acceptodds
Withdrawn from ICLR 2027

KnockGS: Interaction-Grounded Calibration of Physical Gaussian Representations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Physics-integrated 3D Gaussian representations now allow reconstructed deformable objects to be simulated and rendered under explicit material models. Existing pipelines, however, assume that material parameters are known or manually specified, limiting their applicability when these parameters must be inferred from observed object dynamics. We propose KnockGS, an interaction-response PhysicalGS framework that estimates the elasticity and density scales of a 3D Gaussian object from its dynamics under a known applied force. Rather than treating physical simulation only as a forward process, we turn the force-induced response into a calibration signal: temporal response features are extracted from the observed dynamics, the two material scales are estimated from those features, and the estimate is then frozen and written back into the same simulator so that it can be tested on an interaction it was never fitted to. We evaluate the framework on both parameter recovery and response-level fidelity. The estimated scales are compared against hidden ground truth, and the re-simulated object is measured against the target using 3D particle trajectories, response-curve statistics, and rendered-frame quality. Across five held-out material targets, our method recovers the scales substantially more accurately than response retrieval, global regression, or a fixed default material, and the frozen estimate remains predictive under interactions that differ in direction and in magnitude. Interaction response therefore carries enough information to calibrate material scales in physically grounded 3D Gaussian representations. Our study is a first step toward interactive PhysicalGS systems that calibrate a Gaussian asset whose rendered appearance and simulated response are consistent.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Physical AI, 3D Gaussian Splatting, material parameter calibration, active interaction, material point method

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_9dvmOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents KnockGS, a controlled material-calibration procedure for physics-integrated Gaussian assets. For each object and fixed simulator configuration, the method precomputes a library of 54 Probe-A responses over Young's-modulus and density scales. A deterministic five-dimensional descriptor summarizes displacement and velocity curves, after which the ten nearest library responses are used in a local ridge regression to estimate the two continuous scales. The estimate is frozen and written back into the same MPM or FEM simulator, where it is evaluated under held-out force directions or magnitudes. The paper reports parameter error, particle-trajectory error, response statistics, and rendered PSNR/SSIM, and extensively analyzes probe selection, library size, observation quality, discretization mismatch, and failure boundaries.

Critical strengths
  1. Careful experimental protocol and unusually transparent scope. The separation between calibration Probe A and held-out Probe B is appropriate, and the paper carefully documents target exclusion, candidate-only tuning, per-target results, alternative splits, sensitivity, and numerical boundaries. The authors explicitly report failures under object, discretization, observation-noise, and correspondence changes.

  2. Consistent results within the declared simulator contract. Local ridge improves over response retrieval, KNN, global ridge, and fixed defaults on the main aggregate results. The frozen estimates remain predictive under direction and magnitude changes, and a separately constructed FEM library reproduces the basic finding.

Critical weaknesses
  1. The technical contribution is limited to object-specific response-table interpolation. The core estimator consists of five handcrafted response statistics, nearest-neighbor selection, and a standard local ridge fit for two scalar parameters. A separate 54-simulation library must be constructed for every object and numerical configuration. The method fails under cross-object sharing and large fill or grid mismatch unless a matching library is rebuilt. Moreover, the comparisons do not include standard response-surface alternatives such as radial-basis interpolation, Gaussian-process regression, kernel ridge, local polynomial regression, or a small learned regressor. Consequently, the experiments do not establish a sufficiently distinctive methodological contribution beyond classical local interpolation of a smooth simulator response map.

  2. The evaluation does not substantiate the motivating visual PhysicalGS calibration setting. The primary estimator observes privileged simulator particle states with exact correspondence, while targets, candidates, and held-out predictions are all generated by the same simulator and lie within the covered two-parameter range. Clean synthetic depth retains useful evidence, but calibration degrades rapidly with noise and loses its advantage with ID-free LK tracks; no real RGB/RGB-D observations, measured forces, real material parameters, or sim-to-real experiments are provided. The Gaussian representation is also not essential to the estimator, which operates on generic MPM/FEM trajectories. Rendered images differ only subtly, and the large PSNR improvements mostly reflect small pixel-level misalignments in a largely static scene rather than clearly distinguishable physical behavior. Thus, held-out Probe-B prediction mainly verifies same-simulator self-consistency after interpolation, not practical physical reconstruction from interaction observations.

Questions for authors
  1. What capability specifically depends on the Gaussian representation, rather than applying the same response descriptor and local interpolation to an arbitrary particle or mesh simulator?

  2. How does the estimator compare with stronger classical response-surface methods such as RBF interpolation, Gaussian processes, kernel ridge, and local polynomial regression under equal library budgets, noise levels, and target splits?

