Dataset Viewer
Duplicate
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
rule: list<item: string>
  child 0, item: string
P1: list<item: string>
  child 0, item: string
T3_dropped: list<item: string>
  child 0, item: string
P2: list<item: string>
  child 0, item: string
P3_used: list<item: string>
  child 0, item: string
selected: list<item: string>
  child 0, item: string
note: string
removed: list<item: string>
  child 0, item: string
episodes: list<item: struct<dir: string, src: string, ep: string, scene: string, forced: bool, succ: bool, rol (... 11 chars omitted)
  child 0, item: struct<dir: string, src: string, ep: string, scene: string, forced: bool, succ: bool, role: string>
      child 0, dir: string
      child 1, src: string
      child 2, ep: string
      child 3, scene: string
      child 4, forced: bool
      child 5, succ: bool
      child 6, role: string
to
{'note': Value('string'), 'removed': List(Value('string')), 'episodes': List({'dir': Value('string'), 'src': Value('string'), 'ep': Value('string'), 'scene': Value('string'), 'forced': Value('bool'), 'succ': Value('bool'), 'role': Value('string')})}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 544, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3052, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 558, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 361, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2381, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              rule: list<item: string>
                child 0, item: string
              P1: list<item: string>
                child 0, item: string
              T3_dropped: list<item: string>
                child 0, item: string
              P2: list<item: string>
                child 0, item: string
              P3_used: list<item: string>
                child 0, item: string
              selected: list<item: string>
                child 0, item: string
              note: string
              removed: list<item: string>
                child 0, item: string
              episodes: list<item: struct<dir: string, src: string, ep: string, scene: string, forced: bool, succ: bool, rol (... 11 chars omitted)
                child 0, item: struct<dir: string, src: string, ep: string, scene: string, forced: bool, succ: bool, role: string>
                    child 0, dir: string
                    child 1, src: string
                    child 2, ep: string
                    child 3, scene: string
                    child 4, forced: bool
                    child 5, succ: bool
                    child 6, role: string
              to
              {'note': Value('string'), 'removed': List(Value('string')), 'episodes': List({'dir': Value('string'), 'src': Value('string'), 'ep': Value('string'), 'scene': Value('string'), 'forced': Value('bool'), 'succ': Value('bool'), 'role': Value('string')})}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

alin-baseline-vlac-acro: ArrangeTea training data for two trained failure-detection baselines

This repo has the training data and code for two trained baselines that we compare against. Both were first built for our PickPlaceCounterToCabinet ("cabinet") setup. Here they are re-targeted to the RoboCasa composite task ArrangeTea: "Pick the kettle from the counter and place it on the tray. Then pick the mug from the cabinet and place it on the tray. Then close the cabinet doors." The policy is the GR00T N1.5 RoboCasa multitask checkpoint (gr00t_n1-5/multitask_learning/checkpoint-120000). It runs alone, with no monitor and no intervention.

Scope: this repo has code only: the ArrangeTea conversion scripts, the original baseline code, and the rollout/post-processing code. There are no built training files. You can regenerate them from the raw rollouts with the commands below. The raw ArrangeTea GR00T rollouts (about 45 MB per episode) are not included. They come from our cluster (dataset/gr00t1.5N/ArrangeTea_rec_base_pseed{,2}) and can be shared separately on request. Known limitation of the current pool: every usable non-eval episode has 0 sim-GT failure events (the eval-20 selection took all of them). So VLA-C has no risk positives until more GR00T rollouts are added. Also, no usable episode finishes the full task. For that reason ACRO is converted at the sub-goal level.

The two baselines

  • VLA-Corrector verifier (VLA-C) is our re-implementation of the learned verifier from arXiv 2609.06508. One sample is taken every 2 env steps.
    • Inputs: two RGB views (right third-person + wrist), each reduced to a 54-d pooled color feature; 8-d proprioception; the 7-d executed action.
    • Model: a GRU over the last 8 samples.
    • Outputs: risk (a failure happens within 10 steps or less; the teacher label comes from privileged sim state) and a 5-way stage (APPROACH/ALIGN/GRASP/MOVE/PLACE).
    • Cabinet code: code/vlacorr/. build_data.py builds labels/features, train_c100v2.py trains, vc_runtime.py is the online runtime, detect*.py produce offline alarms.
  • ACRO critic is a survival Q/V critic plus a Q-history gate. It uses the original ACRO code (fit_survival_simpler_qv.fit, imported unchanged).
    • Features: GR00T N1.5 backbone features (vl_mean / backbone_mean + normalized robot state). The action is the next 10 env actions. The query stride is 10.
    • Label: success within horizon H.
    • Cabinet code: code/acro/. acro_features.py extracts features on GPU by re-simulating saved states. acro_fit.py + acro_fit_c100v2.py fit with roles from the manifest. acro_alarms*.py produce offline alarms.
    • Requires the original ACRO repo (acro-robocasa-groot/robocasa_pi05/code) at the path set in acro_fit.py.

