SENC: Handling Self-collision in Neural Cloth Simulation

TL;DR

SENC addresses the persistent problem of cloth self-collision in self-supervised neural cloth simulation by introducing a Global Intersection Analysis (GIA) based penetration-volume loss and a self-collision-aware Graph Neural Network that builds distance-based edges to capture non-local collisions. The model is trained with a composite energy that includes self-collision, body-cloth collision, stretching, bending, external forces, inertia, and friction, enabling end-to-end learning without ground-truth garment data. Experiments on AMASS sequences show significant reductions in self-collision across diverse garments while preserving animation quality, outperforming state-of-the-art self-supervised methods like HOOD, SNUG, and NCS. Ablation studies validate the necessity of the volume-based penetration loss and the augmented self-collision graph, and qualitative results demonstrate robust performance under varied external forces. This work advances practical neural cloth simulation by enabling robust, self-supervised handling of self-collision with potential extensions to multi-layer clothing and deformable characters.

Abstract

We present SENC, a novel self-supervised neural cloth simulator that addresses the challenge of cloth self-collision. This problem has remained unresolved due to the gap in simulation setup between recent collision detection and response approaches and self-supervised neural simulators. The former requires collision-free initial setups, while the latter necessitates random cloth instantiation during training. To tackle this issue, we propose a novel loss based on Global Intersection Analysis (GIA). This loss extracts the volume surrounded by the cloth region that forms the penetration. By constructing an energy based on this volume, our self-supervised neural simulator can effectively address cloth self-collisions. Moreover, we develop a self-collision-aware graph neural network capable of learning to handle self-collisions, even for parts that are topologically distant from one another. Additionally, we introduce an effective external force scheme that enables the simulation to learn the cloth's behavior in response to random external forces. We validate the efficacy of SENC through extensive quantitative and qualitative experiments, demonstrating that it effectively reduces cloth self-collision while maintaining high-quality animation results.

Figures (8)

• Figure 1: Our method effectively addresses cloth self-collision, compared to existing state-of-the-art neural cloth simulator grigorev2023hood. Note the inner side of the cloth is painted pink.
• Figure 2: Method overview.
• Figure 3: An example showing how to close the garment.
• Figure 4: The leftmost picture shows the case of the loop vertex baraff2003untangling (enclosed by a red box), where one intersection point is the vertex shared by the two triangles. The middle two pictures show the other cases of two triangles intersecting, generating two intersection points (green circles) respectively. The two green points in these three cases become neighbors in the intersection path (the red line is a segment of the intersection path). These three cases are all the possible cases of two triangles intersecting. The right-most picture shows one intersection point can be represented by the three vertices (yellow circles) of the face using barycentric coordinates.
• Figure 5: Here we show two cases of self-collision, where the penetration volume is composed of one (a)(b) and two (c)(d) interaction paths. (a) and (b) show a severely bent elbow with self-collisions happening inside the elbow. Figure (a) shows the intersection path. Figure (b) shows the vertices inside the self-collision and the intersection path after unbending the arm. Similarly, (c) and (d) show a case where a torus intersects with itself, resulting in two intersection paths and two separate penetration surfaces.