PaperlayerView exact revision on arXiv

Source-linked AI summary

Explicit Frictional Contact Modeling in Neural Garment Simulation

Thanh-Nguyen Truong, Vu-Hoang Tran, Ching-Chun Huang

arXiv:2610.33492v1cs.CVcs.GRcs.LG

TL;DR

Neural garment simulators leave body–garment friction implicit or as a fixed training-time penalty, without freely setting µ at inference. The paper adds a Modular Contact Operator after frozen prediction to apply explicit Coulomb friction without retraining or an iterative solve. On HOOD and ContourCraft, sticking rises and slip falls with µ while penetration stays low and stretching tracks the reference at substantially lower runtime cost.

  • Problem

    Neural garment predictors lack an explicit per-contact Coulomb condition and freely variable inference-time µ, while coefficient-specific retraining leaves the tested backbones nearly insensitive to µ.

    Source: §2.3 · p. 2, §3.2 · p. 3

  • Method

    A Modular Contact Operator attached after frozen prediction estimates normal load, applies a closed-form Coulomb correction, and propagates induced motion through the cloth without retraining or an iterative solve.

    Source: §1 · p. 2

  • Results

    On HOOD and ContourCraft, the operator raises sticking and lowers slip monotonically in µ, with low penetration and reference-consistent stretching at 8.1× lower end-to-end cost.

    Source: §1 · p. 2, §5.2.1 · p. 6

  • Takeaways & Limitations

    The tested frozen predictors can receive explicit, inference-time-controlled friction while retaining the one-pass prediction and avoiding an iterative cloth solve.

    Source: §1 · p. 2

    The operator remains a one-pass body–garment approximation; it does not model cloth–cloth friction, guarantee collision-free trajectories, or recover membrane-tension-induced normal traction.

    Source: §6 · p. 8

Abstract

from arXiv · show

Graph neural networks (GNNs) simulate garments efficiently through physics-based self-supervision, but leave body-garment friction implicit or inside training energies, with no friction coefficient $\mu$ to set at inference; retraining the tested GNNs with a coefficient-specific friction term barely responds to $\mu$. Optimization-based solvers decide each contact's stick-slip state from the reaction that governs it, but only by re-optimizing the garment state at every step; a friction term in a one-pass training objective acts as a fixed-weight penalty on sliding, and the trained weights apply no per-contact Coulomb condition at inference. We therefore resolve friction explicitly after prediction: a Modular Contact Operator, attached to a frozen predictor without retraining or an iterative solve, estimates the normal load, applies a closed-form Coulomb correction along a path across body triangles, and propagates the induced motion through the cloth as a linearized membrane wave. On HOOD and ContourCraft, it raises sticking and lowers slip monotonically in $\mu$, tracking a reference solver's response while coefficient-specific retraining stays flat, with low penetration and reference-consistent stretching at an eighth of the solver's runtime.

1. Introduction

Neural garment simulators improve geometric feasibility but leave body–garment tangential friction under-specified and lack inference-time control of a Coulomb coefficient. The paper adds a post-prediction contact operator that equips frozen predictors with explicit friction, producing coefficient-dependent responses while retaining low penetration and reference-consistent stretching at lower runtime cost.

1. Introduction

  • Neural garment methods improve normal geometric feasibility, but their tangential response—whether cloth sticks, slips, remains supported, or migrates—remains under-specified.Classical solvers use Coulomb friction, whereas neural simulators leave friction implicit or penalize sliding during training with µ fixed as a hyperparameter.Source: §1 · p. 1
  • The Modular Contact Operator estimates normal load, applies a closed-form Coulomb correction along a body-mesh path, and propagates induced motion through the cloth as a linearized membrane wave.It attaches after prediction, leaving the predictor frozen and µ available at inference without retraining or an iterative solve.Source: §1 · p. 2
  • The operator raises sticking and lowers slip monotonically in µ on HOOD and ContourCraft, while keeping penetration low and stretching consistent with the reference at 8.1× lower end-to-end cost.Coefficient-specific retraining of the tested native friction candidates remains nearly flat; the operator is validated with component ablations and normal-load comparisons.Source: §1 · p. 2
  • Friction is poorly suited to a whole-garment training objective because it acts only where garment and body touch.The paper therefore places frictional response outside that objective.Source: §1 · p. 2

2. Related work

Classical cloth solvers provide normal-load-dependent Coulomb friction and conditional stick–slip behavior through iterative contact solves, while learned garment methods generally lack explicit, controllable body–garment friction. Existing neural approaches either omit tangential contact or encode sliding as a fixed training-time penalty.

