acceptodds
Under review as a conference paper at ICLR 2027

FracFusion: A Heterogeneous Simulation-Rendering Framework for Vehicle Impact and Structural Breakage

Abstract

High‑fidelity simulation of simultaneous structural breakage in vehicle‑obstacle collisions is critical for autonomous‑driving safety testing and neural world modeling. Real crash data is scarce and costly to acquire, making simulation the main source of diverse accident scenarios. Existing pipelines face a fundamental trade‑off: node‑beam solvers capture realistic vehicle‑component structural fracture but cannot generate arbitrary continuum fragmentation for obstacles, whereas MPM‑based 3D Gaussian Splatting (3DGS) produces debris but discards calibrated vehicle mechanics when vehicles are converted to particles. Naively combining solvers cannot resolve heterogeneous contact, consistent bidirectional momentum transfer, and persistent inter‑penetration artifacts. We present FracFusion, a heterogeneous simulation‑rendering framework. It keeps calibrated node‑beam vehicle assets intact and uses MPM‑augmented 3DGS for photorealistic brittle obstacle fracture. Lightweight shell proxy particles enable stable bidirectional impulse coupling within Collision‑MPM without redundant vehicle dynamics. We further incorporate particle‑triangle anti‑penetration correction and physics‑aware depth compositing to mitigate cross‑representation artifacts. Experiments and ablations validate our design. Physical checks verify momentum and energy stability under high‑energy impacts, and perceptual studies show improved visual plausibility. FracFusion delivers crash simulation featuring both vehicle‑component structural fracture and obstacle continuum fragmentation for safety‑critical autonomous simulation and world‑model applications.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Hybrid simulation, Node-beam, MPM, Fracture Simulation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_AcsiOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents FracFusion, a simulation-rendering framework for vehicle-obstacle collisions. It combines node-beam vehicles with Gaussian/MPM obstacles through shell proxies that transfer contact impulses back to the vehicle nodes. The system preserves beam-level vehicle damage while supporting obstacle fragmentation, and adds penetration correction and depth compositing to improve the resulting images. Evaluation includes comparisons with node-beam and FractureGS baselines, a 30-participant user study, component ablations, and momentum/energy checks.

Critical strengths
  1. The motivation is practical. Keeping existing node-beam vehicle assets while using Gaussian/MPM obstacles allows the system to combine structured vehicle deformation with detailed obstacle fragmentation. The shell-proxy interface provides a concrete way to connect the two representations without introducing separate vehicle-particle dynamics.

  2. The results show useful complementary behavior. In Table 1, the hybrid retains a structural-response rating close to node-beam (3.8 versus 3.9), while improving the material-fracture rating from 1.2 to 3.9. The component ablations also support the usefulness of the contact and rendering corrections.

Critical weaknesses
  1. I find the technical contribution insufficiently established. The framework largely integrates existing node-beam dynamics, MPM contact, and Gaussian rendering. The vehicle can deform and its beams can break, but this behavior is inherited from the input asset; body-mesh tearing and surface cracks are not modeled. The fractureMPM is mainly borrowed from PhysGaussian and GaussianFluents. It's not clear what the key problem is when combining the two representations.

  2. The main comparison of overall quality relies on a 30-participant user study on 3 cases. There are more qualitative examples in the Appendix, but they do not demonstrate realistic simulations. For example, the door is crashed into particles in Fig. 9(b); the watermelon is torn into pieces in Fig. 10(a). The timing results, penetration ablations, and momentum/energy checks provide additional evidence, but they do not resolve this uncertainty in the quality comparison.

  3. The method figures in the paper are hard to comprehend, including complicated symbols but simple procedure illustrations. Repeated missing word separators, such as “Endtoend GPU simulationrendering and occlusion handling” (line 88) and “GPUaccelerated” (line 120), further interrupt the reading.

Questions for authors

What is the key technical advance of the shell-proxy coupling over the earlier coupling approaches discussed in Section 2.4?

6 comments
AuthorsOpenReview ↗

This response highlights our novel heterogeneous coupling pipeline, clarifies our implemented large scale vehicle deformation and beam level fracture supported by extensive supplementary experiments including a dedicated two car extreme collision case, distinguishes our innovations from prior MPM Gaussian works, and explains that beam level failure already covers common real world crash phenomen.

