hackhackhack66666/aaai-datasets

AAAI paper datasets — OAT training data (HDF5 sources + Zarr)

Canonical training datasets used to reproduce our AAAI submission
"Spend Compute on Selection, Not Reduction" (Consensus Selection / CS-D on OAT policies).

Hub repo: hackhackhack66666/aaai-datasets
Code / eval artifacts: separate repos (mipt_paper/oat on cluster, HF model/eval repos for RoboCasa Wave1–2).
This dataset repo contains only raw + converted training data (HDF5 where retained, Zarr for all suites).

---

Quick map — what is used in the paper?

Suite	Zarr (train)	HDF5 (source)	In paper Table P?	Paper role
RoboMimic Lift	robomimic/zarr/lift_N200.zarr	robomimic/hdf5/lift_mh_image.hdf5	Yes (Lift run B ep-1400)	Generalization baseline
RoboMimic Can	robomimic/zarr/can_N200.zarr	robomimic/hdf5/can_mh_image.hdf5 (+ raw can/mh/*)	Yes	Generalization baseline
RoboMimic Square	robomimic/zarr/square_N200.zarr	robomimic/hdf5/square_mh_image.hdf5	Yes	Generalization baseline
MetaWorld coffee-pull	metaworld/zarr/coffee-pull_N50.zarr	— (generated)	Yes	Generalization baseline
MetaWorld stick-pull	metaworld/zarr/stick-pull_N50.zarr	—	Yes	Generalization baseline
MetaWorld disassemble	metaworld/zarr/disassemble_N50.zarr	—	Yes	Generalization baseline
MetaWorld box-close	metaworld/zarr/box-close_N50.zarr	—	Yes	Generalization baseline
MetaWorld MT4 multitask	metaworld/zarr/mt4_N50.zarr	—	No (exploratory only)	Early multitask probe; not Table P
RoboCasa close_drawer	robocasa/zarr/close_drawer_N200.zarr	partial human HDF5	Yes	Generalization baseline
RoboCasa coffee_press_button	robocasa/zarr/coffee_press_button_N200.zarr	partial human HDF5	Yes	Generalization baseline
RoboCasa turn_off_sink_faucet	robocasa/zarr/turn_off_sink_faucet_N200.zarr	— (zarr only)	Yes	Generalization baseline
RoboCasa turn_off_microwave	robocasa/zarr/turn_off_microwave_N200.zarr	— (zarr only)	Yes	Generalization baseline

LIBERO is not included here (separate diagnosis track; different n_test protocol).

---

Directory layout

robomimic/
  zarr/{lift,can,square}_N200.zarr/     # OAT training format (200 demos each)
  hdf5/
    lift_mh_image.hdf5
    can_mh_image.hdf5
    square_mh_image.hdf5
    can/mh/demo_v15.hdf5                  # raw multi-human before image extract
    can/mh/image_v15.hdf5

metaworld/
  zarr/
    mt4_N50.zarr                          # 200 eps = 50×4 tasks (multitask)
    {box-close,coffee-pull,disassemble,stick-pull}_N50.zarr
  METAWORLD_GENERATION.md                 # how demos differ from upstream sim-env

robocasa/
  zarr/<task>_N200.zarr/                  # each includes ROBOCASA_SOURCE.txt
  hdf5/
    CloseDrawer/human/demo_gentex_im128_randcams.hdf5
    CoffeePressButton/human/demo_gentex_im128_randcams.hdf5

Approximate sizes (cluster, 2026-08-13):

Path	~Size
RoboMimic HDF5 total	10.5 GB
RoboMimic Zarr total	1.0 GB
MetaWorld Zarr total	1.4 GB
RoboCasa Zarr total	4.6 GB
RoboCasa HDF5 (retained)	0.6 GB
Total	~18 GB

---

Common Zarr schema (OAT)

All Zarr stores follow the OAT pipeline layout:

• data/ — time-major arrays (action, RGB, proprioception)
• meta/episode_ends — cumulative step indices per episode
• RoboCasa only: ROBOCASA_SOURCE.txt inside each zarr (provenance + sha256 in paper log)

Validate locally (from oat/ repo):

python scripts/validate_robomimic_data.py
python scripts/validate_robocasa_data.py
python scripts/validate_metaworld_data.py data/metaworld/box-close_N50.zarr --num-tasks 1
python scripts/validate_metaworld_data.py data/metaworld/mt4_N50.zarr --num-tasks 4 --require-subtask-counts

---

1. RoboMimic (official mh demos → Zarr)

Source

• Benchmark: RoboMimic multi-human (mh) image demonstrations.
• Tasks: Lift, Can, Square.
• 200 demonstrations per task (subsampled from 300 mh demos).
• Lift: direct download lift_mh_image.hdf5.
• Can / Square: download raw demo_v15.hdf5, replay through robosuite 1.5 to extract image_v15.hdf5 (see scripts/extract_robomimic_mh_image.sh).

Zarr keys (train)

Key	Shape (per step)
action	7
agentview_image	84×84×3
robot0_eye_in_hand_image	84×84×3
(+ proprio keys per converter)

Paper linkage

• Tokenizer: top MSE checkpoint per task (ep-0690 square, ep-1970 lift, etc.).
• Policy TopK @ test_start_seed=1000; Table P report @ test_start_seed=10000, -n 5, OAT8.
• Locked Table P ckpts: Can ep-1700, Square ep-0700, Lift ep-1400 (run B).

Notes

• Do not use legacy OAT-RoboMimic-Fine-tune/BLT-OAT/data/robomimic/ (old ph layout) — unrelated to this paper track.

