositions 1 and 2 do not depend on the excitation; the excitation only determines which modes are observable.

Setup. We excited 10 trees with band-limited Gaussian noise (flat to 140 Hz, at least 80 dB attenuation above 156 Hz) for 11 s, and analyzed the last 10 s. The same signal was applied with two spatial patterns: uniform along X, and localized to an upper-crown subtree holding about 34% of the tree's mass. We replaced the single FFT with Welch-averaged spectra [2] to reduce the variance of a finite random realization; everything else is unchanged.

E/ρ recovery, 10 treesMedian ALDE ↓Median ALRE ↓
Original pulse0.0550.0080
Random, uniform X0.0540.0078
Random, localized X0.0490.0070

Paired Wilcoxon tests show no detectable difference from the pulse (p = 0.92 and 0.36), although 10 trees give limited statistical power.

Damping without free decay. We use an output-only estimator based on the Random Decrement Technique [1], applied to the motion projected onto its dominant spatial pattern. It uses only the observed trajectories, with no access to the force, the ground-truth damping or the FEM modes. Over 60 runs (10 trees × 2 patterns × 3 random signals), it returned an estimate in 60/60 runs and was within ±20% of the ground truth in 59/60, with a median relative error of 5.2%; 53/60 of its 95% confidence intervals contain the ground truth. It takes a median of 6 s.

Force recovery without a prescribed profile. To further demonstrate that DynaBeam is not limited to predefined excitation profiles, we test whether it can recover unknown, time-varying modal forces directly from observed motion. On one tree (oak_05) under continuous excitation, using high-rate observations (3156 fps), we recover per-frame modal forces without assuming any spatial or temporal force profile. The recovered forces reproduce the motion of six held-out joints through forward integration with Rel-L2 0.037, while the dominant modal force correlates with the ground truth at 0.997–0.999, with only 5–11% amplitude error. The estimated damping ratio has a relative error of only 5.8%. These results demonstrate the potential of DynaBeam to recover unknown time-varying modal forces and reproduce the resulting motion without a prescribed excitation profile.

Scope. Our random excitation keeps a fixed spatial pattern, so it does not capture the full spatio-temporal structure of real wind; we state this in the limitations.

AuthorsOpenReview ↗

W2: Matched optimization-based baselines

Summary. DynaBeam requires fewer additional inputs than the baselines in Table 1. Compared with DynaBeam-OPT (C3), which uses the same beam representation and identical inputs but estimates physical parameters through iterative optimization rather than our direct recovery procedure, DynaBeam achieves slightly better material recovery (median ALDE 0.098 vs. 0.113) and identical damping accuracy, without iterative optimization. This also reduces computational cost, requiring only 2.07 s compared with 4.3–4.9 s for the optimization-based approach. While optimization improves trajectory reconstruction, it comes at the cost of additional computation and can be used as an optional refinement stage. Importantly, both approaches benefit from our efficient beam representation, avoiding the much more expensive volumetric inverse simulations required by methods such as PhysFlow (30 min–5 h).

Ground truth received by each method in Table 1.

MethodGround truth receivedRecovered from motion
VoMPDamping and full force (at re-simulation)Nothing (predicted from appearance)
PhysFlowFull force (at re-simulation)E/ρE/\rho
PhysFlow†Density, damping, full force, E search rangeEE
DynaBeamPulse profile onlyE/ρE/\rho, ζ\zeta, force amplitude

Table 1 uses the prescribed-profile variant of DynaBeam: the force direction, spatial pattern and pulse timing are given, as they are to every baseline within the full ground-truth force, and only the amplitude is recovered. E/ρE/\rho is estimated before force recovery and never uses the force; ζ\zeta uses only the pulse end time to select the free-decay window, which W1 shows is not required. Without the force direction, the held-out error on all 72 trees changes by only −0.0006 Rel-L2 (median direction error 10.1°). The general per-frame formulation (App. D.4) is used for the real plants and in the W1 case study. These details are now explicitly stated in Table 1 and Sec. 3.4.

Matched baseline: DynaBeam-OPT (C3). C3 uses the same beam model, observations, simulator and force assumptions as DynaBeam. It jointly optimizes E/ρ, ζ and the force amplitude with L-BFGS-B to minimize the training trajectory error. We initialized it either randomly (E/ρ within 0.1–10× of DynaBeam's estimate, ζ = 0.3, zero force) or from DynaBeam's estimates; both converged to the same solution on all 10 trees.

Median over 10 treesDynaBeamC3, random init.C3, DynaBeam init.
E/ρE/\rho ALDE ↓0.0980.1130.113
Damping log10 error ↓0.0120.0120.012
Force relative L2 ↓0.1520.1540.154
Force-amplitude error ↓0.0850.0640.064
Held-out Rel-L2 ↓0.2740.1580.158
Total runtime, cold (s) ↓2.074.923.48

The C3 runtimes include the DynaBeam step used to set the initialization or search range; without it, random-initialized C3 takes 4.29 s.