Repo layout

README.md
code/convert/      ArrangeTea conversion (NEW, this release):
                     select_at.py        episode selection + eval/scene-overlap/nav/messy exclusion + roles -> train_at_v1.json
                     build_vlac_at.py    VLA-C data (cabinet features; labels adapted to the composite task, see below)
                     train_vlac_at.py    VLA-C training (= train_c100v2.py with --manifest/--data/--out; model/hyper-params unchanged)
                     acro_fit_at.py      ACRO fit at the next-sub-goal level (acro_fit.py imported unchanged)
                     build_all_at.sbatch the whole chain [manifest -> VLA-C build+train -> ACRO features -> ACRO fit]
code/vlacorr/      VLA-C cabinet code (unchanged copies)
code/acro/         ACRO cabinet code (unchanged copies; acro_features.py is used as-is for ArrangeTea)
code/rollout/      GR00T-alone rollout runner: rollout_recover.py + rollout_recover.sbatch (+ online_monitor_rec.py, acro_online.py);
                   online_monitor_full.py is a STUB (our own monitor is not released; unused with MON=0)
code/postproc/     postproc_at.sbatch = trim_success -> composite_flags -> make_gt_composite -> nav_check_qpos;
                   messy_filter.py; find_stuck_composite.py / gt_events.py (sim-GT failure events); step0_usable.py (how eval eps were chosen)
eval_exclusion/    items20.txt / selected20.json = the 20 ArrangeTea eval episodes (never used for training);
                   cabinet_train100_v2_reference.json = the cabinet manifest these builders were made for
(outputs of the commands below: data/train_at_v1.json, data/vlacorr_at/, data/acro_at/; not uploaded)

Source rollouts

  • ArrangeTea_rec_base_pseed/ep000000-000038, ep000079-000118 and ArrangeTea_rec_base_pseed2/ep000039-000078 (120 dirs; pseed/ep000039 is empty).
  • Scene seed = 1000 + ep, i.e. 1000-1118. rollout_recover.sbatch with TASK=ArrangeTea MON=0 PERTURB= SEED0=1000 POLICY_SEED=1, horizon 2250 steps (the task default).
  • Files per episode used by the conversion: log.npz, meta.json, ep_meta.json, model.xml, agentview_right.mp4, eye_in_hand.mp4, composite_flags.npz, composite_gt.json, state_gt.json. Plus nav_episodes.json (navigation list).
  • The raw rollouts are not in this repo. They can be shared separately. The conversion scripts read them from --root.

Eval-exclusion rule and the 20 eval episodes

An episode is excluded from training if it overlaps an eval episode (eval_exclusion/items20.txt) in any of these ways:

  • the same episode id;
  • the same scene seed;
  • the same signature (layout_id, style_id, object models, robot base position rounded to 1 cm);
  • the same (layout, style, robot base position).

The last condition is stricter than the cabinet c100v2 rule, which used seed + signature only. Duplicate scenes inside the training set are dropped too. Also excluded:

  • navigation episodes: the robot base moved > 10 cm or > 15 deg (nav_check_qpos.py, list nav_episodes.json);
  • messy GT episodes: state-GT progress drops by >= 10 and comes back 3+ times (messy_filter.py);
  • episodes that are not fully post-processed.

The 20 eval episodes (seed = 1000 + ep):

pseed  ep000000 ep000003 ep000013 ep000015 ep000018 ep000019 ep000020 ep000021 ep000023 ep000029
       ep000031 ep000034 ep000088 ep000102 ep000103 ep000106
pseed2 ep000041 ep000043 ep000047 ep000074

How the 20 were chosen (selected20.json, step0_usable.py): from the usable episodes (not nav, not messy, GR00T failed):

  1. Table-3 episodes;
  2. then episodes with at least one sim-GT failure event;
  3. then the rest, in seed order.

On the current 120 rollouts, select_at.py skips 71 navigation, 20 eval, 7 not post-processed or empty, and 4 messy episodes. That leaves 18 training episodes (scene seeds 1044-1109). Roles are assigned by seed order: i%5==2 -> validation, i%5==4 -> calibration, otherwise train. That gives 11 / 4 / 3.