  3. Can the method be evaluated using captured RGB-D trajectories and measured interactions without simulator particle identities? Larger deformation close-ups, displacement overlays, and videos would also help establish whether the reported image-metric differences correspond to meaningful physical-response differences.

Reviewer_jvoJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes partial prediction of material properties for the PhysicalGS framework. Through response observations after a known interaction which the authors refer to as “Probe-A”, KnockGS estimates the elasticity and density scale property of the observed material, which can further be used to predict either an unknown interaction referred as “Probe-B” or the response from “Probe-B”. Empirical comparison is made against some straightforward methods as KnockGS’s problem setting is unique. The quantitative comparisons, focused on the Pillow scene in the main paper, show the best accuracy in material parameter estimation and relevant estimations. However, in qualitative comparison the differences visualized in Final-frame abs. error seem minimal as the response of the Gaussians, or the difference between Mid and Final frame is not discernible even on the Target renderings.

Critical strengths
  1. Uniqueness and clarity in problem statement of estimating material properties based on known interaction.
Critical weaknesses
  1. Lack of explanation of why this method is important compared to visual-prior approach, passive-video inverse, or PhysTwin. More examples should be given where visual-prior approach, passive-video inverse methods are not applicable, and that the use of known interaction is applicable. Otherwise, the problem statement that this paper is solving does seems applicable only to a narrow target situation and hence is not effective in persuading the significance of the given problem statement.

  2. What is the significance of estimating just the elasticity and density scale properties and assuming that other material properties are given and fixed? Similar to above statement, more explanations or examples showing on why this setting is a significant problem to solve should be presented because otherwise, the problem setting seems to be for cases when we there is already some information of the object of interest that makes this task less challenging.

Questions for authors
  • Figure 2 shows hardly any difference in renderings for for Mid frame and Final frame. Is this method only applicable for very small perturbations? Either use zoom-in boxes to highlight the differences in the renderings as well, or use larger, more visible interactions with larger deformations and present qualitative comparisons that clearly show superiority.

  • Why is the evaluation reported in the main paper concentrated on Pillow? Other scenes including Sofa were mentioned but the quantitative or qualitative comparisons should be included to show the generalizability of this method to other scenes than Pillow.

Some suggestions for future work: Some remarks for fully contained paper:

  • Full definitions of some variables are missing. Such as simulator contract C, s_early(r), or xhat^t or x^{*,t}
  • Explanation of how the 3D reconstruction is made should be included.
Reviewer_GaP5OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents KnockGS, which calibrates the Young's-modulus and density scales of a fixed physics-integrated Gaussian asset from its response to a known applied force (Probe A). It builds an object-specific simulation, extracts five temporal response features, selects ten nearest neighbors, and fits local affine ridge regressions to estimate the two scales. The estimate is frozen before predicting held-out interactions (Probe B), such as changes in force direction or magnitude, under the same simulator setup. Evaluation centers on a reconstructed Pillow asset, with separate Ficus and Vasedeck calibrations, observation and stability diagnostics, and a rebuilt FEM contact library.

Critical strengths

Implementing calibration on physical assets and supporting Probe-B evaluation tests are useful functions for simulation and downstream tasks. On the primary Pillow benchmark, local ridge reduces the mean joint parameter error from 2.37% for KNN to 1.13% and improves both held-out trajectory metrics. The supplementary evidence helps characterize both repeatability and failure.

Critical weaknesses
  1. The overall presentation makes the paper harder to follow than necessary. For example, Figure 1 packs the procedure into small nested boxes with unrelated figures and symbols. A simpler overview with larger labels and clearer separation of offline construction, per-target estimation and evaluation would make the method substantially easier to understand.

  2. The technical contribution is limited. Within a fixed asset and simulator, the estimator combines a hand-designed five-dimensional descriptor, nearest-neighbor selection and a local affine ridge fit for two material scales, with limited methodological development beyond those components. A clearer account of the distinct technical contribution would be necessary.

  3. The experimental scope remains limited despite the additional assets and FEM study. Each asset or solver uses a separately constructed library with the physical setup fixed. Practical robustness remains unresolved: ID-free tracking raises local-ridge error to 8.67%, worse than global ridge at 6.73%. The experiments are insufficient to demonstrate the effectiveness, generalizability, and stability of the method.

Questions for authors
  • Can you show whether the local estimator retains a frozen Probe-B advantage with ID-free tracks or realistic depth perturbations, and identify the observation regime where it ceases to outperform global ridge?

  • What will happen when the physical parameter is close to the limit of the plausible range for Young's modulus and density? Will the optimization be stable?

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

Related papers

Discussion (0)

Sign in to comment.