PhySPRING: Structure-Preserving Reduction of Physics-Informed Twins via GNN

Yixiong Jing, Xingyuan Chen, Guangming Wang, Olaf Wysocki, Haibing Wu, Brian Sheil · 2026

Read PhySPRING: Structure-Preserving Reduction of Physics-Informed Twins via GNN by Yixiong Jing, Xingyuan Chen, Guangming Wang, Olaf Wysocki, Haibing Wu, Brian Sheil (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-07687.txt.

Introduction

Yixiong Jing and five co-authors simplify physics-based digital twins, the spring-mass models used to predict how real objects deform and move for robotics. Twins rebuilt from vision inherit the reconstruction's fine resolution, far more detail than the task dynamics need, which makes repeated simulation rollouts needlessly expensive.

Their method, PhySPRING, uses a graph neural network to learn coarser graph versions of the twin together with their mechanical parameters, merging nodes whose learned dynamic responses are similar. Every reduced layer remains an explicit spring-mass system. On the PhysTwin benchmark it improved reconstruction and prediction accuracy while running up to about two and a third times faster. Substituted into robot policy evaluation, the reduced twins kept manipulation success rates comparable.

Catalog record

CREATORYixiong Jing, Xingyuan Chen, Guangming Wang, Olaf Wysocki, Haibing Wu, Brian Sheil
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.07687