---

2. MetaWorld (locally generated expert demos → Zarr)

⚠️ Important — not the official MetaWorld HDF5 download

All MetaWorld data here was generated on our cluster with:

python scripts/gen_metaworld_data.py --task_name <task> --num_episodes 50 --force

(log: logs/metaworld_single_data_regen.log; multitask: logs/gen_metaworld_mt4_N50.log)

Environment code is ported from sim-env into oat/oat/env/metaworld/ (MuJoCo 2.1.0 / MetaWorld v2 stack — different from RoboMimic/LIBERO robosuite 1.4).

Demo acceptance rule (differs from stricter filters)

An episode is kept iff info["success"] == True on at least one timestep during the expert rollout:

episode_success_count += int(bool(info.get("success", False)))
if episode_success_count == 0:
    continue  # reject episode, retry with new seed

We do not require success only on the terminal step. This matches mt4_N50 generation and the four single-task regen runs used for Table P.

Reset / seed fix (July 2026)

Single-task regen originally hung because MetaworldEnv.reset() without seed restored a fixed MuJoCo snapshot. Fix:

roll_seed = episode_idx * 1_000_000 + attempt_idx
obs_dict, _ = env.reset(seed=roll_seed)

Without incrementing attempt_idx on reject, retries repeat the same failed init.

Tasks & episode counts (validated 2026-08-13)

Zarr	Episodes	Action dim	Cameras	Steps (total)
mt4_N50.zarr	200 (50×4)	4	4×128² RGB + agent_pos 9D	20 316
box-close_N50.zarr	50	4	same	5 616
coffee-pull_N50.zarr	50	4	same	4 088
disassemble_N50.zarr	50	4	same	5 220
stick-pull_N50.zarr	50	4	same	6 101

MT4 task order in subtask_counts: [box-close, coffee-pull, disassemble, stick-pull].

Single-task files can also be obtained deterministically by splitting mt4_N50.zarr (scripts/split_metaworld_mt4_zarr.py, round-robin i % 4); paper Table P uses freshly regen single-task zarr, byte-identical to split for the episodes that were completed before regen.

How this differs from “original sim-env / paper MetaWorld”

1. Success timing: our collector accepts any-time success during the demo rollout (see above). A stricter terminal-only filter was tested and rejected for compatibility with mt4_N50.
2. Implementation port: same intent as sim-env, but our wrapper had the reset-without-seed bug (fixed before final single-task zarr). Treat MW numbers as valid within this port, not as a claim of bit-identical reproduction of Chaoqi sim-env demo files.
3. Paper limitation (explicit): “MetaWorld demo port — controlled limitation; interpret MW within our implementation.” (see RESULTS.md / RESOLUTIONPLAN.md in code repo).

Paper linkage

• Table P uses single-task specialists (one zarr → one tokenizer → one policy per task).
• mt4_N50.zarr is exploratory multitask (shared model); not reported in Table P.

---

3. RoboCasa (official v0.2 HDF5 → Zarr)

Source (G0 protocol)

• RoboCasa v0.2 registry: human_im + mg_im from UT Austin Box (demo_gentex_im128_randcams.hdf5).
• Mix per task: 50 human + 150 MimicGen, subsample seed 0, action_dim = 12.
• Converter: scripts/convert_robocasa_dataset.py.

Tasks (paper Table P)

Task slug	Zarr	Human+MG eps	TopK lock (selection @ seed 2000)
close_drawer	close_drawer_N200.zarr	200	ep-0500 @ SR 0.700
coffee_press_button	coffee_press_button_N200.zarr	200	ep-0500 @ SR 0.600
turn_off_sink_faucet	turn_off_sink_faucet_N200.zarr	200	ep-0500 @ SR 0.580
turn_off_microwave	turn_off_microwave_N200.zarr	200	ep-0500 @ SR 0.620

Each zarr contains ROBOCASA_SOURCE.txt listing exact human/MG HDF5 keys used before MG files were deleted on cluster (disk policy).

HDF5 retention policy on cluster

After Zarr conversion we deleted MimicGen HDF5 to save disk. Human HDF5 for CloseDrawer and CoffeePressButton were kept (~614 MB).
Microwave and sink human HDF5 were also removed post-convert — full training data is in Zarr only for those two tasks.

Obs keys (Zarr)

action, robot0_agentview_{left,right}_rgb, robot0_eye_in_hand_rgb, robot0_eef_pos, robot0_eef_quat, robot0_gripper_qpos (128×128 RGB).

Paper eval protocol (not in this repo)

• Selection TopK: test_start_seed=2000, n_test=50.
• Table P report: literal seeds 10000…10004, each -n 1 --n_test 50, OAT8 + BoN8 vote.

---

Regeneration pointers (code repo)

Suite	Script
MetaWorld gen	scripts/gen_metaworld_data.py, scripts/cluster_gen_metaworld_single_data.sh
MetaWorld split	scripts/split_metaworld_mt4_zarr.py
RoboMimic convert	scripts/convert_robomimic_dataset.py, scripts/prepare_robomimic_{lift,can,square}.sh
RoboCasa convert	scripts/convert_robocasa_dataset.py

Cluster path when uploaded (2026-08): /home/askhabaliev_gs/mipt_paper/oat/data/.

---

Citation

If you use these datasets, cite OAT (ordered action tokenization) and our AAAI paper (anonymous submission at upload time). RoboMimic / MetaWorld / RoboCasa have their own benchmark citations — see respective papers.

---

Changelog

Date	Note
2026-08-13	Initial public upload: all paper Zarr + retained HDF5 from MIPT cluster mipt_paper/oat.