cs.LGOct 9, 2026

Continuous-Time Cloth Dynamics with Physics-Informed Neural ODEs

Authors: Yingdong Ru, Lipeng Zhuang, Zhuo He, Gerardo Aragon Camarasa

Organizations: University of Glasgow, Glasgow, United Kingdom

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

We present ContinuousCloth, a continuous-time, physics-informed neural solver for deformable cloth dynamics that enforces a spring--mass ODE within a neural trajectory model. Unlike classical time-stepping simulators that return trajectories only on a discrete integration grid, ContinuousCloth supports querying at arbitrary timestamps and computing displacement, velocity, and acceleration via automatic differentiation. We evaluate ContinuousCloth along two dimensions. In a physics-only setting, ContinuousCloth achieves trajectory accuracy comparable to high-precision solvers across six diverse boundary-condition scenarios while reducing runtime by up to 17.4 times. When fused with sparse and noisy observations, the spring--mass ODE acts as a strong inductive bias for downstream tasks: interpolation and parameter estimation. That is, ContinuousCloth reduces interpolation error by 7.3 times over a data-only neural baseline in simulation and 2.4 times on real-world cloth sequences, and achieves 2.6 times lower parameter error than a differentiable-solver baseline under high-noise conditions, demonstrating consistent robustness under realistic measurement imperfections.

Figures & tables

Figure 1: (a) Classic simulators accumulate errors on discrete grids and can become unstable, whereas ContinuousCloth represents cloth motion as a continuous function queryable at any tt . (b) Given only the first and last frames, ContinuousCloth reconstructs a physically consistent trajectory across the full intermediate horizon.
Figure 2: Overview of ContinuousCloth. We represent cloth dynamics as a continuous-time trajectory field conditioned on boundary conditions and physical parameters, and enforce spring–mass ODE constraints via physics residuals computed with automatic differentiation. ContinuousCloth can be queried at arbitrary timestamps and the same framework supports data-scarce applications, including trajectory interpolation from sparse observations and inverse physical-parameter estimation.
Figure 3: Average of per-rollout runtime across scenarios: RK4 (fixed step size) vs. our ContinuousCloth (CC) method.
MethodS1S2S3S4S5S6
PR ↓\downarrowRMSE ↓\downarrowPR ↓\downarrowRMSE ↓\downarrowPR ↓\downarrowRMSE ↓\downarrowPR ↓\downarrowRMSE ↓\downarrowPR ↓\downarrowRMSE ↓\downarrowPR ↓\downarrowRMSE ↓\downarrow
QRes(w/o MFPSG)4.39×10−34.39\times 10^{-3}7.67×10−17.67\times 10^{-1}4.88×10−34.88\times 10^{-3}5.72×10−15.72\times 10^{-1}8.13×10−28.13\times 10^{-2}4.65×10−14.65\times 10^{-1}9.31×10−29.31\times 10^{-2}3.75×10−13.75\times 10^{-1}4.92×10−24.92\times 10^{-2}7.06×10−17.06\times 10^{-1}4.91×10−24.91\times 10^{-2}7.01×10−17.01\times 10^{-1}
QRes1.17×10−31.17\times 10^{-3}1.24×10−11.24\times 10^{-1}1.73×10−31.73\times 10^{-3}1.01×10−11.01\times 10^{-1}7.98×10−27.98\times 10^{-2}4.66×10−14.66\times 10^{-1}9.17×10−29.17\times 10^{-2}3.76×10−13.76\times 10^{-1}3.93×10−23.93\times 10^{-2}6.44×10−16.44\times 10^{-1}4.11×10−24.11\times 10^{-2}6.67×10−16.67\times 10^{-1}
FLS(w/o MFPSG)4.51×10−34.51\times 10^{-3}7.93×10−17.93\times 10^{-1}4.68×10−34.68\times 10^{-3}5.55×10−15.55\times 10^{-1}4.86×10−34.86\times 10^{-3}3.90×10−13.90\times 10^{-1}4.95×10−34.95\times 10^{-3}2.73×10−12.73\times 10^{-1}5.01×10−35.01\times 10^{-3}4.72×10−14.72\times 10^{-1}5.15×10−35.15\times 10^{-3}5.41×10−15.41\times 10^{-1}
FLS4.37×10−34.37\times 10^{-3}6.68×10−16.68\times 10^{-1}4.53×10−34.53\times 10^{-3}4.39×10−14.39\times 10^{-1}4.77×10−34.77\times 10^{-3}3.64×10−13.64\times 10^{-1}4.87×10−34.87\times 10^{-3}2.55×10−12.55\times 10^{-1}4.76×10−34.76\times 10^{-3}3.84×10−13.84\times 10^{-1}5.92×10−35.92\times 10^{-3}5.34×10−15.34\times 10^{-1}
CC-MLP(w/o MFPSG)2.02×10−32.02\times 10^{-3}1.19×10−11.19\times 10^{-1}1.59×10−31.59\times 10^{-3}1.01×10−11.01\times 10^{-1}2.57×10−32.57\times 10^{-3}9.58×10−29.58\times 10^{-2}3.21×10−33.21\times 10^{-3}1.12×10−11.12\times 10^{-1}2.55×10−22.55\times 10^{-2}8.19×10−18.19\times 10^{-1}2.27×10−22.27\times 10^{-2}7.30×10−17.30\times 10^{-1}
Table 1: Quantitative comparison across S1–S6 scenarios. We report the mean physics residual (PR) in newtons and RMSE in meters over the full rollout horizon for each method. ’w/o MFPSG’ denotes an ablated variant without MFPSG. ’CC’ represents ContinuousCloth.
Figure 4: Qualitative results in the physics-only forward simulation. (a) Trajectory overlay (blue: ground truth, orange: prediction) of ContinuousCloth solver and an RK4 reference [ 39 ] under the same evaluation step size. Three scenarios are S2, S3 and S4. (b) Boundary-mask conditioning enables a single model to handle diverse constraint patterns. (c) Parameter conditioning allows the solver to generalize across different material parameters.
Figure 5: Oscillatory cloth dynamics in the four-corner pinned scenario (S4). Blue shows the high-precision RK4 reference trajectory, and orange shows the ContinuousCloth prediction. The close alignment demonstrates that our model captures the underlying spring–mass oscillations over time.
Figure 6: Qualitative interpolation results across simulation (D2 Moving clean, D3 Swing large-noise) and real data (R1 Shaking clean, R2 Fling outliers). Compared with the MLP baseline, our ContinuousCloth yields more stable and physically plausible trajectories, especially under noisy observations.
More in this topic

