PhysSkin: Real-Time and Generalizable Physics-Based Animation via Self-Supervised Neural Skinning

Yuanhang Lei; Tao Cheng; Xingxuan Li; Boming Zhao; Siyuan Huang; Ruizhen Hu; Peter Yichen Chen; Hujun Bao

cs.GRMar 24, 2026

PhysSkin: Real-Time and Generalizable Physics-Based Animation via Self-Supervised Neural Skinning

Authors: Yuanhang Lei, Tao Cheng, Xingxuan Li, Boming Zhao, Siyuan Huang, Ruizhen Hu, Peter Yichen Chen, Hujun Bao,

Organizations: State Key Laboratory of CAD&CG, Zhejiang University · BIGAI · Shenzhen University · University of British Columbia

Abstract

Achieving real-time physics-based animation that generalizes across diverse 3D shapes and discretizations remains a fundamental challenge. We introduce PhysSkin, a physics-informed framework that addresses this challenge. In the spirit of Linear Blend Skinning, we learn continuous skinning fields as basis functions lifting motion subspace coordinates to full-space deformation, with subspace defined by handle transformations. To generate mesh-free, discretization-agnostic, and physically consistent skinning fields that generalize well across diverse 3D shapes, PhysSkin employs a new neural skinning fields autoencoder which consists of a transformer-based encoder and a cross-attention decoder. Furthermore, we also develop a novel physics-informed self-supervised learning strategy that incorporates on-the-fly skinning-field normalization and conflict-aware gradient correction, enabling effective balancing of energy minimization, spatial smoothness, and orthogonality constraints. PhysSkin shows outstanding performance on generalizable neural skinning and enables real-time physics-based animation.

Figures & tables

Figure 2: Given a static 3D shape, we first sample volumetric cubature points for animation and surface points for shape encoding. A shape encoder processes the surface points to produce shape latents, from which a set of learnable queries extract handle latents via cross-attention. Subsequently, spatial queries attend to the handle latents through another cross-attention to derive point-wise features, which are decoded into neural skinning fields. A geometry-only, physics-informed learning strategy optimizes the network in a self-supervised manner under physical and orthogonal constraints, enabling the learned skinning fields to support real-time, physics-based animation.

Figure 3: Evolution of neural skinning fields during optimization. Starting from disordered initial representations, our physics-informed self-supervised learning progressively organizes them into physically consistent, geometrically orthogonal, and spatially smooth skinning results. Each skinning weight $i$ is scaled by $\max(abs(\bm{W}_{i}))$ to fall within $[-1,1]$ and centered around 0.

Method	$\Omega_{\text{orth}}\times 10^{-2}$ $\downarrow$	$\kappa_{\log}$ $\downarrow$	$H_{\text{spec}}$ $\uparrow$
RigNet [ 52 ]	0.5324	2.7997	0.9762
M-I-A [ 15 ]	1.4098	27.7357	0.7224
Anymate [ 11 ]	1.5737	2.6093	0.9682
Puppeteer [ 43 ]	0.5615	5.5605	0.9798
Ours	0.0033	1.0453	0.9999

Table 1: Quantitative comparison on RigNet [ 52 ] dataset.

Figure 4: Qualitative comparisons on RigNet [ 52 ] dataset. We visualize combined skinning fields obtained by blending all skinning weights together, providing an overall view of deformation influences and smoothness compared with baselines [ 52 , 15 , 11 , 43 ] .

Method	$\Omega_{\text{orth}}\times 10^{-2}$ $\downarrow$	$\kappa_{\log}$ $\downarrow$	$H_{\text{spec}}$ $\uparrow$
RigNet [ 52 ]	0.5130	4.5417	0.9648
M-I-A [ 15 ]	1.5736	27.9573	0.7006
Anymate [ 11 ]	5.3520	4.9221	0.8858
Puppeteer [ 43 ]	1.9528	1.4317	0.9799
Simplicits [ 36 ]	0.2621	1.5205	0.9941
Ours	0.0098	1.0460	0.9997

Table 2: Quantitative comparison on ShapeNet dataset [ 5 ] .

Figure 5: Qualitative results on unseen 3D shapes from the RigNet [ 52 ] test set. We visualize blended skinning fields to demonstrate our method’s generalization to novel 3D shapes. All results are produced by a single unified PhysSkin model.

Figure 6: Qualitative comparisons on ShapeNet [ 5 ] dataset. We compare our methods with Simplicits [ 36 ] , Anymate [ 11 ] , and Puppeteer [ 43 ] . Zoom in for more details.

Figure 7: Qualitative skinning results of our method on unseen objects from the ShapeNet [ 5 ] dataset. Zoom in for more details.

Figure 8: Skinning results for the deformable Lego shape family, where the deformable geometries are from Mani-GS [ 14 ] .

Figure 9: Animation results on ShapeNet [ 5 ] mesh objects.

