acceptodds
Under review as a conference paper at ICLR 2027

NeuralCut: Data-Free Learning of Neural Reduced-Order Cut Simulation

14 views

Abstract

Cutting deformable objects is common in daily life, but the resulting displacement discontinuities pose significant challenges for reduced-order simulation. We introduce NeuralCut, a novel data-free learning framework for reduced-order cut simulation of deformable 3D objects. NeuralCut uses neural skinning to represent full-space deformation in a compact subspace. To represent cut-induced displacement discontinuities, we develop a geometry-derived lifting-and-restriction formulation that augments spatial coordinates with a cut-dependent scalar lifting coordinate. This coordinate separates opposite sides of the cut while smoothly vanishing at its boundary, enabling discontinuous deformation across the cut and continuous deformation at the boundary. We learn the skinning fields through physics-informed self-supervision without costly full-space simulation trajectories. Our training objective minimizes expected potential energy estimated by cut-aware Monte Carlo importance sampling, promotes distinct weights across the cut through a cross-side jump constraint, and encourages orthogonality among the skinning weight functions. Experiments demonstrate real-time reduced-order cut simulation across diverse 3D objects and cut configurations, supporting both meshes and Gaussian splats. We further demonstrate NeuralCut in robot interaction simulation. Moreover, coupling NeuralCut with a transformer-based model enables cross-object generalization of cut-conditioned neural skinning.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Physics-Informed Self-Supervised Learning, Neural Reduced-Order Modeling, Real-Time Cutting Simulation of Deformable Objects

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_W3pROpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Simulating the cutting of deformable 3D objects is notoriously difficult because the resulting gaps break traditional fast simulation methods. NeuralCut solves this by enabling real-time cut simulations without needing expensive, pre-recorded training data. It compresses complex deformations using neural skinning, paired with a clever geometric trick that cleanly separates the two sides of a cut while keeping the edges perfectly smooth. Instead of relying on heavy simulation datasets, the model learns directly from physics principles by minimizing energy and applying smart constraints. The result is a fast, versatile tool that works on standard 3D meshes and Gaussian splats alike, making it highly practical for applications like robot interaction and easily adaptable to entirely new objects when paired with a transformer model.

Critical strengths

Enables real-time reduced-order simulation with high computational speed. Utilizes data-free, physics-based self-supervised learning, eliminating the need to generate costly simulation datasets.

Employs a simple geometric technique—coordinate lifting—to effectively handle separation on both sides of a cut while maintaining smooth edges.

Supports both traditional 3D meshes and state-of-the-art 3D Gaussian splats; when integrated with a Transformer, it can generalize to entirely new objects. Proven effective for robotic interaction simulations, with significant potential for practical application.

Overall, the results are very good, and the figures are presented excellently.

Critical weaknesses

While it eliminates the need for external data, the self-supervised training process based on physical energy minimization and Monte Carlo sampling can be computationally expensive or difficult to converge.

As a "reduced-order" simulation, it may offer lower physical precision compared to traditional, high-fidelity, full-space simulations when dealing with extremely complex deformations.

Not very sure? Because the approach relies on specific geometrically derived formulas, it may face modeling constraints when encountering highly complex or non-standard dynamic cutting topologies.

Free to include in the reference:

Reduced-order neural modeling with differentiable simulation for high-detail tactile perception. In 2026 IEEE 9th International Conference on Soft Robotics (RoboSoft). IEEE.

RealSimLoop: Online real-to-sim adaptation via differentiable reduced-order simulation with vision feedback. In Proceedings of the 2026 ACM SIGGRAPH Asia Conference. Association for Computing Machinery.

Questions for authors

None

3 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We sincerely thank Reviewer W3pR for the thoughtful feedback and strong endorsement of our work. We are encouraged by the recognition of our conceptual and technical contributions, the strength of our experimental results, and the broad potential for practical applications. Below, we address the reviewer's comments.

Q1. Training Cost and Convergence.

Response: We appreciate the reviewer's concern regarding the computational cost and convergence of physics-informed self-supervised training.

Computational Efficiency. Our implementation is built upon NVIDIA Warp [1], leveraging GPU-accelerated computation for Monte Carlo sampling and physics-based optimization. We further profile the computational cost of the key training components, as reported below.

ComponentAverage Time / Step
Monte Carlo Sampling1.09 ms
Physical Energy Evaluation0.37 ms
Physics Loss Forward38.01 ms
Physics Loss Backward75.37 ms

These results demonstrate that our GPU-accelerated implementation enables efficient physics-based optimization.

Convergence Stability. We agree that energy-based, data-free training can be difficult to converge. In our experiments, we observed similar difficulties when adopting the optimization strategy of Simplicits [2]. However, by incorporating the on-the-fly skinning-field normalization and conflict-aware gradient correction introduced in PhysSkin [3], we effectively address these convergence issues and achieve stable training in NeuralCut.

