Data Format Specification

Grasp Poses

Per-scene files

One HDF5 file per scene:

• File name: {scene_id}.h5
• Datasets (float arrays):
  • grasp_qpos: shape (N, D)
  • pre_grasp_qpos: shape (N, D)
  • post_grasp_qpos: shape (N, D)

Where:

• N = number of stored grasps for this scene (varies per scene; may include appended results from multiple runs)
• D = dimension of the configuration. For floating base robots, this includes 3D position, 3D euler angles (rxyz), and hand joint positions.

Merged file

• File name: {output_dir}.h5
• Structure: one group per scene /{scene_id}/ containing the same datasets as per-scene files