Deformable Object Simulation

Deformable object simulation models how soft materials such as cloth, cables, tissue, and solids change shape under forces and contact.

See all 84 papers

Explore similar work

CardsList
May 27, 2026

ClothTransformer: Unified Latent-Space Transformers for Scalable Cloth Simulation

Unified and scalable Transformers have recently achieved remarkable success in modeling diverse phenomena traditionally associated with computer graphics, such as 3D visual effects, rendering processes, and motion in videos. In this work, we take a step further by investigating whether modern Transformer techniques can tackle the challenging task of cloth simulation. To this end, we present ClothTransformer, a framework that reformulates cloth simulation as autoregressive sequence modeling in a learned latent space. Existing neural cloth simulators are largely specialized to single scenarios, intrinsically coupled to the mesh discretization, and lack robust collision handling. Our approach addresses these limitations through three contributions: (1) a unified Transformer architecture that handles diverse scenarios -- body-driven garments, robotic manipulation, and free-fall collisions -- under a single model and achieves approximately 44--9×9{\times} lower error than prior state-of-the-art methods across all scenarios; (2) a scalable latent-space formulation that compresses arbitrary-resolution meshes into a fixed-size set of latent tokens, making temporal dynamics computation independent of mesh resolution; and (3) a diverse-scenario high-fidelity penetration-free dataset of ∼{\sim}493.4k frames spanning all three settings, which enables a differentiable Continuous Collision Detection (CCD) module to suppress penetration artifacts. Project Page: https://yucrazing.github.io/clothtransformer/

Yu Zhang, Yidi Shao, Wenqi Ouyang +5

1S-Lab, Nanyang Technological University, Singapore · 2Feeling AI · University of Oxford +2

TransformerPhysics-Based SimulationDeformable Object Simulation
AskPDF
Jul 22, 2026

ODeform: Learning Continuous 4D Motion for Shape Deformation with Neural ODEs

Modeling continuous object deformation is important for many computer vision and robotics tasks, such as manipulation and simulation. Existing approaches rely on learning-based methods or physics simulators to model shape deformations. However, these approaches either use discrete time steps or are too computationally intensive for real-time applications. We present ODeform, a novel extension of Neural Ordinary Differential Equations to continuous 4D dynamics of deformable objects in 3D space. Our method transforms 3D point clouds and physical conditions (like material properties) into a unified latent space. By solving the resulting ordinary differential equations over time, we model deformations as continuous flows within this learned embedding, eliminating the need for discrete time steps while maintaining computational efficiency. We evaluate our approach on unseen physical parameter configurations, showing improved motion prediction accuracy over baseline methods. Our experiments further demonstrate a successful transfer to real 3D captured objects with novel shapes, along with effective interpolation and extrapolation of the learned dynamics. Our code and data will be made publicly available.

Yordanka Velikova, Mahdi Saleh, Liming Kuang +1

Technical University of Munich · Munich Center for Machine Learning (MCML)

Neural ODEsDeformable Object Simulation
AskPDF
Oct 7, 2026

Same Action, Different Outcome: Variability in Dynamic Cloth Manipulation

Although cloth is known to exhibit different outcomes under repeated fast dynamic motions, even when the same trajectory is applied, this variability has not yet been systematically characterized. Quantifying it is essential to assess the reliability of learned manipulation policies and the extent to which simulation can reproduce real-world behavior. To study this, we execute the same trajectory ten times across four dynamic tasks, two of which are novel, each tested with three cloths of very different properties and at up to three execution speeds, with a total of 269 recorded rollouts. For all of them, we record small marker positions on the cloth and synchronized stereo camera. We then formalize different metrics to quantify variability, and our results show how it is significant in every test condition, in most cases one to three orders of magnitude above the repeatability inherent to the robot and sensing noise. Our results also show variability is driven mainly by the cloth physical properties but also grows with speed. By replaying all the rollouts in four calibrated modern cloth simulators, we show that none of them can reproduce the variability magnitude we observed in real cloth, nor its ordering. Together with our analysis, we publish the dataset with synchronized OptiTrack, vision and robot logs and their corresponding simulator twins.

Mahed Dadgostar, Guillem Alenyà, Júlia Borràs

Institut de Robòtica i Informàtica Industrial, CSIC-UPC, Spain

Deformable Object ManipulationRobotic ManipulationDynamic Robotic ManipulationDeformable Object Simulation
AskPDF