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

PhySPRING: Structure-Preserving Reduction of Physics-Informed Twins via GNN Yixiong Jing, Xingyuan Chen, Guangming Wang, Olaf Wysocki, Haibing Wu, Brian Sheil Abstract Physics-based digital twins aim to predict the dynamics of real-world objects under interaction, enabling real-to-sim-to-real applications in robotics. Current approaches reconstruct such twins as explicit physical models (such as spring-mass systems) to predict the dynamics, but the resulting models often inherit the resolution of the visual reconstruction rather than being reduced to the physical complexity required to reproduce task-relevant dynamics. This mismatch introduces redundant topology, making repeated forward-dynamics rollouts unnecessarily expensive. To address this challenge, we present PhySPRING, an fully differentiable GNN-based method to reduce complexity in spring–mass digital twins. PhySPRING jointly learns a hierarchy of coarsened graph topologies and their mechanical parameters from observations. At each reduction level, PhySPRING merges nodes with similar learned dynamic responses to optimize the topology, while maintaining every reduced layer as an explicit spring–mass system. On the PhysTwin benchmark, PhySPRING improves dense reconstruction and prediction accuracy over PhysTwin, while reduced models retain stable physical and visual fidelity with up to a 2.30 × \times speed-up. We further demonstrate the effectiveness of PhySPRING in a Real2Sim robot policy-evaluation pipeline, where the reduced models are substituted zero-shot into ACT and π 0 \pi_{0} evaluations, maintaining comparable manipulation success rates across downsampling levels while improving action-sampling effectiveness. Together, PhySPRING enables efficient and structure-preserving spring–mass reduction without sacrificing fidelity or robotic utility. 1 Introduction Reconstructing physics-informed digital twins from real-world objects enables accurate prediction of how objects move and deform under interaction 9 ; 20 , supporting downstream applications such as robot manipulation planning 29 ; 23 ; 25 ; 18 ; 16 and data generation for policy learning 8 ; 34 ; 49 . However, precise physical reconstruction from observations remains non-trivial. Previous graphics and engineering pipelines rely on carefully designed constitutive models and discretizations 40 ; 42 that assume comprehensive prior knowledge of object physics, making them expensive and difficult to generalize. Leveraging recent advances in differentiable physics 31 and 3D reconstruction 21 , PhysTwin 20 learns a spring–mass physical model directly from RGB-D video via differentiable simulation, with appearance reconstructed by 3D Gaussian Splatting (3DGS) 21 . However, current physics-driven reconstruction approaches directly inherit physical simulation resolution from visual reconstruction resolution. The resolution needed to represent surface geometry and appearance can be much higher than the number of degrees of freedom needed to reproduce task-relevant dynamics for downstream robotics applications. Planning and policy evaluation require repeated forward-dynamics rollouts under varied interactions, so simulating a dense physical representation inside the robot loop can introduce unnecessary computational overhead. Therefore, we identify that the core challenge is to obtain a reduced-order physical representation that accelerates simulation while preserving physical fidelity and visual/geometric consistency. One possible approach is to learn a reduced topology with GNN-based simulation models 36 ; 11 . These methods relax discrete topology-reduction decisions into end-to-end trainable architectures, enabling adaptive graph representations that reduce physical-system complexity for dynamics prediction. However, they typically require ground-truth dynamic states from numerical solvers and learn black-box forward dynamics rather than explicit physical parameters. As a result, the learned reduced models do not preserve the same physical structure as the original system, and therefore cannot be directly deployed in external simulators under unseen interactions such as actions from robot arms. Classical model order reduction (MOR) provides a physics-based approach for reducing physical systems through nodal clustering based on analytically computed dynamic similarity 10 ; 1 . This preserves the second-order mechanical form of the system and produces reduced graphs that can be integrated by standard numerical solvers. However, these methods assume known physical matrices and material parameters, which requires pre-optimization with observation data. Moreover, computing dynamic-similarity matrices over thousands of reconstructed nodes is computationally expensive, limiting their practicality for downstream applications. To address this gap, we present PhySPRING a structure-preserving reduction method for spring–mass digital twins reconstructed from sparse-view RGB-D observations as shown in Fig. 1 . PhySPRING combines GNN-based topology reduction with differentiable physics so that the reduced model remains an explicit physical system rather than a black-box dynamics predictor. Our contributions are as follows. (i) We introduce an end-to-end physics-informed GNN that jointly learns topology reduction and mechanical parameters from observation-supervised rollouts. On PhysTwin, PhySPRING improves dense reconstruction/prediction over PhysTwin, and reduced models achieve controlled fidelity–efficiency trade-offs with up to 2.30 \times speed-up. (ii) We preserve the explicit spring–mass form of the digital twin across all reduction levels. Galerkin projection induces coarse spring-mass representation while a GNN decoder refines the resulting mechanical parameters through differentiable rollouts. (iii) We evaluate PhySPRING in a Real2Sim 49 policy-evaluation pipeline. The reduced models can be substituted zero-shot into ACT 52 and \pi_{0} 5 without retraining, preserving manipulation success rate, while improving action-sampling throughput by up to 1.23 \times over the original model. 2 Related works Physics-driven 4D Reconstruction Recent 4D reconstruction methods achieve high-fidelity geometry and appearance from visual observations 21 ; 32 ; 37 ; 35 ; 44 ; 27 . However, they are primarily descriptive and lack physical priors for predicting future dynamics or supporting grounded interaction. To improve physical fidelity, several works combine reconstructed scenes with explicit numerical simulation 46 ; 51 ; 12 ; 45 or infer material properties from video via differentiable physics 24 ; 14 . These methods often rely on predefined materials or implicit geometry and tend to prioritize visual plausibility over accurately recovering the dynamic behavior of real-world objects. To jointly recover both 3D structure and dynamics, other methods 53 ; 20 ; 50 reconstruct physically grounded spring-mass models from observations to create physics-based digital twins. Nevertheless, these models inherit the geometric resolution of the observations, yielding redundant and over-constrained systems whose simulation cost limits their use inside robot planning loops. Figure 1: We propose PhySPRING, a novel method for structure-preserving order reduction of physics-informed digital twins reconstructed from sparse-view RGB-D observations. An end-to-end multi-hierarchical GNN jointly predicts the reduced topology and the per-edge mechanical parameters at every level l . Since each reduced-order model remains an explicit second-order spring–mass system, it is directly deployable in a downstream Real2Sim robot-arm manipulation simulator 49 . Model order reduction Classical model order reduction (MOR) compresses a dynamical system with known mechanical properties into a lower-order surrogate whose dynamic response closely approximates that of the original system 33 ; 15 ; 2 ; 41 ; 43 ; 4 . For second-order linear time-invariant systems, second-order Gramians reduce system complexity while preserving the dynamic structure (e.g., same form of governing equations) 7 ; 38 ; a previous method 10 adapts the method on spring-mass systems to conduct MOR via nodal clustering based on controllability-Gramian similarity, yielding reduced graphs compatible with a standard numerical integrator. However, these methods require the mechanical properties of the system to measure nodal dynamic similarity, which cannot be directly acquired from observation. The resulting topology, once optimized, remains static throughout the dynamic sequence and is not updated temporally. Moreover, computing the dynamic-similarity matrices introduce expensive matrix operations, which become prohibitively costly as the system size grows. Graph topology optimization for physical simulation NN-based simulators ( 39 ; 36 ; 26 ) learn structural dynamics from ground-truth solver rollouts, offering substantial speedups over traditional numerical

Read the complete text