UCBProject/baseline_3_v4_oakink362_thumb_middle

OakInk 362 ep — thumb+middle (baseline_3 v4 DP3 training data)

OakInk grasp episodes for DP3 (3D-Diffusion-Policy) training on Franka parallel-jaw. Sim collection via sim/run_grasp_sim_baseline3_v4.py (UCB_Project).

What's in here

data/
├── oakink__<obj_id>_<sub>__<ts>__<subj>__<cam>.hdf5             # orig yaw=0
├── oakink__<obj_id>_<sub>__<ts>__<subj>__<cam>_yaw{90,180,270}.hdf5
└── ...
RESCUE_LOG.md   # which ep are from the rescue (thumb+INDEX) layer

• 268 ep collected fresh with PINCH_FINGER=middle (J=12)
• 94 ep rescued from a prior thumb+INDEX (J=8) run to fill gaps where thumb+middle couldn't grasp — see RESCUE_LOG.md
• Total: 362 ep covering 57 / 74 use=true OakInk obj
• yaw augmentation: 3 yaws per source ep (90, 180, 270) + orig

⚠️ DexYCB data (162 ep) is NOT included here

Use the existing repo: UCBProject/baseline_3_v4_dexycb162_oakink207_dp3 (the DexYCB hdf5 are unchanged between this run and the prior one).

For training: concatenate both — DexYCB 162 ep + this OakInk 362 ep → 524 ep total.

Provenance

OakInk retarget	Baseline1/oakink/retarget_oakink.py (UCB_Project)
PINCH_FINGER default	middle (= MANO J=12) — empirically ~2× collection rate vs index
Onset selection	OLD (= argmax(d ≤ d_min+4cm))
Yaw aug	3 yaws (90/180/270) at collection time
Sim runtime	IsaacSim 5.1 (env_isaaclab), RTX 5090, PAR=3
Collector	sim/run_grasp_sim_baseline3_v4.py
Mass	0.05 kg hardcoded (PhysX-stable)
cuRobo fallback	enabled (solve_plan_with_fallback)
Collection date	2026-05-26 → 2026-05-27
Wallclock	5h53m at PAR=3 on RTX 5090
Errors	0 sanity_fail / 0 abort / 0 OOM

How rescue worked

Smoke comparison on A01001 showed thumb+middle had a ~2× collection rate. After the full74 batch finished, we filled gaps from the prior thumb+INDEX run (Baseline1/data/episodes_b3_v4_oakink89_2026-05-26, 207 ep) where:

• the same (src_ep, yaw) filename did not exist in the new thumb+middle run
• → copied OLD/F → NEW/F (transparent merge, same filename)

This introduces a small mixed-convention noise: rescued ep have a target EE position ~2.8 cm offset from the new ep (different finger midpoint), but DP3 PC→EE training tolerates this for the coverage benefit. RESCUE_LOG.md lists the 94 rescued files for full traceability.

Manifest

74 use=true OakInk obj (out of 89 in Baseline1/oakink/class_id_map.json):

• 50 obj with thumb+middle ep
• +7 obj added by rescue → 57 total
• 17 obj 0-ep (mostly contactpose tools + S20* knives — geometry not parallel-jaw friendly)

How to train (A6000 instructions)

# 1. Pull DexYCB 162 ep
hf download UCBProject/baseline_3_v4_dexycb162_oakink207_dp3 \
    --repo-type model  # ckpt only; DexYCB data is in DP3_DexYCB_training_data dataset

# Actually for the DexYCB hdf5:
hf download UCBProject/DP3_DexYCB_training_data --repo-type dataset --local-dir <DEX_DIR>

# 2. Pull OakInk 362 ep (this repo)
hf download UCBProject/baseline_3_v4_oakink362_thumb_middle \
    --repo-type dataset --local-dir <OAKINK_DIR>

# 3. Merge into single training dir
mkdir -p Baseline1/data/episodes_b3_v4_dexycb162_oakink362_combined
cp <DEX_DIR>/data/*.hdf5     Baseline1/data/episodes_b3_v4_dexycb162_oakink362_combined/
cp <OAKINK_DIR>/data/*.hdf5  Baseline1/data/episodes_b3_v4_dexycb162_oakink362_combined/
ls Baseline1/data/episodes_b3_v4_dexycb162_oakink362_combined/*.hdf5 | wc -l   # should be 524

Train with task_baseline1_b3_v4_*.yaml (see prior run's config for params).