High-quality, intersection-free simulated cloth data removes a common bottleneck for learning-based cloth simulators: noisy or interpenetrating ground truth. This dataset supplies dense, mesh-level trajectories suitable for CCD-aware losses and models that operate in vertex/mesh latent spaces.
What Sets It Apart
- Penetration-free ground truth: 2,056 trajectories × 240 frames (493,440 frames total), generated with a GPU incremental-potential contact solver (GIPC), enabling training with Continuous Collision Detection (CCD) losses.
- Diverse scenarios: three subsets — human garments on animated SMPL bodies, robotic gripper manipulation over 1,000+ garments, and 1,000 diverse object collision drops — supporting transfer and generalization experiments.
- Mesh-level realism: per-frame vertex positions and displacements, triangle/edge topology, and redundant collision triangle fields; mesh resolutions vary (1k–4k verts) so models can be tested across scales.
- Reproducible parameters: material and solver priors (stretching, bending, density, friction) provided, with CC BY 4.0 license and provenance notes for third-party geometry.
Who it's for + Trade-offs
Great fit if you develop or evaluate learning-based cloth simulators, mesh latent-space transformers, or CCD-aware loss functions and need large, collision-free ground truth at mesh resolution. Look elsewhere if you require real-world capture data, deformable objects beyond cloth, or very high-resolution textile microstructure — this dataset focuses on simulated cloth physics and mid-scale meshes (1k–4k vertices).
