Cloth-splatters/unfolding-meshes

Synthetic cloth-unfolding trajectories for training graph-based dynamics and state-estimation models. Same simulator, schema, and cloth set as Cloth-splatters/folding-meshes, but the action sampler reverses fold motions: trajectories start from a folded state and pull the cloth flat.

Used by:

• Cloth-splatters/unfolding-dynamics-gns
• Cloth-splatters/unfolding-dynamics-gps
• Cloth-splatters/unfolding-state-est-gps

Files

File	Size	Train / Val cloths	Cameras	Notes
unfold_seed_1397.h5	1.91 GB	105 / 45	cam_0	full set
unfold_seed_1397_filtered.h5	1.19 GB	—	cam_0	quality-filtered subset of the same generation run

Schema

Identical to folding-meshes; see that dataset card for the full HDF5 layout. Top-level groups: metadata/, training/, validation/. Each cloth carries edges, faces, rest_positions, and a set of trajectory_<i> groups whose step_* entries contain positions, velocities, gripper_pos, plus per-camera pointclouds, images, depth, segmentation.

Loading

from huggingface_hub import hf_hub_download
path = hf_hub_download(
    "Cloth-splatters/unfolding-meshes",
    filename="unfold_seed_1397.h5",
    repo_type="dataset",
)

License

MIT.