DP3_OakInk_training_data

207 successful grasp + lift trajectories collected in IsaacSim 5.1 + cuRobo 0.8 from OakInk hand-pose sequences, retargeted onto a Franka 2-finger gripper. Used as an add-on to the DexYCB-sourced training set (UCBProject/DP3_DexYCB_training_data) for the next round of baseline_3 v4 Diffusion Policy 3D training.

Source code: UCB_Project @ gate3-curobo-ik — collector sim/run_grasp_sim_baseline3_v4.py, retarget Baseline1/oakink/retarget_oakink.py, orchestrator scripts/baseline_3_v4/oakink_full89_queue_resume.sh.

Important: this dataset is meant to be combined with the DexYCB-sourced 162-ep set on the training side. See the "Combined Training on A6000" section below for the new output paths that must be used so the previous DexYCB-only checkpoint and train/test split on the A6000 are preserved (we still plan to evaluate the previous model).

---

Why OakInk (not just more DexYCB)

DexYCB has 21 (mostly food-box / can / drill) objects. OakInk adds 89 (mostly container / mug / tool) categories with hand annotations recorded across multiple subjects and grasp styles. Adding OakInk to the DP3 training set increases category and grasp-style diversity well beyond what DexYCB alone provides.

The collection on RTX 5090 ran 89 objects × 846 source sessions × 4 yaws = 3 384 attempts in ~6 h 20 min wall time. Final yield: 207 successful trajectories (6.16 % overall) — lower than DexYCB's ~25 % because OakInk contains many wide containers (min-dim > Franka 8 cm span) and very thin tools (aspect-ratio > 5) that are physically un-graspable by a 2-finger parallel gripper. 45 of the 89 objects contributed ≥1 trajectory; 44 contributed 0 (see "Per-object Breakdown" below).

---

Per-object Breakdown

45 objects with ≥1 successful trajectory:

obj_id	cid	category	orig	yaw aug	total
C03001	1058	container	4	10	14
O01000	1001	container	1	10	11
A01009	1012	container	0	10	10
A01002	1008	container	0	9	9
A01005	1011	container	1	8	9
A01010	1009	container	1	8	9
A01026	1072	container	0	9	9
A01001	1003	container	2	6	8
S16001	1017	container	2	6	8
S16002	1039	container	3	5	8
A02015	1015	container	2	5	7
A15027	1025	container	0	7	7
A01008	1007	container	0	6	6
A01023	1071	container	1	5	6
C14001	1050	container	1	5	6
Y27035	1020	maniptools	0	6	6
A02021	1037	container	2	3	5
O03001	1014	container	0	5	5
S16005	1036	container	0	5	5
A02011	1022	container	1	3	4
O03003	1029	container	0	4	4
O21001	1044	maniptools	1	3	4
A01027	1073	container	0	3	3
C10001	1076	container	0	3	3
C22001	1055	maniptools	1	2	3
C37001	1056	maniptools	1	2	3
O02001	1021	container	1	2	3
S15004	1032	container	2	1	3
S20005	1041	maniptools	0	3	3
Y03021	1023	container	0	3	3
A02012	1033	container	0	2	2
A02032	1026	container	0	2	2
A15015	1027	container	0	2	2
C42001	1060	wearable	0	2	2
C90001	1078	geometry	0	2	2
O03002	1018	container	0	2	2
S10017	1087	container	1	1	2
S16003	1049	container	0	2	2
A02014	1024	container	0	1	1
A02030	1031	container	1	0	1
A16026	1019	container	0	1	1
C15001	1067	container	1	0	1
S10005	1081	container	0	1	1
S10008	1006	container	0	1	1
S10021	1002	container	1	0	1
		TOTAL	31	176	207

class_id 1001–1089 corresponds to the OakInk slot in Baseline1/oakink/class_id_map.json (DexYCB occupies cid 1–21; cid > 1000 is reserved for OakInk).

Yaw augmentation contributes 85 % of the data (176/207). For many objects MANO-derived grasp poses fail Franka kinematic reachability at the recorded orientation but succeed at one of the 90°/180°/270° rotated variants.

