YuCrazing1/ClothTransformer-dataset

ClothTransformer Dataset

Paper (arXiv:2605.27852) | Project Page | Code

Official dataset of ClothTransformer: Unified Latent-Space Transformers for Scalable Cloth Simulation. It contains 2,056 penetration-free cloth simulation trajectories (240 frames each, 493,440 frames in total, ~33 GB) across three scenarios: free-fall collision onto diverse objects, garments on animated human bodies, and robotic cloth manipulation. Each trajectory is a cloth mesh interacting with a static or moving collider, stored as one self-contained NumPy .npz file.

Preview

Human Garment	Robotic Manipulation	Diverse Object Collision

Dataset Structure

ClothTransformer-dataset/
├── sims_collision_all_static_1k/   # Diverse Object Collision: cloth dropped onto a static object (1,000 sims)
├── sims_garment_anim_all/          # Human Garment: garment worn by an animated body              (56 sims)
└── sims_grasp_all_case1/           # Robotic Manipulation: cloth grasped by a robotic gripper     (1,000 sims)

Each simulation i comes with two files:

File	Content
sim_{i:05d}.npz	The full simulation arrays (see below). This is all you need.
sim_{i:05d}_processed_cloth_00000_uv.*	Cloth rest mesh with UVs — .obj in sims_garment_anim_all, Houdini .bgeo.sc in the other two scenarios. Optional: the same geometry is already inside the .npz.

The array schema is identical across scenarios. Mesh resolution varies per sample — always read shapes from the file.

Subset	Cloth #Verts	Cloth #Faces	Collider #Faces	Sequences	Frames
Human Garment	1k–3.6k	2k–7.1k	1k–5.1k	56	13,440
Robotic Manipulation	1k–4k	1.9k–7.9k	0.6k	1,000	240,000
Diverse Object Collision	3.6k	7k	1k–4k	1,000	240,000
Total	1k–4k	1.9k–7.9k	0.6k–5.1k	2,056	493,440

Data Fields

Notation: N_v / N_c = cloth / collider vertex count, F / E = triangle / edge counts, T = 240 frames at dt = 1/60 s (4 s). Coordinates are world units, Y-up. Velocity fields store per-frame displacements; divide by dt for physical velocity.

Key	Shape	Dtype	Description
initial	(N_v, 6)	float64	Cloth rest state: [0:3] rest position (equals traj[0]), [3:6] UV as (u, v, 0).
traj	(T, N_v, 3)	float32	Cloth vertex positions per frame.
traj_vel	(T, N_v, 3)	float32	Cloth per-frame displacement: traj_vel[t] == traj[t] − traj[t−1].
triangles	(F_cloth, 3)	int32	Cloth triangle vertex indices (0-based).
edges	(E_cloth, 2)	int32	Cloth edge vertex-index pairs.
collision_vertices	(T, N_c, 3)	float32	Collider vertex positions per frame (constant for static colliders).
collision_vel	(T, N_c, 3)	float32	Collider per-frame displacement.
collision_triangles	(F_col, 3)	int32	Collider triangle vertex indices.
collision_edges	(E_col, 2)	int32	Collider edge vertex-index pairs.
collision	(T, F_col, 9)	float32	Redundant convenience field: per-frame triangle corner positions, exactly collision_vertices gathered by collision_triangles.

Topology is fixed over time within a sample and differs between samples.

Usage

import numpy as np

data = np.load("sims_collision_all_static_1k/sim_00000.npz")  # no allow_pickle needed
cloth_pos = data["traj"]                       # (240, N_v, 3)
cloth_vel = data["traj_vel"] / (1 / 60)        # world units per second

# Reconstruct the redundant `collision` field from raw arrays:
cv, tri = data["collision_vertices"], data["collision_triangles"]
collision = cv[:, tri.reshape(-1)].reshape(cv.shape[0], -1, 9)

See example_load_dataset.py for a complete script that loads a sample and exports frames to OBJ.

Data Generation

All trajectories are ground-truth simulations of the Baraff–Witkin cloth model produced with GIPC (Huang et al., ACM TOG 2024), a penetration-free GPU solver based on incremental potential contact. The dataset is strictly intersection-free, making it suitable for training with Continuous Collision Detection (CCD) losses. Material parameters: stretching Young's modulus 1e6 Pa, bending Young's modulus 1e5 Pa, Poisson's ratio 0.49, shear stiffness 5e6 Pa, density 200 g/m², friction coefficient 0.4.

Subset	Cloth meshes	Collider
Human Garment	T-shirts and skirts	Animated SMPL avatars; all motions (walking, running, dancing, jumping, etc.) generated with Make-It-Animatable, no external mocap data involved
Robotic Manipulation	1,000+ garments from the Dataset of 3D Garments with Sewing Patterns	Robotic gripper
Diverse Object Collision	Square cloth sheets	1,000+ rigid objects sampled from Objaverse

In the paper, each subset is split into train / validation / test sets at an 8:1:1 ratio.

License

Released under CC BY 4.0. Portions of the underlying geometry derive from third-party assets with their own terms:

• SMPL-Body (CC BY 4.0) — human body colliders, courtesy of the Max Planck Institute for Intelligent Systems.
• Make-It-Animatable (Apache 2.0) — used to generate the body motions in the Human Garment subset.
• Dataset of 3D Garments with Sewing Patterns (CC BY 4.0) — garment meshes in the Robotic Manipulation subset.
• Objaverse (ODC-BY 1.0) — collider meshes in the Diverse Object Collision subset; each object retains its own Creative Commons license as recorded in the Objaverse metadata.

Citation

@article{zhang2026clothtransformer,
  title   = {ClothTransformer: Unified Latent-Space Transformers for Scalable Cloth Simulation},
  author  = {Zhang, Yu and Shao, Yidi and Ouyang, Wenqi and Lan, Yushi and Liang, Zhexin and Wu, Chengrui and Xu, Xudong and Pan, Xingang},
  journal = {arXiv preprint arXiv:2605.27852},
  year    = {2026}
}