Interpretation. Both initializations reach the same solution, so C3 does not depend on DynaBeam's estimate within the tested range. Its closer trajectory fit comes from directly minimizing that error. Both methods run in seconds because of our reduced beam representation, whereas PhysFlow's dense volumetric inverse simulation takes 30 min–5 h; this representation is what the paper's efficiency claim rests on. DynaBeam followed by C3 refinement (3.48 s) combines both strengths, and we will add it as an optional stage.

The 10 trees used in W1 and W2 were randomly selected independently, resulting in different subsets, which explains the difference in their median ALDE values (0.055 vs. 0.098). All DynaBeam damping estimates normalize the measured decay rate by the observed oscillation frequency rather than the beam-model frequency; Table 1 already uses this, as now clarified in Sec. 3.4 and App. D.3.

AuthorsOpenReview ↗

W3: Quantitative real-world validation

Summary. Measured-force ground truth on real plants requires instrumented capture (calibrated force sensors and ground-truth 3D motion), which, to our knowledge, no existing vision benchmark provides. Following standard practice, we therefore establish physical accuracy on synthetic data with FEM ground truth. We nevertheless add two forms of quantitative real-world evidence: on three real plants, DynaBeam's trajectory error is about 3× lower than ModalNeRF's and VoMP's, and participants preferred its novel-force responses in all three comparisons of a user study.

Real-world reconstruction. We compared how well each method reproduces the video-derived 3D trajectories over a common post-release interval, with identical reference trajectories, joints and metrics. DynaBeam recovers E/ρE/\rho, ζ\zeta and time-varying modal forces with no prescribed profile. ModalNeRF uses its learned modal basis, and VoMP simulates a volumetric FEM built from the annotated skeleton with its predicted materials; both fit the free-decay response using DynaBeam's damping estimate.

Real plantDynaBeam Rel-L2 ↓ModalNeRF Rel-L2 ↓VoMP Rel-L2 ↓
Plant 10.0780.3430.262
Plant 20.1220.5470.488
Plant 30.2060.4530.545
Mean0.1350.4480.432

DynaBeam is driven by its recovered time-varying modal forces, whereas the baselines fit only the free-decay response. This comparison therefore measures reconstruction consistency on real data, not prediction; prediction under unseen forces is evaluated on synthetic data.

User study. We conducted a preliminary user study with 57 participants, each evaluating three plant–force configurations. For each configuration, participants viewed three anonymized simulations (DynaBeam, ModalNeRF, and VoMP) under the same applied force, indicated by an arrow, and selected the response that best matched the expected physical behavior.

MethodComparison 1Comparison 2Comparison 3
DynaBeam61.4%87.7%70.2%
ModalNeRF33.3%7.0%24.6%
VoMP5.3%5.3%5.3%

DynaBeam was preferred in all three comparisons. One-sided binomial tests against chance (33.3%, corresponding to three choices) yielded p < 0.001 in every comparison. Method order varied between comparisons but remained fixed across participants within each comparison.

Given the rebuttal timeline, this study is small; we will extend it in the final version with more plants and force configurations.

References. [1] Cole, "On-line failure detection and damping measurement of aerospace structures by random decrement signatures," NASA, 1973. [2] P. D. Welch, "The use of fast Fourier transform for the estimation of power spectra: A method based on time averaging over short, modified periodograms," IEEE Transactions on audio and electroacoustics, 1967

Dear Reviewer JdPW,

Thank you again for your review. Each of your concerns led to a new experiment (App. L):

  • W1: E/ρ and damping are recovered under continuous, unknown excitation with no free decay.

  • W2: A matched optimization baseline on our beam model reaches comparable material and damping accuracy at about twice the runtime.

  • W3: On real plants, DynaBeam’s reconstruction error is about 3× lower than the baselines’, and a 57-participant user study preferred it.

We’d be glad to clarify anything before the discussion closes, and would be grateful if you took these results into account in your final assessment.

Best regards,  The authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response to all reviewers

We thank the reviewers for their careful reading and constructive feedback. They recognized the beam-based design as physically grounded and well suited to vegetation (JdPW, mQcY), highlighted the efficiency of our recovery method (all three), and found the novel-force generalization convincing (mQcY).

