UCBProject / ProcessedData
Phase 1A outputs of the Affordance2Grasp pipeline — used by downstream phases:
- Phase 2 (training the main method's PointNet++ contact-prediction net) reads
training_fp/{dataset}/{object}.hdf5. - Phase 3 (inference, grasp sampling, sim execution, Sim2Real deployment) reads
per-object affordance priors from
human_prior_fp/{object}.hdf5.
The upstream object meshes (SAM3D reconstructions) live in a sibling repo,
UCBProject/ObjMesh.
Layout
training_fp/
oakink/ 100 files (A* C* O* S* Y* — OakInk object IDs)
dexycb/ 20 files (ycb_dex_01 ... ycb_dex_20 — DexYCB grasp objects)
human_prior_fp/
A01001.hdf5 ... Y35037.hdf5 100 oakink files
ycb_dex_01.hdf5 ... ycb_dex_20.hdf5 20 dexycb files
HDF5 schema
training_fp/{ds}/{obj}.hdf5 (training-ready)
| key | shape | dtype | meaning |
|---|---|---|---|
point_cloud |
(4096, 3) | float32 | 4096 surface samples of the SAM3D object mesh (metric m, mesh canonical frame) |
normals |
(4096, 3) | float32 | per-point unit normals |
human_prior |
(4096,) | float32 | per-point contact probability in [0, 1] (Gaussian-smoothed, per-object max-normalised) |
robot_gt |
(4096,) | float32 | all-zero placeholder (no robot ground truth for this regime) |
force_center |
(3,) | float32 | centroid of mesh verts with contact_smooth >= 80th percentile |
attr object |
str | — | object id (e.g. ycb_dex_14 or A01001) |
human_prior_fp/{obj}.hdf5 (inference-ready, indexed by object only)
Same five fields as above. Same content as training_fp/{ds}/{obj}.hdf5, but flattened
across datasets so that inference/predictor.py can read by object id alone.
Provenance
Generated by:
data/batch_depth_pro.py→data/batch_haptic.py→tools/batch_obj_pose.py→data/batch_align_mano_fp.py- Source raw data: DexYCB (NVIDIA, CC-BY-NC-4.0) + OakInk (CVPR 2022).
- Phase 1A run on lab RTX 5090; OakInk completed 2026-05-10, DexYCB completed 2026-05-14.
HO3D-v3 and ARCTIC are not included here — those will be added when the partner runs their Phase 1A on those datasets.