Thank you for your comment. We do not adopt an off the shelf complete node beam simulator. Only topology and material parameters are parsed from community assets. All large scale deformation, plastic evolution and beam fracture logic are implemented within our custom GPU resident explicit node beam solver. Stably integrating discrete beam elements on Warp GPU under high speed impact, handling plasticity, severe distortion and component level fracture poses substantial numerical challenges and cannot be obtained by merely loading asset files. Extensive experiments and videos are provided in revised supplementary material. We have also designed an extreme high speed two car collision case (see Suppl. Video S-5), where large scale distortion and component fracture on bumpers, chassis and other components can be directly observed.

Our core innovation lies in a novel heterogeneous bidirectional coupling paradigm. Via the shell proxy interface, fully dynamic node beam vehicles can directly interface with split field Collision MPM and perform bidirectional impulse exchange with MPM 3DGS obstacles. Existing MPM Gaussian works (PhysGaussian, GaussianFluents, Fracture GS) are restricted to all particle scenes. They cannot accept external structured non particle bodies into dual field contact solves. Naive combination of separate systems yields severe penetration, distorted impulse transfer and cross representation rendering conflicts. Our ablation studies in Fig.5 and Suppl. video S-2.2 demonstrate that removing shell proxy leads to failed component fracture and severe inter penetration. These coupling challenges have not been addressed in prior literature. Our shell proxy interface, particle triangle anti penetration correction and physics aware depth compositing are specifically proposed to resolve these coupling artifacts.

We would clarify that our framework targets component wise beam level fracture. Large scale distortion, plastic deformation and structural failure of bumpers, chassis and other components are fully reproduced. In real world vehicle crashes, dominant failure modes are bending, buckling and joint fracture of chassis and bumper components. Fine grained triangle mesh tearing on car body panels occurs only in relatively extreme and rare crash scenarios. Therefore our beam component level simulation already covers most common real world crash phenomena. Fine mesh level surface tearing requires dedicated additional mesh fracture algorithms and is out of scope for this paper, which we leave for future work. We adopt MPM 3DGS as the obstacle side rendering backend; our contribution is not the MPM Gaussian rendering itself, but enabling the heterogeneous simulation pipeline connecting structured node beam and particle based MPM.

AuthorsOpenReview ↗

This response clarifies the role of the user study, explains our multi faceted validation pipeline with explicit pointers to where each experiment is located within paper / appendix / supplementary material, and addresses reviewer’s concerns regarding perceptual evaluation limitations.

Thank you for your feedback. This user study, presented in Section 4 of main paper, is designed only for subjective perceptual comparison, to evaluate visual plausibility of outputs from different approaches, rather than proving absolute physical correctness of our simulation. To comprehensively evaluate our method, we adopt a multi faceted validation pipeline instead of relying solely on the user study. The location for each set of experiments is explicitly listed below:

  1. Component wise ablation studies: reported in Fig.5 of the main text, verifying the necessity of core modules including shell proxy particles and anti penetration correction for suppressing penetration artifacts and preserving component fracture behaviour. Key sub plots are Fig.5(a)(d), and corresponding dynamic fracture behaviours are shown in supplementary video S-2.2, visually demonstrating component fracture failure and inter penetration after removing key modules.
  2. Physical checks: Basic momentum and energy residual curves and statistics are provided in Appendix C.4. Beyond that, our revised supplementary material contains extensive additional physical evaluation experiments, including momentum energy traces under varied collision velocities, impulse error statistics under different proxy particle densities, and quantitative penetration depth metrics, which further examine physical behaviour and numerical stability across more conditions.
  3. Rich qualitative simulation cases and dynamic video: Static snapshots: Fig.9, Fig.10 in main paper and Appendix D, including door shattering and watermelon fracture examples mentioned by the reviewer. Dynamic sequences: full set videos in revised supplementary material, including our dedicated extreme high speed two car collision case video S-5, plus multiple obstacle fragmentation scenes.

Regarding the door and watermelon examples pointed out by the reviewer: both objects are brittle and shatter completely under strong impact, which is exactly the behaviour represented by MPM particle fragmentation. It should be distinguished that load bearing vehicle components such as chassis and bumpers undergo beam level fracture handled by the node beam solver. When brittle shells like door panels or watermelons are fully destroyed, they are represented as MPM particles for fragment rendering. This mixed representation is one target capability of our heterogeneous framework.