References

[1] Warp: A High-performance Python Framework for GPU Simulation and Graphics. NVIDIA GPU Technology Conference (GTC), 2022.

[2] Simplicits: Mesh-Free, Geometry-Agnostic Elastic Simulation. ACM Transactions on Graphics, 2024.

[3] PhysSkin: Real-Time and Generalizable Physics-Based Animation via Self-Supervised Neural Skinning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026.

Q2. Reduced-Order Limitations under Complex Deformations.

Response: We thank the reviewer for highlighting this limitation. We acknowledge that reduced-order simulation may offer lower physical precision than high-fidelity full-space simulations when handling extremely complex deformations. One promising future direction is to augment the reduced-order representation with a complementary neural deformation model to capture residual deformations beyond the reduced subspace. We have further discussed this limitation and potential future direction in the revised paper.

Q3. Complex Cutting Topologies.

Response: We appreciate the reviewer's concern regarding complex cutting topologies. Our geometry-derived lifting formulation also enables generalization to curved cut surfaces without additional training, as demonstrated in video_curved_cut.mp4 in the updated Supplementary Material.

Q4. Additional References.

Response: We thank the reviewer for suggesting these relevant works. We agree that both papers are closely related to NeuralCut and have incorporated the corresponding citations into the revised paper.

Thank you for your detailed and satisfactory rebuttal. You have adequately addressed my concerns regarding the training cost, convergence stability, and the limitations of the reduced-order simulation. The additional clarifications and results provided are highly appreciated.

I have no further questions and will maintain my positive assessment of the paper. I wish you the best of luck with the final decision.

Additionally, I highly encourage you to open-source your code and models if the paper is accepted. This would be a valuable contribution to the community.

We sincerely thank the reviewer for the encouraging feedback and continued support. We are glad that our responses have addressed your concerns. We will release our code and models upon acceptance.

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

This paper proposes a data-free framework for neural reduced-order simulation of deformable 3D objects with finite cuts. Its geometry-derived lifting coordinate conditions neural skinning fields on the prescribed cut, allowing displacement discontinuities across the cut surface while maintaining continuity at its boundary. The framework combines physics-informed optimization with cut-aware sampling and a cross-side separation constraint. Experiments include comparisons with full-space solvers, mesh and 3D Gaussian representations, interactive manipulation, and a transformer-based extension for cross-object skinning prediction.

Critical strengths

(1) A meaningful problem and a well-motivated representation. Supporting cut-induced discontinuities is a genuine challenge for continuous neural reduced-order representations. The lifting construction in Sec. 3.2 provides a clear mechanism for separating opposite cut sides while bringing their lifted coordinates together at the cut boundary. Its integration with data-free neural skinning is a useful contribution.

(2) Promising computational efficiency with quantitative evidence. Table 2 reports simulation times of 4.3–8.8 ms per step across eight objects with approximately 29K–208K vertices, substantially faster than the reported full-space solvers. Table 1 additionally reports low final-state errors relative to FEM, suggesting a promising accuracy–efficiency tradeoff in the evaluated settings.

(3) Broad practical demonstrations. The paper demonstrates the framework on both meshes and 3D Gaussian representations, with additional interactive and robot-coupled simulations. Importantly, Appendix Fig. 12 compares eight objects under two cut configurations each against FEM and MPM. These results provide meaningful evidence beyond isolated examples.

Critical weaknesses

(1) The cut-configuration evaluation protocol and generalization evidence are insufficiently specified. The paper demonstrates multiple objects and cut configurations, but it does not clearly specify the parameterization or sampling distribution of the training cuts, or their relationship to the evaluation cuts. Sec. 4.1 states that a cut configuration is sampled at each training iteration, but it does not describe the ranges of positions, orientations, or finite extents. Furthermore, Table 1 reports only one NMSE/RDE result per method, without identifying the evaluated object–cut combinations or explaining how their errors are combined. Consequently, it is difficult to assess performance variation across configurations or generalization to unseen cuts.

(2) The ablations do not establish the contribution of each component to simulation accuracy. Tables 3–4 report UCD, CSS, and OLS, which measure the continuity and separation properties of the skinning fields. These are appropriate mechanism-level diagnostics, but they do not directly measure deformation accuracy, as NMSE and RDE do.

(3) The presentation could be improved for accessibility. Since many ICLR readers may not have specialized backgrounds in graphics and physics-based simulation, domain-specific concepts and terminology should be explained more intuitively. This is why I selected "Generally understandable, but requires substantial effort in places." In addition, I suggest simplifying Fig. 2 by removing the detailed mathematical formulas and focusing instead on the conceptual workflow.

