Cloth-splatters/lifting-meshes

Synthetic cloth-lifting trajectories for training graph-based dynamics models. The action sampler picks a single vertex and lifts it vertically while the rest of the cloth deforms under gravity.

Currently only the no-hole cloth subset is published; with-hole variants may be added later.

Used by:

• Cloth-splatters/lifting-dynamics-gns
• Cloth-splatters/lifting-dynamics-edge-gnn

Files

File	Size	Train / Val cloths	Cameras
lift_seed_338.h5	540 MB	30 / 15	cam_0
lift_seed_6438.h5	524 MB	—	cam_0

Cloth names are of the form cloth_without_hole_<id>.

Schema

Mostly identical to folding-meshes, with one difference: lifting files do not carry per-cloth rest_positions — only edges and faces. Use step_0000/positions as the initial flat-cloth state if you need a rest configuration.

metadata/rendering_parameters/...
training/
  <cloth_name>/
    edges  (E, 2) int32
    faces  (F, 3) int32
    trajectory_<i>/
      actuated_vertices (V, 1) float64
      goal_vertices     (V, 1) float64
      step_<n>/
        positions    (V, 3) float64
        velocities   (V, 3) float64
        gripper_pos  (1, 3) float64
        pointclouds/cam_0  (N, 3)
        images/cam_0       (H, W, 3)
        depth/cam_0        (H, W)
        segmentation/cam_0 (H, W)
validation/...

Loading

from huggingface_hub import hf_hub_download
path = hf_hub_download(
    "Cloth-splatters/lifting-meshes",
    filename="lift_seed_338.h5",
    repo_type="dataset",
)

License

MIT.