In revision, we will explicitly state the scope and limitation of the user study in the experiment section, separate subjective perceptual evaluation from ground truth physical validation, and cite ablation figures, appendix sanity check results, additional physical evaluations in supplementary material and all supplementary videos as our full evidence chain.

AuthorsOpenReview ↗

This response confirms full revision of formatting typos, diagrams and symbol consistency in our updated manuscript.

Thank you for pointing out these formatting and readability issues. We will thoroughly fix all concatenated word typos, refine method diagrams, and improve label, data flow and symbol consistency to enhance readability in our revised manuscript.

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

FracFusion is a heterogeneous simulation and rendering framework for vehicle–obstacle collisions. The main idea is to keep the node–beam representation for vehicles while using MPM and 3D Gaussian Splatting to simulate and render obstacle fragmentation. The two solvers are coupled through shell proxy particles and a shared CollisionMPM grid, allowing bidirectional momentum transfer. A particle–triangle penetration correction and a depth-based compositing strategy are introduced to address contact and rendering artifacts. The results demonstrate simultaneous vehicle deformation and obstacle fragmentation, with improved visual quality and lower simulation cost than the FractureGS baseline.

Critical strengths

The shell proxy interface allows bidirectional coupling without introducing additional vehicle degrees of freedom.

Critical weaknesses
  1. The paper is more about combining and applying existing computer graphics techniques. Contributions to machine learning or its downstream applications are limited. Although the results look better than the baselines, and the simulation shows reasonable conservation properties, these do not mean the simulation is physically accurate.
  2. The methodological novelty is narrow, as most of the components are based on existing methods. The paper shows that combining two different representations is beneficial, but has not explained what makes the proposed coupling strategy unique or better than other possible approaches.
Questions for authors
  1. The authors claim that the node–beam representation can capture realistic vehicle component fracture. However, in the results, only the obstacles seem to break apart.
  2. High-fidelity dynamic simulation is a main focus of the paper, but no supplementary videos containing dynamic results seem to be provided.
5 comments
AuthorsOpenReview ↗

This response contrasts shell proxy against alternative coupling options, explains why other candidate strategies suffer practical drawbacks supported by our experiments, and adds comparative discussions in revised related work section.

Thank you for the feedback. Besides our shell proxy strategy, several alternative heterogeneous coupling options exist, yet each carries clear practical drawbacks:

  1. Convert everything to MPM particles (e.g. Fracture GS [2]): vehicles also become particles. Reproducing component wise heterogeneous structural fracture of real vehicles purely with MPM particles is intrinsically difficult. It requires meticulous per particle tuning and manual reconstruction of layered mechanical structure, without mature off the shelf tooling, leading to prohibitive manual cost. This is also experimentally validated in Fig.5(b) of our main paper: converting the vehicle fully into MPM particles fails to reproduce the native component level fracture behaviours from node beam assets. Moreover, this discards well tuned component wise fracture behaviour from node beam assets. The Rigs of Rods ecosystem hosts a large library of community calibrated node beam vehicles with tuned stiffness, plasticity and fracture thresholds. Converting vehicles fully into MPM particles makes component level fracture responses hard to reproduce and wastes this widely available pre calibrated asset base, which contradicts our core motivation.

  2. Direct point grid contact: node beam nodes directly interact with the MPM grid without full surface sampling. This yields unstable contact constraints, local penetration and spurious impulse noise. This failure mode is demonstrated in Fig.5(d) and supplementary video S-2.2: without shell proxy particles and using only beam nodes for contact, severe inter penetration occurs between vehicle and wall obstacles.

  3. FEM MPM interaction nodes from reference [1]: designed for single field MPM and FEM meshes. Beyond incompatibility with split field dual field Collision MPM and lack of native node beam support, the underlying FEM formulation struggles with large distortion and massive structural fracture; FEM meshes degrade numerically under severe deformation, which mismatches our target crash scenarios. The uniqueness of our shell proxy lies in read only surface proxy particles that introduce zero additional vehicle dynamical degrees of freedom, are compatible with dual field MPM contact, keep vehicles un particleized, and reliably scatter impulses back to node beam nodes, avoiding drawbacks of above alternatives.

[1] Cheon, Y.-J., & Kim, H.-G. (2018). An efficient contact algorithm for the interaction of material particles with finite elements. Computer Methods in Applied Mechanics and Engineering, 335, 631–659.