Figure 10: Animation results on RigNet [ 52 ] mesh objects.

Figure 11: Animation results on 3DGS models [ 34 ] .

3D Shape	Vertex Count	FEM [ 7 ] Step Cost (ms)	MPM [ 56 ] Step Cost (ms)	Ours Step Cost (ms)
Airplane	10K	79.83	141.83	12.26
Bag	121K	3012.47	233.79	13.39
Camera	80K	2121.02	203.38	12.52
Chair	52K	1772.70	178.46	12.48
Pillow	127K	3170.93	251.81	13.74

Table 3: Comparison of per-step simulation cost for physics-based animation with full-space simulators FEM [ 7 ] and MPM [ 56 ] .

Config.	$\Omega_{\text{orth}}\times 10^{-2}$ $\downarrow$	$\kappa_{\log}$ $\downarrow$	$H_{\text{spec}}$ $\uparrow$
$\mathit{w/\!o}$ Skinning Normalization	6.5533	8.5492	0.8113
$\mathit{w/\!o}$ ONI Layer	0.0081	1.0844	0.9997
$\mathit{w/\!o}$ ConFIG Optimization	8.9247	11.8595	0.7594
$\mathit{w/\!o}$ $\mathcal{L}_{\text{orth}}$	100.0	29.18	NaN
$\mathit{w/\!o}$ $\mathcal{L}_{\text{smooth}}$	0.0050	1.0567	0.9998
Full Model	0.0033	1.0453	0.9999

Table 4: Ablation studies on the RigNet [ 52 ] dataset.

Figure 12: Top: Trained without $\mathcal{L}_{\text{pot}}$ . Bottom: Trained with $\mathcal{L}_{\text{pot}}$ .

Physics-Based Simulation

Physics-based simulation models physical systems with equations and numerical solvers, increasingly augmented by learned models for speed and control.

See all 228 papers

PhysSkin++: Real-Time and Generalizable Physics-Based Animation via Self-Supervised Neural Skinning
Oct 19, 2026Yuanhang Lei, Baofei Liu, Xingxuan Li +9Physics-Based Simulation Physics-Informed ML
State Key Laboratory of CAD&CG, Zhejiang University · Ant Group, Hangzhou 310000, China · Beijing Institute for General Artificial Intelligence (BIGAI) +2
LUNA: Learning Universal 3D Human Animation Beyond Skinning
Jun 30, 2026Peng Li, Rawal Khirodkar, Junxuan Li +6Human Motion Generation Human Avatar Animation
The Hong Kong University of Science and Technology · Work was done during an internship at Meta. · Codec Avatars Lab, Meta
AniGen: Unified $S^3$ Fields for Animatable 3D Asset Generation
Apr 9, 2026Yi-Hua Huang, Zi-Xin Zou, Yuting He +6Articulated 3D Asset Generation 3D Asset Generation
The University of Hong Kong, China · VAST, China · The Chinese University of Hong Kong, China +1

Oct 19, 2026J/K move · Enter open · S save

PhysSkin++: Real-Time and Generalizable Physics-Based Animation via Self-Supervised Neural Skinning

Yuanhang Lei, Baofei Liu, Xingxuan Li, Tao Cheng+8

State Key Laboratory of CAD&CG, Zhejiang University · Ant Group, Hangzhou 310000, China · Beijing Institute for General Artificial Intelligence (BIGAI)+2

Physics-Based Simulation Physics-Informed ML Reduced-Order Modeling

Open paper AskPDF arXiv

Generalizing real-time physics-based animation across diverse 3D shapes and spatial discretizations remains challenging. In this paper, we introduce PhysSkin++, a physics-informed framework that learns continuous neural skinning fields from static geometries to construct compact deformation subspaces. Specifically, we propose a Neural Skinning Fields Autoencoder that predicts shape-dependent deformation subspaces without per-object network training. To learn these subspaces without simulation trajectories, we develop a Physics-Informed Self-Supervised Learning strategy that combines skinning-field normalization with conflict-aware gradient correction, stabilizing joint optimization under physical and orthogonality constraints. The learned representation supports efficient reduced dynamics with user control and contact interactions, together with direct animation binding to meshes, point clouds, and 3D Gaussians. For prescribed rig animation, we further introduce a self-supervised complementarity objective to learn complementary elastic responses without enforcing complementarity constraints at runtime. Experiments demonstrate cross-shape generalization of the learned skinning fields, real-time reduced simulation, and applications to controllable, collision-aware, and rig-complementary animation.

Abstract

Physics-Based Simulation

Explore similar work

PhysSkin++: Real-Time and Generalizable Physics-Based Animation via Self-Supervised Neural Skinning

LUNA: Learning Universal 3D Human Animation Beyond Skinning

AniGen: Unified S3S^3S3 Fields for Animatable 3D Asset Generation

AniGen: Unified $S^3$ Fields for Animatable 3D Asset Generation