DP3_DexYCB_training_data

162 successful grasp + lift trajectories collected in IsaacSim 5.1 + cuRobo 0.8 from DexYCB hand-pose sequences, retargeted onto a Franka 2-finger gripper. Used to train the baseline_3 v4 Diffusion Policy 3D (DP3) policy in the UCB_Project repo (gate3-curobo-ik branch).

10 YCB objects after dropping foam and scissors (cuRobo could not plan a single successful grasp on either shape). Each source DexYCB episode was collected at its original object yaw and one randomly-selected augmented yaw in {90°, 180°, 270°} around world-Z (a task-symmetric transform — gravity, table and contact geometry are unchanged by yaw rotation).

2026-05-26 update — A complementary OakInk-sourced dataset is now available at UCBProject/DP3_OakInk_training_data (207 ep, 45 obj). For the next DP3 round we will train on the combined 369-ep dataset (DexYCB 162 + OakInk 207). See the "Combined Training (DexYCB + OakInk → fresh DP3 model)" section below.

Per-object Breakdown

ycb_class_id	object	orig	yaw aug	total
03	sugar	14	8	22
04	tomato	14	6	20
05	mustard	11	4	15
06	tuna	11	6	17
07	pudding	9	9	18
08	gelatin	11	4	15
09	potted_meat	11	7	18
12	bleach	17	8	25
15	drill	3	4	7
18	marker	3	2	5
	TOTAL	104	58	162

Total size: ~238 MB. Each .hdf5 is ~1.5 MB.

Per-episode Schema (HDF5)

Field	Shape	dtype	Notes
state	(31, 8)	float32	[x,y,z, qw,qx,qy,qz, gripper] in object-centric G-frame, retarget-quat convention
action	(31, 8)	float32	state[1:] (shifted by 1)
point_cloud	(31, 4096, 3)	float32	Static CAD surface samples in G-frame (all 31 frames identical; object is static during collection)
obj_origin_G	attr (3,)	float64	Object frame origin in G-frame
obj_quat_G_wxyz	attr (4,)	float64	Object orientation in G-frame
ycb_class_id	attr scalar	int	DexYCB class id (e.g. 03 = sugar)

File Naming

dexycb__<session>__<sub-session>__<camera_id>__ycb_dex_NN[_yawDDD].hdf5

• No _yaw suffix → original DexYCB yaw
• _yaw90 / _yaw180 / _yaw270 → yaw-augmented variant

Download

# Option 1: huggingface-cli
huggingface-cli download UCBProject/DP3_DexYCB_training_data \
    --repo-type dataset \
    --local-dir Baseline1/data/episodes_b3_v4_full12_yaw

# Option 2: snapshot_download from Python
from huggingface_hub import snapshot_download
snapshot_download(repo_id="UCBProject/DP3_DexYCB_training_data",
                  repo_type="dataset",
                  local_dir="Baseline1/data/episodes_b3_v4_full12_yaw")

The 162 hdf5 files land under data/ inside this repo, so the --local-dir above ends up with Baseline1/data/episodes_b3_v4_full12_yaw/data/<162 hdf5>. For the UCB_Project pipeline, move them up one level so the path matches the README:

mv Baseline1/data/episodes_b3_v4_full12_yaw/data/*.hdf5 \
   Baseline1/data/episodes_b3_v4_full12_yaw/
rmdir Baseline1/data/episodes_b3_v4_full12_yaw/data

Training Pipeline

Full retrain instructions (DexYCB-only, original run): Baseline1/RETRAIN_V4_FULL12.md in the UCB_Project repo.

---

Combined Training (DexYCB + OakInk → fresh DP3 model)

We are now training a new DP3 model that combines this 162-ep DexYCB set with the 207-ep OakInk set at UCBProject/DP3_OakInk_training_data.

Important — preserve previous DexYCB-only artefacts:

• The A6000 already has the previous DexYCB-only DP3 checkpoint (v4_sml experiment, 3000-epoch run) and the corresponding train/test split saved on disk. We still intend to evaluate that model. The new combined run MUST use distinct output paths so nothing is overwritten.
• Sim collection for this round was completed entirely on the dev box (RTX 5090). The earlier plan to also run sim collection on A6000 was abandoned because the system glibc (2.31) is incompatible with IsaacSim 5.1's requirement (glibc 2.35). The A6000 is training-only this round. (See UCBProject/baseline_3_v4_collection_assets for the deprecated A6000 collection instructions, kept for reference only.)

Step 1 — Layout the combined dataset in a fresh dir

cd $HOME/UCB_Project   # the A6000 repo clone

# Fresh dir — do NOT reuse Baseline1/data/episodes_b3_v4_full12_yaw which holds
# the 162-ep DexYCB set and is the training input for the existing model.
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 —
# do NOT re-download).
cp Baseline1/data/episodes_b3_v4_full12_yaw/*.hdf5 "$NEW/"

# 1.2 Download the new OakInk 207 ep
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 a FRESH zarr (do not overwrite the existing one)

conda activate dp3   # same env A6000 already has

python Baseline1/convert_to_zarr.py \
    "$NEW" \
    --output_zarr Baseline1/data/dp3_train_v4_dexycb162_oakink207.zarr

Existing zarr (DexYCB-only) at Baseline1/data/dp3_train_v4_sml.zarr remains untouched.

Step 3 — Fresh train/test split

The previous split lives in third_party/3D-Diffusion-Policy/.../experiments/v4_sml/data_split/. Do not touch it. Make a new experiment dir:

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

EXP=dexycb162_oakink207_2026-05-26

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 — Fresh DP3 config + output dir

Copy the prior config and adjust:

cp config/v4_sml.yaml config/${EXP}.yaml

# Edit config/${EXP}.yaml:
#   task.dataset_zarr_path:   Baseline1/data/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)
#   training.num_epochs:      3000

Step 5 — Launch

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 — Output lands in a fresh dir

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

The previous experiment (experiments/v4_sml/) and its checkpoint remain untouched. To re-evaluate the previous model later:

python eval.py --config-name=v4_sml   # unchanged from before

---

Collection Details

Collected by sim/run_grasp_sim_baseline3_v4.py (gate3-curobo-ik branch) with:

• IsaacSim 5.1, PhysX TGS solver, GPU dynamics + CCD
• cuRobo 0.8 motion planner (per-phase mesh toggle for pre-grasp WITH mesh, final/lift WITHOUT mesh)
• mass = 0.05 kg (hardcoded; real per-class mass triggers PhysX overflow)
• chunked-5 + retry wrapper to recover from PhysX corruption events

License

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

Citation

If you use this dataset, please also cite DexYCB:

@inproceedings{chao2021dexycb,
  title = {DexYCB: A Benchmark for Capturing Hand Grasping of Objects},
  author = {Chao, Yu-Wei and Yang, Wei and Xiang, Yu and Molchanov, Pavlo
            and Handa, Ankur and Tremblay, Jonathan and Narang, Yashraj S
            and Van Wyk, Karl and Iqbal, Umar and Birchfield, Stan and others},
  booktitle = {CVPR},
  year = {2021}
}