[2] Wang, X., Wu, H., Song, W., & Xu, K. (2026). Fracture-GS: Dynamic Fracture Simulation with Physics-Integrated Gaussian Splatting. International Conference on Learning Representations (ICLR).

AuthorsOpenReview ↗

This response points to newly provided supplementary videos demonstrating key dynamic simulation behaviours.

Thank you for your valuable suggestion. We provide extensive dynamic simulation videos in the revised supplementary material, covering vehicle obstacle collisions, vehicle component fracture, obstacle fragmentation, ablation comparisons, and Rigs of Rods reference simulations. Please refer to the supplementary attachments.

AuthorsOpenReview ↗

This response explains when vehicle fracture occurs and points to supplementary videos including a dedicated two car high speed collision case demonstrating clear vehicle structural breakage.

Thank you for the question. Component fracture of node beam vehicles is governed by whether beam element stress exceeds breaking thresholds. In several test scenarios, collision energy is mostly absorbed by obstacles. Vehicles undergo dominant plastic bending without beam element stress reaching fracture thresholds, so visible structural breakage does not appear.

In the revised supplementary material, we provide extensive simulation videos where vehicle component fracture can be clearly observed. To better illustrate this behaviour, we specifically design a high speed two car collision case showing beam failures on bumpers (see Suppl. Video S-5), chassis and other parts, with simultaneous breakage on both vehicles and obstacles. Please refer to the attached supplementary videos .

Reviewer_1qtvOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a heterogeneous simulation-rendering framework that combines node-beam vehicle dynamics with MPM-based fracture simulation to capture both vehicle structural deformation and obstacle fragmentation. Shell proxy particles are introduced to enable bidirectional collision interactions between the two representations. Experiments demonstrate visually plausible collision dynamics, including vehicle deformation, structural fracture, and debris generation.

Critical strengths
  • The proposed framework combines the complementary advantages of node-beam and MPM simulations, enabling simultaneous vehicle structural deformation and obstacle fragmentation within a unified framework.
  • The qualitative results demonstrate visually compelling collision and fracture effects in the selected simulation snapshots.
Critical weaknesses
  1. Limited methodological novelty. The proposed framework primarily integrates existing simulation and rendering techniques, including node-beam dynamics, CollisionMPM-based fracture simulation, and 3D Gaussian Splatting. In particular, the proposed shell-proxy-particle interface shares substantial conceptual similarities with the distributed interaction nodes introduced in [1]. Both approaches follow a similar pipeline involving surface sampling, grid-based contact computation, and feedback of contact forces or impulses to structural nodes, without introducing additional independent degrees of freedom. Although the proposed method adopts split-field CollisionMPM, operates on node-beam structures rather than finite elements, and integrates an MPM-3DGS fracture pipeline, these differences appear primarily implementation-oriented. The authors should more clearly distinguish their coupling formulation from [1] and demonstrate the specific methodological advances or numerical advantages of the proposed approach. Furthermore, the MPM fracture simulation and rendering components are largely inherited from existing work such as Fracture-GS [2], making the overall contribution appear more like a system-level integration of established techniques than a substantial methodological innovation.

  2. Insufficient justification of momentum consistency in the shell-proxy representation. In Section 3.1.1, the authors explicitly impose the mass-matching constraint ∑smsproxy=∑rmr\sum_s m_s^{\mathrm{proxy}}=\sum_r m_r to ensure that the total mass of shell proxy particles equals that of the original vehicle. However, this constraint alone does not guarantee that the proxy representation preserves the original vehicle's momentum. Specifically, according to Eq. (8), the proxy velocity is computed as vs=∑rwsrvr\mathbf{v}_s=\sum_r w_{sr}\mathbf{v}_r. Consequently, the total momentum represented by the shell proxies is