Questions for authors

(1) How are cut configurations parameterized and sampled during training and evaluation? Are evaluation configurations independently sampled or otherwise held out? What cut geometries are included?

(2) Which object–cut combinations contribute to Table 1, and how are their errors combined? Could per-object/per-cut results and variability be provided?

(3) Could the ablated models in Tables 3–4 also be evaluated using simulation accuracy metrics like NMSE/RDE?

3 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We sincerely thank Reviewer QMJS for the thoughtful and constructive feedback. We are encouraged by the recognition of the importance of the problem, the well-motivated design, and the promising experimental results. Below, we provide detailed responses to each of the reviewer's comments.

Q1. Cut Configuration Sampling and Generalization.

Response: Each planar cut configuration used for training is defined by its center, plane orientation, and finite extent. Each object is normalized to fit within [−1,1]3[-1,1]^3. During training, we randomly sample a new cut center within the object, a plane orientation in 3D, and a finite extent bounded by the object's bounding-box dimensions. All evaluation cut configurations are explicitly held out and excluded from training-time cut sampling, as detailed in Q2.

Although training uses planar cuts, our lifting formulation also enables generalization to curved cut surfaces without additional training. In the updated Supplementary Material, we provide video_curved_cut.mp4 to demonstrate this capability.

We have also incorporated further methodological details and the new results in Fig. 10 into the revised paper.

Q2. Per-Object and Per-Cut Quantitative Results.

Response: In the original submission, Table 1 reports errors averaged over eight objects, each evaluated under three held-out cut configurations (24 cases in total). These configurations were excluded from training-time cut sampling, with corresponding qualitative comparisons presented in Fig. 7 and Fig. 12 (with Fig. 12 renumbered as Fig. 13 in the revised paper).

To further assess variability across cut configurations, we extended the evaluation to five cuts per object, yielding 40 object–cut cases. The tables below report the per-object average NMSE and RDE.

NMSE (↓)

MethodOrangeCarSpongeStarfishBirdPigBagTomatoAverage
MPM15.825×10⁻³9.918×10⁻³7.940×10⁻³4.328×10⁻³3.240×10⁻³19.765×10⁻³6.785×10⁻³16.436×10⁻³10.530×10⁻³
Ours1.684×10⁻³2.546×10⁻³0.011×10⁻³0.440×10⁻³0.117×10⁻³0.965×10⁻³0.960×10⁻³1.133×10⁻³0.982×10⁻³

RDE (↓)

MethodOrangeCarSpongeStarfishBirdPigBagTomatoAverage
MPM11.844%9.687%7.414%6.071%5.112%10.521%7.504%11.616%8.721%
Ours3.898%4.478%0.293%1.669%0.704%2.404%2.663%2.965%2.384%

We provide the complete results for all 40 object–cut cases in per_cut_results.csv and the per-object averages for each method in per_object_averages.csv, both in the updated Supplementary Material. We have also updated the corresponding results and experimental details in the revised paper.

Q3. Simulation Accuracy Ablations.

Response: As the reviewer noted, the UCD, CSS, and OLS metrics in Tables 3–4 primarily characterize the properties of learned skinning fields rather than simulation accuracy. Following this suggestion, we extended the ablation study to include NMSE and RDE, evaluated on the same 40 object–cut cases as in Q2.

MethodNMSE ↓RDE ↓
MPM10.530×10⁻³8.721%
w/o PE1.745×10⁻³3.265%
w/o Smooth3.331×10⁻³4.697%
w/o Sampling1.880×10⁻³3.430%
w/o Jump1.729×10⁻³3.369%
Ours0.982×10⁻³2.384%

The results demonstrate that the full model achieves the lowest average simulation errors across all variants. Detailed per-case results are provided in per_cut_results.csv, with per-object averages in per_object_averages.csv, both in the updated Supplementary Material. We have also updated the corresponding ablation results in Table 5 and evaluation details in the revised paper.

Q4. Clarity and Accessibility.

Response: We appreciate the reviewer's suggestion to improve accessibility for readers less familiar with graphics and physics-based simulation. We will clarify domain-specific terminology and provide more intuitive explanations of the key concepts and methodology. We will also refine the accompanying description of Fig. 2 to better highlight the overall conceptual workflow.

Thank you for the detailed rebuttal. I believe my concerns have been well addressed, and I am raising my initial rating to 3. I hope the authors will improve the presentation in the revised version so that the paper is more accessible to a broader readership.

We sincerely thank the reviewer for the encouraging feedback and for reconsidering the assessment. We appreciate your suggestion and will continue improving the clarity and accessibility of the paper for a broader readership.