2.1. Contact and friction in cloth solvers

  • Iterative cloth solvers enforce normal-load-dependent Coulomb friction and conditional stick–slip behavior, but revise deformation and contact variables at each step, keeping per-frame cost far from interactive rates.Recent work improves solver robustness, speed, or differentiation for inverse problems.Source: §2.1 · p. 2

2.2. Learned garment animation and draping

  • Learned garment animation and draping methods efficiently reproduce appearance or equilibrium, but body–garment interaction is learned from data or handled by normal collision constraints rather than an explicit dynamic friction law.Consequently, whether garments stick or slip is neither controllable nor bounded by normal load in these methods.Source: §2.2 · p. 2

2.3. Self-supervised neural garment dynamics

  • Self-supervised neural garment simulators do not provide an explicit normal-load-dependent Coulomb bound with conditional stick–slip behavior and freely variable inference-time µ.HOOD’s sliding-work objective uses a gravity-based load proxy and fixes µ during training; coefficient-specific retraining leaves tested backbones nearly insensitive to µ.Source: §2.3 · p. 2

3. Background and problem formulation

The background contrasts iterative optimization, which decides stick–slip from contact reactions, with one-pass neural prediction, whose training-time friction penalty does not impose a per-contact Coulomb condition at inference. The paper keeps the fast predictor and restores explicit friction after prediction.

3.1. Optimization-based frictional contact

  • Coulomb friction bounds each contact’s tangential force by µ times its compressive normal reaction; sticking has zero tangential relative velocity, while sliding lies on the bound and opposes motion.The contact-local bound varies with the normal reaction at each vertex.Source: §3.1 · p. 2
  • Optimization-based simulators minimize a one-step incremental potential over the current state, combining inertia, elastic, gravity, and contact energies with a friction potential.The energy forms depend on the method and the recent states and body motion.Source: §3.1 · p. 2
  • In the described IPC formulation, the friction term depends on reaction and tangent-frame values refreshed from the new state, and minimization repeats until the step residual is met.The per-step iterative solve incurs energy, derivative, and linear-system evaluations across iterates and lag updates.Source: §3.1 · p. 3

3.2. Self-supervised neural garment simulation

  • Self-supervised neural garment simulators learn autoregressive transitions without solver-generated target trajectories, using physical objectives evaluated at network predictions.The graph network communicates deformation along cloth edges and body interaction through proximity relations.Source: §3.2 · p. 3
  • The neural friction term penalizes approximate tangential displacement using a gravity-based load proxy, while inference replaces per-step minimization with one network evaluation.The proxy uses vertex mass and local body slope; amortization enables fast autoregressive rollout.Source: §3.2 · p. 3
  • The training penalty uses a fixed µ-weighted load proxy that the network never computes, and the shared trained parameters trade sliding cost against other energies across contacts and training states.Because friction applies only to contacting cloth while the network predicts the garment as a whole, the contacting fraction’s friction may remain unresolved.Source: §3.2 · p. 3
  • At inference, a single predictor evaluation does not compare its output with the Coulomb bound or decide stick–slip per contact.The coupled loading induced in surrounding cloth is also absent from the prediction.Source: §3.2 · p. 3

3.3. Explicit contact modeling after prediction

  • The design retains one-pass prediction and restores the stick–slip decision afterward, with the frozen predictor’s tentative state corrected by the runtime-µ operator.The operator estimates reaction, applies the bound per contact, keeps the endpoint along the body surface, and carries corrected motion into surrounding fabric without iterative minimization or retraining.Source: §3.3 · p. 3

4. Modular contact operator

The Modular Contact Operator corrects a frozen predictor’s tentative state through four stages: estimating normal load, applying a Coulomb correction, traversing the body surface, and propagating motion through the cloth.

4.1. Overview

  • The operator corrects each tentative prediction by estimating load, applying Coulomb friction, traversing the body surface, and propagating induced motion through the fabric.At µ = 0, every vertex retains the baseline prediction exactly.Source: §4.1 · p. 3, Figure 2 · p. 4, §4.5 · p. 5

4.2. Normal-Load Estimation

  • The normal-load estimate retains the compressive component of a gravity-compensated force residual and uses proposal kinematics rather than an exact coupled-system reaction.It captures gravitational support by surface orientation and impact load from normal deceleration; the estimate requires no barrier-stiffness calibration.Source: §4.2 · p. 4

4.3. Closed-Form Coulomb Correction

  • A closed-form Coulomb correction converts the friction bound into an arrest length, producing sticking or residual slip without a contact solve.The correction allows runtime control of µ; its static–sliding transition is regularized by a velocity-scale band.Source: §4.3 · p. 4