    =\sum_s m_s^{\mathrm{proxy}}\mathbf{v}_s =\sum_s m_s^{\mathrm{proxy}}\left(\sum_r w_{sr}\mathbf{v}_r\right), =\sum_r\left(\sum_s m_s^{\mathrm{proxy}}w_{sr}\right)\mathbf{v}_r,$$ whereas the original node-beam vehicle momentum is $$\mathbf{P}_{\mathrm{beam}}=\sum_r m_r\mathbf{v}_r.$$ To guarantee $\mathbf{P}_{\mathrm{proxy}}=\mathbf{P}_{\mathrm{beam}}$ for arbitrary nodal velocities, the following condition is required: $$\sum_s m_s^{\mathrm{proxy}}w_{sr}=m_r,\qquad \forall r.$$ However, the paper does not explicitly enforce or establish this condition. Therefore, the conservation of momentum is not guaranteed.
  3. Unclear fidelity of the reused vehicle dynamics. In Section 3 (lines 158–159), the authors state that the vehicle is represented by a node-beam formulation that captures component-level structural deformation and fracture. Furthermore, Appendix B.3 (lines 734–746) states that the vehicle's mass, beam topology, stiffness, damping, plastic yield thresholds, and breaking strengths are parsed from existing Rigs of Rods vehicle assets. However, Section 3.2 implements an explicit node-beam dynamics model with a simplified axial spring-damper force formulation in Eq. (7). This raises the question of whether the proposed simulator can faithfully reproduce the mechanical behaviour of the original vehicle simulator, particularly regarding plastic deformation and structural fracture. Reusing the original physical parameters does not necessarily guarantee equivalent dynamic responses if the underlying dynamics and constitutive formulations differ. Since retaining the calibrated mechanical behaviour of existing vehicle assets is one of the central motivations of the paper, the authors should provide quantitative comparisons between their node-beam implementation and the original vehicle simulator under identical loading conditions.

  4. Insufficient evidence supporting the claimed real-world applicability. The paper motivates the framework through high-fidelity crash simulation, autonomous-driving safety testing, and neural world modeling. However, the physical fidelity of the resulting simulations remains insufficiently demonstrated, particularly given the concerns regarding the reused vehicle dynamics discussed in Weakness 3. Moreover, the authors explicitly acknowledge in Section 5 (lines 476–477) that the material parameters are visually tuned rather than calibrated against real crash measurements. While visually plausible animations may be useful for graphics applications and synthetic data generation, visual plausibility alone does not establish physically accurate collision responses or structural damage predictions. Without validation against real-world measurements or trusted reference simulations, it is difficult to assess whether the generated scenarios are sufficiently reliable for the claimed safety-critical applications. The authors should either provide stronger physical validation or more clearly distinguish applications requiring visual realism from those requiring quantitative physical accuracy. An evaluation of whether the generated crash data benefits downstream perception or planning tasks would also strengthen the application claims.

  5. Insufficient experimental evaluation. Although the paper provides qualitative comparisons, component ablations, a user study, and basic physical sanity checks, the experimental evidence remains limited relative to the claimed contributions. Several additional evaluations would substantially strengthen the paper:

    • More relevant baselines. The comparison should include existing heterogeneous coupling approaches, particularly [1], which shares a similar surface-sampling and grid-contact strategy.
    • Validation against the original vehicle simulator. As discussed in Weakness 3, experiments should verify whether the proposed node-beam dynamics reproduce the deformation, plasticity, and fracture behaviour of the source vehicle simulator. Otherwise, the claim that reusing existing vehicle assets preserves their calibrated mechanical properties remains insufficiently supported.
    • Robustness and sensitivity analysis. The paper should systematically investigate how shell-proxy sampling density, MPM grid resolution, simulation timestep, and penetration penalty strength affect contact accuracy, numerical stability, computational efficiency, and fracture behaviour. Such experiments are particularly important because the shell-proxy interface is the main methodological contribution. Comparisons against reference deformation and damage patterns would also help establish physical fidelity.
    • More comprehensive quantitative metrics. The current evaluation relies substantially on a perceptual user study involving only 30 participants, without reporting uncertainty estimates or statistical significance. Although the paper includes momentum and energy plots, visual inspection of these curves is insufficient to rigorously establish numerical accuracy or conservation properties. The authors should report explicit quantitative measurements, such as deformation error, contact impulse error, penetration depth, linear and angular momentum deviation, and energy-balance residuals, ideally using controlled experiments with known reference solutions.
    • Dynamic visualization and temporal consistency. Although Appendix D provides additional qualitative results, the submitted material does not include accompanying video demonstrations. Since the paper focuses on dynamic simulation of high-energy collisions and structural fracture, selected static frames are insufficient to thoroughly assess temporal consistency, collision responses, fracture propagation, and dynamic plausibility. Video demonstrations would provide important additional evidence for the claimed simulation quality.
  6. Presentation and formatting issues. The manuscript contains frequent formatting problems, particularly missing spaces or hyphens in compound technical terms. For example, "shellproxyparticle" appears in line 84, and similar issues occur throughout the paper, reducing readability.

