Team Ai
Datasetpublic

Kit-Key/libero-3d-pointcloud-complete

LIBERO 3D Point-Cloud GT — complete object meshes (sim) Per-frame, object-level 3D point-cloud ground truth for the LIBERO benchmark. We replay each official LIBERO human demonstration's recorded MuJoCo states and, at every frame, sample each object's point cloud directly from its visual mesh (area-weighted, posed by the live state) — giving complete, view-independent clouds (full object incl. back/occluded surfaces), not a single-view 2.5D shell. Exact metric geometry, no… See the full description on the dataset page: https://huggingface.co/datasets/Kit-Key/libero-3d-pointcloud-complete.

sourceHugging Facemitupdated 3mo agoView on Hugging Face
0likes139downloads
Dataset Card

LIBERO 3D Point-Cloud GT — complete object meshes (sim)

Per-frame, object-level 3D point-cloud ground truth for the LIBERO benchmark. We replay each official LIBERO human demonstration's recorded MuJoCo states and, at every frame, sample each object's point cloud directly from its visual mesh (area-weighted, posed by the live state) — giving complete, view-independent clouds (full object incl. back/occluded surfaces), not a single-view 2.5D shell. Exact metric geometry, no estimation.

A depth-unprojected (single-view 2.5D, RGB-aligned) variant is at Kit-Key/libero-3d-pointcloud.
  • —Source: LIBERO base demos (yifengzhu-hf/LIBERO-datasets), 130 tasks × 50 demos across the 5 suites (spatial / object / goal / long / 90).
  • —Object selection: each task's obj_of_interest from its BDDL (target + destination objects; distractors excluded; same-class instances disambiguated). Abstract placement regions map to their parent fixture's cloud.
  • —Teacher: sim depth unprojection (metric).

Format

One HDF5 file per (task, demo) episode. Each frame group "%06d" holds:

keyshape / dtypemeaning
object_pcf32 (4, 512, 3)scene-normalized (scheme B) per-object point clouds
object_namesstr (4,)LIBERO instance names ("" = padding)
object_kindsstr (4,)graspable \fixture \""
validbool (4,)slot valid this frame (False when object fully occluded)
norm_paramsf32 (4,)shared centroid c(3) + radius r(1); metric = pc·r + c
rgb_jpeguint8 (N,)JPEG bytes of the 256×256 agentview RGB

Root attrs: task, language, bddl, teacher="sim", coord="scene_normalized_B", view="agentview", n_frames.