---

Per-episode Schema (HDF5)

Identical to DP3_DexYCB_training_data:

Field	Shape	dtype	Notes
state	(T, 8)	float32	[x,y,z, qw,qx,qy,qz, gripper] in object-centric G-frame
action	(T, 8)	float32	state[1:] (shifted by 1)
point_cloud	(T, 4096, 3)	float32	Static CAD surface samples in G-frame
obj_origin_G	attr (3,)	float64	(0, 0, obj_z) — table-relative offset
obj_quat_G_wxyz	attr (4,)	float64	Obj orientation in G-frame
ycb_class_id	attr scalar	int64	1001+ for OakInk
obj_id	attr str	-	e.g. "A01001"
dataset	attr str	-	"oakink" (lets training code disambiguate)

T ≈ 31 frames per episode (4 hover waypoints + 12 approach + 6 grasp + 9 lift, ±a few). state[T-1] is the post-lift gripper pose.

---

File Naming

oakink__<seq_id>__<ts>__<subj_flag>__<cam>[_yawDDD].hdf5

• seq_id = <obj_id>_<trial_indices> (e.g. A01001_0001_0000)
• ts = recording timestamp (2021-09-26-19-59-58)
• subj_flag = 0 (primary) — all 207 saved episodes are from subj=0
• cam = 0 — single camera retained per session; OakInk's 4-camera redundancy was de-duplicated since all 4 produce bit-identical world-frame trajectories
• _yawDDD suffix → yaw-augmented variant (90/180/270); no suffix → original

---

Download

# Option 1: huggingface-cli (recommended, parallel)
huggingface-cli download UCBProject/DP3_OakInk_training_data \
    --repo-type dataset \
    --local-dir Baseline1/data/episodes_b3_v4_oakink89_2026-05-26 \
    --include "data/*.hdf5"

# Move from data/ subdir up to root
mv Baseline1/data/episodes_b3_v4_oakink89_2026-05-26/data/*.hdf5 \
   Baseline1/data/episodes_b3_v4_oakink89_2026-05-26/
rmdir Baseline1/data/episodes_b3_v4_oakink89_2026-05-26/data

Or via Python:

from huggingface_hub import snapshot_download
snapshot_download(repo_id="UCBProject/DP3_OakInk_training_data",
                  repo_type="dataset",
                  local_dir="Baseline1/data/episodes_b3_v4_oakink89_2026-05-26",
                  allow_patterns="data/*.hdf5")

---

Combined Training on A6000 (DexYCB + OakInk → new DP3 model)

Important constraints:

• The A6000 already has the previous DexYCB-only DP3 checkpoint and the 162-ep train/test split saved on disk. We still plan to evaluate the previous model, so the new run MUST NOT overwrite those paths.
• Sim collection on A6000 was abandoned earlier this round due to glibc 2.31 (system) vs 2.35 (IsaacSim 5.1 requirement) mismatch. Do not attempt sim collection on A6000. All collection now happens on the dev box (RTX 5090); A6000 is training-only.

Step 1 — Layout the combined dataset (use a fresh dir)

cd $HOME/UCB_Project   # the A6000 repo clone

# Pick a FRESH name that does NOT collide with the prior run's
# episodes_b3_v4_full12_yaw/ (which holds the 162 DexYCB ep and is the
# input to the existing model). Suggested:
NEW=Baseline1/data/episodes_b3_v4_dexycb162_oakink207_2026-05-26
mkdir -p "$NEW"

# 1.1 Copy DexYCB 162 ep from the existing local dir (already downloaded
# from UCBProject/DP3_DexYCB_training_data — do NOT re-download).
cp Baseline1/data/episodes_b3_v4_full12_yaw/*.hdf5 "$NEW/"

# 1.2 Download the new OakInk 207 ep from THIS dataset.
huggingface-cli download UCBProject/DP3_OakInk_training_data \
    --repo-type dataset --local-dir /tmp/oakink_dl --include "data/*.hdf5"
cp /tmp/oakink_dl/data/*.hdf5 "$NEW/"

# 1.3 Verify count
ls "$NEW"/*.hdf5 | wc -l       # expect 162 + 207 = 369

Step 2 — Build zarr (fresh, distinct from the DexYCB-only zarr)

conda activate dp3   # same dp3 env A6000 already has

# Output to a NEW zarr file (do not overwrite the existing one)
python Baseline1/convert_to_zarr.py \
    "$NEW" \
    --output_zarr Baseline1/data/dp3_train_v4_dexycb162_oakink207.zarr