[1] Cheon, Y.-J., & Kim, H.-G. (2018). An efficient contact algorithm for the interaction of material particles with finite elements. Computer Methods in Applied Mechanics and Engineering, 335, 631–659.

[2] Wang, X., Wu, H., Song, W., & Xu, K. (2026). Fracture-GS: Dynamic Fracture Simulation with Physics-Integrated Gaussian Splatting. International Conference on Learning Representations (ICLR).

Questions for authors

Please refer to the weaknesses.

6 comments
AuthorsOpenReview ↗

This response clarifies novelty and distinctions against FEM MPM [1] coupling and Fracture GS [2]

Thank you for your observation. Shell‑proxy and the distributed interaction nodes in reference [1] share superficial overlap in a small set of high‑level procedural steps. Nevertheless, their underlying design constraints and motivations diverge substantially. Apparent step‑level resemblance does not make them comparable solutions, nor can one be adapted to the other via minor implementation tweaks.

Reference [1] targets offline mechanical analysis coupling FEM meshes with single‑field MPM. Three hard incompatibilities block its direct adoption for our heterogeneous‑injection task:

  1. Asset constraint: It natively expects FEM meshes and cannot consume node‑beam assets out‑of‑the‑box. Converting node‑beam assets to FEM discards well‑tuned component‑wise fracture behaviour from community assets. Furthermore, standard FEM formulations struggle with large‑scale deformation and massive structural fracture; mesh elements suffer severe numerical degradation under heavy distortion.
  2. Contact‑backend constraint: Split‑field Collision‑MPM relies on two separate mass‑momentum fields for contact detection, which is incompatible with the single‑field MPM contact logic adopted in [1].
  3. Rendering‑pipeline constraint: Reference [1] outputs purely mechanical simulation results and contains no logic for handling rendering artifacts between meshes and fractured 3D‑Gaussian splats.

We want to clarify an important background: existing split‑field Collision‑MPM + 3D‑Gaussian‑Splatting fracture pipelines (e.g., Fracture‑GS, PhysGaussian) built upon the NACC constitutive model require all colliding bodies to be represented as MPM particles. They provide no native interface for external non‑particle structures such as node‑beam assemblies to join the dual‑field contact solve. Fracture‑GS works under a homogeneous all‑particle assumption. Our goal is different: keep non‑particle pre‑calibrated node‑beam vehicle assets un‑particleized, and enable bidirectional collision between heterogeneous entities: node‑beam vehicles and MPM‑3DGS obstacles.

These intrinsic limitations inherited from the split‑field Collision‑MPM framework motivate our tailored shell‑proxy interface: strictly read‑only proxies for vehicle‑side surface sampling, dual‑field P2G projection, augmented with anti‑penetration correction and physics‑aware depth compositing to close the heterogeneous simulation‑rendering pipeline. Ablation validates its necessity: removing shell‑proxy increases penetration burden from 2.72 to 8.11 and produces severe contact‑failure artifacts such as door embedding into chassis (see Suppl.pdf Sec.2, Video S-2.1 and S-2.2).

While the obstacle‑side fracture constitutive model follows the NACC formulation used in Fracture‑GS, Fracture‑GS only addresses homogeneous all‑particle simulation. Our core contribution is not modifying the NACC fracture model, but solving the unaddressed heterogeneous‑coupling gap: how to inject external non‑particle node‑beam structures into the mature split‑field Collision‑MPM + 3D‑GS stack for bidirectional collision interaction. We will emphasize these underlying constraint differences rather than only superficial procedural steps in revised version.

[1] Cheon, Y.-J., & Kim, H.-G. (2018). An efficient contact algorithm for the interaction of material particles with finite elements. Computer Methods in Applied Mechanics and Engineering, 335, 631–659.

[2] Wang, X., Wu, H., Song, W., & Xu, K. (2026). Fracture-GS: Dynamic Fracture Simulation with Physics-Integrated Gaussian Splatting. International Conference on Learning Representations (ICLR).

AuthorsOpenReview ↗

This response addresses the momentum consistency concern of shell proxy, explains the engineering trade offs of the per node constraint for large scale fracture scenarios, and presents supplementary quantitative residual measurements.