Labels on ArrangeTea (what each builder needs)

  • VLA-C needs per-step failure events. It does not need episode success. The features are exactly the cabinet ones. The labels are adapted to the composite task:

    • risk: the events are the same sim-GT events as the eval GT (find_stuck_composite.events per object, with the same filtering as gt_events.py):

      • gt_drop = a held kettle/mug is lost for 10 frames;
      • gt_empty = the gripper closes for 15 frames with no contact while the hand is within 15 cm of the object;
      • events on an object that is already in the target are ignored.

      risk = 1 if the next event is 10 steps away or less; 0 if it is more than 20 steps away, if there is none, or after full success; -1 (ignored) otherwise.

    • stage is computed w.r.t. the active sub-goal. The active sub-goal is:

      • the object in gripper contact; otherwise
      • the nearest kettle/mug that is not yet in the target; otherwise
      • the cabinet door.

      For a PnP sub-goal, the cabinet rules apply. For the door: PLACE = door closed, MOVE = door joint moving, ALIGN = hand within 10 cm of the handle.

    • An extra field sub (0 kettle, 1 mug, 2 door) is stored. The trainer does not use it.

  • ACRO needs "task success within H" per sequence. GR00T almost never finishes the full ArrangeTea task: 3/120 rollouts, all of them navigation episodes, and 0/18 training episodes. So acro_fit_at.py cuts each episode into next-sub-goal segments using the order-free composite GT (composite_gt.json first_done):

    • segment j runs from the (j-1)-th to the j-th sub-goal completion;
    • a segment that ends in a completion is a success;
    • the last open segment, up to T = 2250, is a failure (censored);
    • H = 750 per segment (the same as cabinet);
    • failure segments shorter than H are skipped (ACRO rule).

    The features are the unchanged acro_features.py output on the full episode. The query times are re-based to the segment start. This is a documented adaptation, not the original episode-level label.

Adding GR00T rollouts

Use new scene seeds. Existing ArrangeTea seeds are 1000-1118, and the eval eps are inside that range. We suggest START >= 200 (seed 1200+). select_at.py re-checks scene overlap with the eval 20.

# from code/rollout (paths inside rollout_recover.sbatch are our cluster's; set D/OUT/CKPT for yours). 10 eps per job:
for S in $(seq 200 10 549); do
  TASK=ArrangeTea MON=0 PERTURB= SEED0=1000 POLICY_SEED=1 N=10 START=$S SUFFIX=_rec_base_new \
    sbatch --qos=<qos> --nice=200 rollout_recover.sbatch
done
# post-process (robocasa env): trim -> composite_flags -> composite GT -> nav check
cd <gr00t1.5N dir> && SRC=ArrangeTea_rec_base_new EP0=200 EP1=549 G=<repo>/code/postproc sbatch <repo>/code/postproc/postproc_at.sbatch
# select (adds messy filter + eval exclusion) and rebuild everything:
python code/convert/select_at.py --root <gr00t1.5N dir> --out train_at_v2.json ArrangeTea_rec_base_pseed ArrangeTea_rec_base_pseed2 ArrangeTea_rec_base_new

composite_flags.py and make_gt_composite.py need the RoboCasa repo layout (composite_flags.py imports robocasa.* from its parent dir).

How to regenerate the training data and train (exact commands)

Run from the repo root.

ROOT=/path/to/gr00t1.5N            # dir holding ArrangeTea_rec_base_pseed{,2}/ep*/ (log.npz, *.mp4, model.xml, ep_meta.json, composite_*.{npz,json}, state_gt.json)
# 1. manifest (eval/scene/nav/messy exclusion, roles, labels)            [groot env]
python code/convert/select_at.py --root $ROOT --nav /path/to/nav_episodes.json --out data/train_at_v1.json
# 2. VLA-C data + training (CPU ok)                                       [groot env]
python code/convert/build_vlac_at.py --manifest data/train_at_v1.json --out data/vlacorr_at/data
python code/convert/train_vlac_at.py --manifest data/train_at_v1.json --data data/vlacorr_at/data --out data/vlacorr_at
# 3. ACRO GR00T features (GPU, re-simulates saved states)                 [groot env, MUJOCO_GL=egl]
python -c "import json;print('\n'.join(e['dir'] for e in json.load(open('data/train_at_v1.json'))['episodes']))" > data/acro_at_dirs.txt
MUJOCO_GL=egl python code/acro/acro_features.py --skip_done --out data/acro_at/feat_at $(cat data/acro_at_dirs.txt)
# 4. ACRO fit (next-sub-goal segments)                                    [acro env]
python code/convert/acro_fit_at.py --manifest data/train_at_v1.json --feat data/acro_at/feat_at --out data/acro_at/fit_at
# or all of it on SLURM:
ROOT=$ROOT OUT=$(pwd)/data sbatch code/convert/build_all_at.sbatch