4.4. Geometry-Aware Contact Traversal

  • Geometry-aware traversal follows crossed body triangles, preserving surface arc length where a straight ambient chord could cut through the body.The resulting path is straight in the unfolded face strip, but is not claimed to be a globally shortest geodesic.Source: §4.4 · p. 5, Figure 3 · p. 5

4.5. Friction-Induced Motion Propagation

  • Friction-Induced Motion Propagation carries contact corrections through the surrounding fabric as a material-dependent wave, removing edge displacement discontinuities.A per-edge strain budget limits added-field strain, and a final clamp restores any body gap reduced by propagation.Source: §4.5 · p. 5

5. Experimental results

Across HOOD and ContourCraft, the operator produces a friction response that changes with μ in the reference direction, unlike coefficient-specific friction-energy retraining. Component ablations and runtime measurements assess garment quality and the cost of adding this response.

5.1. Experimental setup

  • The evaluation compares HOOD and ContourCraft operator attachments against native Efriction retraining on 40 matched garment-motion sequences spanning μ ∈ [0, 0.6].The operator attaches at inference to a backbone retrained without a friction term; native Efriction retrains one checkpoint per μ.Source: §5.1 · p. 6, §5 · p. 6

5.2.1. Quantitative results

  • Across both backbones, the operator increases sticking and reduces slip monotonically with μ, tracking PPF’s response while native Efriction retraining stays nearly flat.The operator’s slip falls less steeply than PPF’s; the shared trend supports transfer between the two tested architectures.Source: §5.2.1 · p. 6, Figure 5 · p. 7
  • The operator has lower mean penetration than native Efriction retraining on each backbone, while its stretching-energy ratio rises with μ as PPF’s does.Native Efriction candidates remain approximately flat in stretching-energy ratio.Source: §5.2.1 · p. 6, Table 1 · p. 6

5.2.2. Qualitative results

  • In two garment-motion examples, the operator follows PPF’s response to higher μ, while friction-energy training changes weakly or moves the cuff in the opposite direction.The examples show a tank-top armhole retaining more forward displacement and a cuff moving upward with the raised forearm.Source: §5.2.2 · p. 6, Figure 6 · p. 7
  • On a retrained ContourCraft backbone, raising μ keeps more of an unseen skirt supported by a raised leg and keeps an unseen hood on the head.The garments and motions were absent from training; these examples demonstrate coefficient control rather than reference agreement.Source: §5.2.2 · p. 6, Figure 7 · p. 8

5.3. Ablation study

  • Adding traversal nearly halves mean body penetration without measurable change in stretching-energy deviation from the reference.Traversal walks corrected endpoints across the body surface and completes them in the normal direction under a non-penetration clamp.Source: §5.3 · p. 7, Table 2 · p. 8
  • At μ = 0.3 and 0.6, propagation without strain limiting increases stretching deviations and visibly over-stretches nearby fabric; the complete operator has the lowest mean and median deviations.At μ = 0.1, all four ablation rows remain closely comparable.Source: §5.3 · p. 7, Table 2 · p. 8, Figure 8 · p. 8

5.4. Runtime

  • At μ = 0.6, the operator runs a sequence in 27.2 s versus PPF’s 220.4 s, or 8.1× faster end-to-end; marginal per-frame cost is 86 ms versus 481 ms.The reported end-to-end comparison includes PPF’s fixed settling prologue.Source: §5.4 · p. 8

6. Conclusion

The Modular Contact Operator adds explicit, runtime-controllable Coulomb friction after prediction while retaining frozen one-pass predictors. Across ContourCraft and HOOD, it produces a friction-dependent response with low penetration and stretching trends consistent with the reference solver, but remains a body–garment approximation with stated modeling limits.

  • Across ContourCraft and HOOD, the operator produces a µ-dependent frictional response while keeping penetration low and stretching trends consistent with the reference solver.Source: §6 · p. 8
  • The operator applies normal-load-dependent Coulomb friction after prediction, resolving stick–slip in closed form and transporting the induced motion through surrounding cloth.It attaches to frozen neural garment predictors and provides runtime control of µ.Source: §6 · p. 8
  • The one-pass body–garment approximation does not model cloth–cloth friction, guarantee collision-free trajectories, or recover membrane-tension-induced normal traction.The wave transports corrected motion without resolving reactions; coupled elastic contact forces would require a local solve or force supervision.Source: §6 · p. 8

Related papers

Loading 2610.33492v1…