Thank you for your rigorous mathematical derivation. We clarify a key design premise: the per node condition ∑smsproxywsr=mr,∀r\sum_{s} m_s^\text{proxy} w_{sr}=m_r,\forall r derived by the reviewer is a sufficient condition for exact momentum equality between the intermediate shell proxy set and node beam vehicle. However, this is not a design objective of our framework.

Shell proxy particles act as read only sampling intermediates with no standalone dynamic evolution. The full physical transfer pipeline works as follows:

  1. At each timestep, proxy positions and velocities are interpolated from underlying beam nodes.

  2. Proxies only project mass momentum onto the Collision MPM grid for contact computation.

  3. Grid level collision impulses are solved from relative velocity of two fields. Impulses are not directly copied from proxy momentum but newly computed from contact constraints.

  4. Grid impulses are scattered back to original beam physical nodes via barycentric weights. Vehicle dynamics are fully integrated within the node beam solver, and proxies receive no G2P update.

Barycentric weights wsrw_{sr} stay constant under ordinary bending and plastic deformation. Nevertheless, our framework targets high speed multi vehicle crashes with massive structural fracture and disintegration (see Video S-5). Re solving the over determined linear system would be required repeatedly whenever beam fracture modifies topology. Frequent re solves introduce non negligible computational overhead. Moreover, solving this over determined system tends to produce zero mass or even negative mass proxies, which trigger severe numerical instability during MPM P2G projection.

Balancing robustness and runtime performance for large fracture scenarios, our current implementation uniformly distributes total mass and only enforces ∑smsproxy=∑rmr\sum_{s} m_s^\text{proxy}=\sum_{r} m_r. This bidirectional mapping introduces inherent interpolation residuals. For further validation, we have implemented a variant that strictly follows the reviewer’s per node mathematical constraint. Our supplementary experiments show that the two schemes yield largely consistent results: the momentum error of our original scheme is small and controllable, and the visual difference between the two variants is negligible (see Suppl.pdf Sec.4, Video S-1.1 and S-1.2).

Original Section 3.1.1 only stated the total mass matching constraint without elaborating the above trade offs, which caused ambiguity. In revision, we will add discussions on this per node condition, its practical costs and interpolation residuals in the main text. Quantitative measurements of momentum residuals are included in supplementary material. No modification to core algorithm is needed. Developing a sampling scheme that satisfies this strict per node condition while keeping all proxies positive mass with low overhead is left for future work.

AuthorsOpenReview ↗

This response addresses concerns about node beam mechanical fidelity; we clarify parameter reuse boundaries, explain why we do not adopt the original solver, present comparative experimental results, and describe manuscript revisions.

Thank you for raising this important concern regarding node beam mechanical fidelity. We first clarify the boundary of asset reuse: we parse and import calibrated parameters from Rigs of Rods assets, including topology, beam geometry, mass, stiffness, damping, plastic yield threshold and breaking strength. However, we do not adopt the full internal solver of Rigs of Rods. The original Rigs of Rods is implemented in C++, while our full MPM simulation stack is built upon the Warp GPU framework. Directly invoking the original solver would introduce heavy overhead from frequent CPU GPU data transfer and bottleneck our heterogeneous pipeline. For this reason, we implement our own explicit GPU resident node beam solver for dynamic integration. The original manuscript wording can be misleading. We do not claim to perfectly reproduce the full dynamic response of the Rigs of Rods solver. We reuse the community calibrated physical parameter set. Even with identical material parameters, differences in integration scheme, damping model and fracture judging implementation will lead to measurable discrepancies in deformation and fracture outputs across different solvers.

Following your suggestion, we run side by side tests under identical vehicle asset and initial impact conditions. we provide qualitative visual evidence including full simulation videos and key frame snapshot comparisons (see Suppl.pdf Sec. 6, Video S-2.1 and S-2.3). Visually, we observe similar global failure patterns, component fracture sequences and deformation tendencies, while local details differ as expected from solver specific implementations.

In revision, we will rephrase misleading sentences to clearly distinguish between “reusing physical parameters from assets” and “reusing the original solver”. The qualitative videos and snapshot comparisons are added to supplementary material. We also explicitly state our target use case: generating perceptually plausible crash scenarios for autonomous driving simulation, rather than pursuing frame level exact alignment with the Rigs of Rods solver.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.