Summary. In response to their questions and concerns, we performed additional experiments that further strengthen the paper's central claims. We demonstrate that DynaBeam requires neither pulse excitation nor a free-decay interval, recovers material properties directly from observed motion rather than relying on priors, and requires fewer additional inputs than the baselines. Moreover, it substantially outperforms all baselines without oracle information in predicting motion under novel forces and during free decay. The key new results are:

  1. No pulse or free decay needed (JdPW). Under continuous random excitation, applied throughout and never given to the method, E/ρE/\rho recovery achieves comparable accuracy to the original pulse protocol, demonstrating robustness to different excitation conditions. An output-only damping estimator recovers ζ\zeta within 20% in 59/60 runs.
  2. The motion, not a prior, drives material recovery (mQcY). DynaBeam outperforms the competing methods in the original benchmark, while three additional experiments further demonstrate its generalization capabilities. The original material range is narrow (1.32×), allowing a constant based on the nominal material mean to achieve low ALDE. However, the constant prior becomes substantially less accurate when the material range expands, and it fails to capture the observed dynamics as accurately as DynaBeam. DynaBeam predicts novel-force motion more accurately than the fixed prior (Rel-L2 0.268 vs. 0.913), tracks 0.25–4× material changes (slope 0.997), and outperforms even a constant derived from the physical properties of eleven real wood species on all 20 geometries (mean ALDE 0.067 vs. 0.151).
  3. DynaBeam relies on fewer inputs than the baselines (JdPW, AicH). In Table 1, VoMP and PhysFlow are re-simulated with the full ground-truth force, and VoMP also with ground-truth damping. DynaBeam recovers E/ρE/\rho, ζ\zeta and the force amplitude from motion, and shares only the pulse profile with the other methods. Without the force direction, held-out error changes by only −0.0006 Rel-L2.
  4. Matched optimization-based and classical baselines confirm our design (JdPW, AicH). We introduce DynaBeam-OPT (C3), which uses our identical beam representation and physical model but recovers the parameters through iterative optimization rather than DynaBeam's direct recovery procedure. It achieves comparable material and damping accuracy (median ALDE 0.113, vs. 0.098 for DynaBeam), but requires iterative optimization (4.3–4.9 s vs. 2.1 s). Because it minimizes trajectory error directly, it fits trajectories more closely, and we add it as an optional refinement stage on top of DynaBeam. Both run in seconds thanks to our beam representation, versus 30 min–5 h for PhysFlow. A classical ERA pipeline recovers E/ρ less accurately (median ALDE 0.087 vs. 0.070, 72 trees).
  5. Damping, force and free decay validated directly (mQcY). Median damping error is 2.0% and force-amplitude error 8.5%. With the force set to exactly zero after the pulse, DynaBeam predicts the free decay far better than VoMP and PhysFlow (held-out Rel-L2 0.54 vs. 4.10 and 17.9). Re-simulation uses the constrained physical force, not the modal residual, and held-out joints are excluded from every estimation step.
  6. Calibration accounts for discretization differences (AicH). The force-scaling factor cFc_F compensates for differences in compliance between the volumetric FEM ground truth and our reduced beam model. In real use, cF=1c_F = 1. Even without calibration, DynaBeam substantially outperforms ModalNeRF on novel distributed forces (Rel-L2 0.312 vs. 0.999), while cosine similarity and MDE remain unchanged (0.977 / 0.043 for distributed forces, 0.905 / 0.099 for point forces).
  7. Quantitative real-world evidence (JdPW, AicH). On three real plants, DynaBeam reproduces video-derived trajectories with a mean Rel-L2 of 0.135, versus 0.448 for ModalNeRF and 0.432 for VoMP. In a 57-participant user study, DynaBeam's novel-force responses were preferred in all three comparisons against both ModalNeRF and VoMP (61–88%, p < 0.001).
  8. Robust to geometric errors (AicH). ±20% radius or scale errors are absorbed into the effective E/ρ, with little change in motion error (held-out Rel-L2 0.243–0.287 vs. 0.267).

Accuracy under measured forces on real plants requires instrumented capture (calibrated force sensors and ground-truth 3D motion), which, to our knowledge, no existing vision benchmark provides. Uncorrelated per-frame trajectory noise is the main sensitivity (AicH Q3). Realistic spatio-temporal wind fields are left to future work.

AuthorsOpenReview ↗

Changes to the paper (marked in blue in the revised version).

  • Clarifications. We have clarified in Table 1 which ground-truth information each method receives. Sec. 3.4 and App. D.3–D.6 now distinguish the prescribed-profile force used in Table 1 from the general per-frame formulation, correct the D.6 notation, and provide more precise descriptions of the damping estimator and point-force mass correction. The results in Tables 1 and 2 remain unchanged.
  • New experiments. We have added experiments on continuous excitation with the RDT damping estimator; comparisons with C1 (ERA), C3 (model updating), and FDD, alongside a discussion of classical structural dynamics; constant-material, material-scaling, and six-species evaluations; direct damping, force, and free-decay evaluation; calibration ablation; sensitivity analysis; and quantitative real-world evaluation and a user study, including side-by-side videos.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.