Reviewer_ARAVOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

NeuralCut is a data-free method for real-time reduced-order simulation of soft 3D objects under finite cuts. It uses neural skinning in a compact subspace, and a geometry-derived lifting coordinate that separates the two sides of a cut and fades out at its boundary, so the model can represent discontinuous motion across the cut. The skinning fields are learned by minimizing expected elastic energy. It could run in real time on meshes and 3dgs objects, and also supports robot cutting interactions.

Critical strengths
  1. The lifting constraction addresses an important limitation of continuous neural deformation fields. It could represent displacement discontinuities across a cut while perserving connectivity at its boundary. The combination of a side indicator and smooth boundary attenuation provides an interpretable geometric mechanism for this behavior.
  2. The physical objective, cross-side separation constraint, and cut-aware sampling are motivated by the challenges of learning a cut-sensitive deformation space without reference trajectories. The reported numerical results and demonstrations across meshes, Gaussain splats, and robot interactions provide evidence of practical versatility.
Critical weaknesses
  1. Table 1 only reports overall errors. It would help to clarify how many objects and cuts were tested and show results for individual cases.
  2. The cross-object results are mostly visual, and it is unclear which objects were unseen during training. A quantitative comparison with the per-object models would make the generalization claim more convincing.
Questions for authors
  1. How many objects and cuts are evaluated in Table 1? Could you report errors for individual cases?
  2. Are the cross-object results on unseen objects without additional training? How does the shared model compare with per-object models?
1 comment
AuthorsOpenReview ↗

Rebuttal by Authors

We sincerely thank Reviewer ARAV for the thoughtful and constructive feedback. We are encouraged by the recognition that our work addresses an important limitation and demonstrates practical versatility. Below, we provide detailed responses to each of the reviewer's comments.

Q1. Per-Object and Per-Cut Quantitative Results.

Response: In the original submission, Table 1 reports errors averaged over eight objects, each evaluated under three held-out cut configurations (24 cases in total). These configurations were excluded from training-time cut sampling, with corresponding qualitative comparisons presented in Fig. 7 and Fig. 12 (with Fig. 12 renumbered as Fig. 13 in the revised paper).

To further assess variability across cut configurations, we extended the evaluation to five cuts per object, yielding 40 object–cut cases. The tables below report the per-object average NMSE and RDE.

NMSE (↓)

MethodOrangeCarSpongeStarfishBirdPigBagTomatoAverage
MPM15.825×10⁻³9.918×10⁻³7.940×10⁻³4.328×10⁻³3.240×10⁻³19.765×10⁻³6.785×10⁻³16.436×10⁻³10.530×10⁻³
Ours1.684×10⁻³2.546×10⁻³0.011×10⁻³0.440×10⁻³0.117×10⁻³0.965×10⁻³0.960×10⁻³1.133×10⁻³0.982×10⁻³

RDE (↓)

MethodOrangeCarSpongeStarfishBirdPigBagTomatoAverage
MPM11.844%9.687%7.414%6.071%5.112%10.521%7.504%11.616%8.721%
Ours3.898%4.478%0.293%1.669%0.704%2.404%2.663%2.965%2.384%

We provide the complete results for all 40 object–cut cases in per_cut_results.csv and the per-object averages for each method in per_object_averages.csv, both in the updated Supplementary Material. We have also updated the corresponding results in Table 1 and experimental details in the revised paper.

Q2. Cross-Object Generalization and Comparison with Per-Object Models.

Response: All objects shown in Fig. 10 of the original submission (renumbered as Fig. 11 in the revised paper) are unseen during Transformer training, and the shared model directly predicts their cut-conditioned skinning fields without additional per-object training or fine-tuning.

Following the reviewer's suggestion, we conducted a quantitative comparison between the shared Transformer and per-object models across 30 object–cut configurations from this cross-object evaluation. We also provide side-by-side visualizations of their corresponding skinning modes in transformer_object_pairs.pdf in the updated Supplementary Material, with case IDs matching those in the table below. The overall average cosine similarity is 0.9358, demonstrating that the shared Transformer can predict skinning modes closely matching those of per-object models on unseen objects.

CaseCosineCaseCosineCaseCosineCaseCosineCaseCosine
10.875220.954430.981740.880650.8966
60.971370.978880.967390.9815100.8994
110.8836120.9299130.9180140.9601150.9343
160.9845170.9318180.9818190.9788200.9016
210.8864220.9093230.9386240.9676250.9519
260.9864270.8810280.9826290.9018300.8758

Comments on the paper

2 comments
Area_Chair_7rfuOpenReview ↗

Discussion encouraged during this period!

Hi all reviewers,

We’re about halfway through the discussion period. This is a friendly reminder to check the authors' replies. If the authors have answered questions or responded to critiques in the reviews, please leave a comment for the authors!

Thank you so much!

Best, Your AC

Thank you for the detailed response. The author have successfully answered my questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.