Step 3 — Create a new DP3 train/test split

The previous split (32 train + 8 test for the 3-obj experiment, OR 130 train

• 32 test for the 162-ep v4_sml run) lives under third_party/3D-Diffusion-Policy/.../experiments/<old_exp_name>/. Do not touch it. Make a fresh experiment dir:

cd third_party/3D-Diffusion-Policy/3D-Diffusion-Policy

# Suggested fresh experiment name
EXP=dexycb162_oakink207_2026-05-26

# 80/20 split — 295 train / 74 test (Baseline1/split_v4_full12.py is the
# splitter used previously; copy + rerun on the NEW zarr file)
python ../../../Baseline1/split_v4_full12.py \
    --zarr ../../../Baseline1/data/dp3_train_v4_dexycb162_oakink207.zarr \
    --train_ratio 0.8 \
    --out_dir experiments/$EXP/data_split

Step 4 — Configure DP3 training (fresh config, output dir)

Copy the prior config and adjust:

# Copy prior config (whatever you used for the 162-ep run)
cp config/v4_sml.yaml config/${EXP}.yaml

# Edit config/${EXP}.yaml:
#   - task.dataset_zarr_path:  point to the NEW zarr (dp3_train_v4_dexycb162_oakink207.zarr)
#   - exp_name:                 ${EXP}
#   - hydra.run.dir:            experiments/${EXP}/${now:%Y-%m-%d_%H-%M-%S}
#   - checkpoint.save_ckpt:     True            ← critical, defaulted False historically
#   - checkpoint.topk.k:        3 (or more)     ← keep last-N best
#   - training.num_epochs:      3000 (matches prior run)

Step 5 — Launch training

WANDB_MODE=online   # or offline if A6000 has no internet
python train.py --config-name=${EXP}

Expected wall time on A6000 at batch_size=128: ~6 h for 3000 epochs on 369 ep (vs ~3 h for 162 ep).

Step 6 — After training, ckpt + logs land in a fresh dir

experiments/${EXP}/{date_time}/checkpoints/
experiments/${EXP}/{date_time}/wandb/

The PRIOR experiment dir (experiments/<prior_v4_sml_exp>/) is left untouched. To re-evaluate the prior model:

python eval.py --config-name=<prior_v4_sml_exp>   # unchanged from before

---

Collection Details (same as DexYCB)

• IsaacSim 5.1, PhysX TGS solver, GPU dynamics + CCD
• cuRobo 0.8 motion planner (per-phase mesh toggle: pre-grasp WITH mesh, final/lift WITHOUT mesh)
• mass = 0.05 kg (hardcoded; real per-class mass causes PhysX overflow with OakInk's photogrammetry-derived non-watertight meshes)
• chunked-5 + retry wrapper to recover from PhysX corruption events
• Yaw augmentation: orig + {90°, 180°, 270°} = 4 attempts per source ep
• 2-parallel orchestrator on RTX 5090 (verified safe); 4-par broke historically due to cuRobo subprocess contention

---

Why the 44 zero-yield objects?

Bucketed analysis on the 37 zero-yield objects in the partner shortlist (+ 7 self-identified):

• 12 obj: min_dim ≥ 8 cm → exceeds Franka 8 cm gripper span (e.g. S10010-S10020 mug body, A02018 teapot, C13001 large kettle)
• 10 obj: aspect_ratio ≥ 5 → extremely thin (e.g. S20021 spoon, Y29040 stick, O24001 spatula) — gripper can't pinch
• 6 obj: severe mesh degeneracy (non-manifold edges > 2 000) where cuRobo's collision approximation diverges from PhysX runtime geometry
• ~9 obj: MANO grasp pose puts the wrist below or to the side of the object such that the Franka 7-DoF cannot match the required wrist orientation (gripper z-axis world component > +0.2)

This is not a pipeline bug — DexYCB succeeds at ~25 % on the same collector because YCB benchmark obj are designed for parallel-jaw grasping. The OakInk objects expose the natural human-vs-Franka kinematic gap.

---

License & citation

Data: CC-BY-4.0. OakInk source data subject to the original OakInk license.

@inproceedings{yang2022oakink,
  title = {OakInk: A Large-scale Knowledge Repository for Understanding
           Hand-Object Interaction},
  author = {Yang, Lixin and Li, Kailin and Zhan, Xinyu and Wu, Fei and Xu,
            Anran and Liu, Liu and Lu, Cewu},
  booktitle = {CVPR},
  year = {2022}
}