ACRO feature extraction is the unchanged cabinet acro_features.py. It reads the task (ArrangeTea) from each episode's meta.json and restores the scene from model.xml + ep_meta.json. With the current 18 episodes, step 2 trains only the stage head, because there are no risk positives (val_auroc = null). Add rollouts first.

Environments

  • robocasa: RoboCasa + robosuite (MuJoCo with EGL). Used for post-processing (composite_flags, composite GT, nav check).
  • groot: Isaac-GR00T N1.5 (our RoboCasa-compat fork Isaac-GR00T-robocasa), robocasa, torch, imageio. Used for rollouts, the VLA-C build/train, and ACRO features.
  • acro: torch + the original ACRO repository. Used for the ACRO fit.

All paths in the scripts are absolute paths on our cluster (/home/mlnha/...). Adjust them (or GROOT_DIR, --root, --nav).

Where the files came from (our tree)

here ours
code/vlacorr/* dataset/gr00t1.5N/vlacorr/
code/acro/* output/progress/sol/gr00t1.5N/acro/
code/rollout/* dataset/robocasa_sim/groot/ (online_monitor_full.py replaced by a stub)
code/postproc/* dataset/robocasa_sim/groot/ (trim_success, composite_flags, make_gt_composite, nav_check_qpos, stuck/find_stuck_composite), output/progress/sol/at_detrec/ (gt_events, step0_usable), messy_filter = drops() of output/progress/sol/agx_eval_jitter.py, postproc_at.sbatch from output/progress/sol/bench20/postproc_atwfc.sbatch
code/convert/* new (output/progress/sol/at_baselines/)
eval_exclusion/* output/progress/sol/at_detrec/items20.txt, selected20.json; dataset/gr00t1.5N/train100_v2.json

ν•œκ΅­μ–΄ μš”μ•½

  • μ½”λ“œλ§Œ 올린 μ €μž₯μ†Œλ‹€:
    • ArrangeTea λ³€ν™˜ 슀크립트 (편 κ³ λ₯΄κΈ° + 평가·μž₯λ©΄ κ²ΉμΉ¨ μ œμ™Έ, composite GT ν•˜μœ„ λͺ©ν‘œ 라벨, VLA-C 데이터 λΉŒλ”, ACRO ν•˜μœ„ λͺ©ν‘œ λ‹¨μœ„ fit)
    • μ›λž˜ 베이슀라인 μ½”λ“œ
    • λ‘€μ•„μ›ƒΒ·ν›„μ²˜λ¦¬ μ½”λ“œ
  • μ›μ‹œ 둀아웃은 ν¬ν•¨ν•˜μ§€ μ•Šμ•˜κ³ , λ”°λ‘œ κ³΅μœ ν•  수 μžˆλ‹€. μœ„ λͺ…λ ΉμœΌλ‘œ λ‹€λ₯Έ μ„œλ²„μ—μ„œ ν•™μŠ΅ 데이터λ₯Ό λ‹€μ‹œ λ§Œλ“€ 수 μžˆλ‹€.
  • ν•™μŠ΅μ—μ„œ λΉΌλŠ” 편:
    • 평가 20편과 편·μž₯λ©΄ (seed / μ„œλͺ… / 배치+μŠ€νƒ€μΌ+λ‘œλ΄‡ μœ„μΉ˜) 이 κ²ΉμΉ˜λŠ” 편
    • 이동 편 (>10cm / 15도)
    • messy GT 편
  • ν˜„μž¬ λ‘€μ•„μ›ƒμœΌλ‘œλŠ” ν•™μŠ΅ 18편이 λ‚¨λŠ”λ‹€. 그런데 이 18νŽΈμ—λŠ” μ‹€νŒ¨ 사건이 μ—†μ–΄μ„œ VLA-C μœ„ν—˜ 양성이 0이닀. 둀아웃 μΆ”κ°€κ°€ ν•„μš”ν•˜λ‹€ (seed 1200+, λͺ…령은 μœ„μ— μžˆλ‹€).
Downloads last month
84