Dataset Viewer
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
task: string
instance_id: int64
rollout_id: int64
steps: int64
fps: int64
sample_every: int64
goal_labels: list<item: string>
child 0, item: string
goal_sentences: list<item: string>
child 0, item: string
goal_status: list<item: struct<step: int64, satisfied: list<item: null>, unsatisfied: list<item: string>>>
child 0, item: struct<step: int64, satisfied: list<item: null>, unsatisfied: list<item: string>>
child 0, step: int64
child 1, satisfied: list<item: null>
child 0, item: null
child 2, unsatisfied: list<item: string>
child 0, item: string
held: list<item: struct<step: int64, left: string, right: string>>
child 0, item: struct<step: int64, left: string, right: string>
child 0, step: int64
child 1, left: string
child 2, right: string
distances: list<item: struct<step: int64, candle.n.01_1: double, butter_cookie.n.01_1: double, candle.n.01_3: d (... 503 chars omitted)
child 0, item: struct<step: int64, candle.n.01_1: double, butter_cookie.n.01_1: double, candle.n.01_3: double, bow. (... 491 chars omitted)
child 0, step: int64
child 1, candle.n.01_1: double
child 2, butter_cookie.n.01_1: double
child 3, candle.n.01_3: double
child 4, bow.n.08_1: double
child 5, swiss_cheese.n.01_3: double
child 6, wicker_basket.n.01_4: double
child 7, butter_cookie.n.01_2: double
child 8, swiss_cheese.n.01_4: double
child 9, wicker_basket.n.01_1: double
child 10, swiss_cheese.n.01_2: double
child 11, floor.n.01_1: double
child 12, swiss_cheese.n.01_1: double
child 13, candle.n.01_4: double
child 14, butter_cookie.n.01_3: double
child 15, bow.n.08_3: double
child 16, bow.n.08_4: double
child 17, bow.n.08_2: double
child 18, wicker_basket.n.01_3: double
child 19, butter_cookie.n.01_4: double
child 20, wicker_basket.n.01_2: double
child 21, candle.n.01_2: double
child 22, table.n.02_1: double
commands: list<item: struct<step: int64, base: list<item: double>>>
child 0, item: struct<step: int64, base: list<item: double>>
child 0, step: int64
child 1, base: list<item: double>
child 0, item: double
unachieved: list<item: null>
child 0, item: null
success: bool
time: struct<simulator_steps: int64, simulator_time: double, normalized_time: double>
child 0, simulator_steps: int64
child 1, simulator_time: double
child 2, normalized_time: double
normalized_agent_distance: struct<base: double, left: double, right: double>
child 0, base: double
child 1, left: double
child 2, right: double
agent_distance: struct<base: double, left: double, right: double>
child 0, base: double
child 1, left: double
child 2, right: double
q_score: struct<final: double>
child 0, final: double
to
{'task': Value('string'), 'instance_id': Value('int64'), 'rollout_id': Value('int64'), 'steps': Value('int64'), 'success': Value('bool'), 'agent_distance': {'base': Value('float64'), 'left': Value('float64'), 'right': Value('float64')}, 'normalized_agent_distance': {'base': Value('float64'), 'left': Value('float64'), 'right': Value('float64')}, 'q_score': {'final': Value('float64')}, 'time': {'simulator_steps': Value('int64'), 'simulator_time': Value('float64'), 'normalized_time': Value('float64')}}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, 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 127, 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 483, 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 2840, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, 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 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, 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 343, 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 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
task: string
instance_id: int64
rollout_id: int64
steps: int64
fps: int64
sample_every: int64
goal_labels: list<item: string>
child 0, item: string
goal_sentences: list<item: string>
child 0, item: string
goal_status: list<item: struct<step: int64, satisfied: list<item: null>, unsatisfied: list<item: string>>>
child 0, item: struct<step: int64, satisfied: list<item: null>, unsatisfied: list<item: string>>
child 0, step: int64
child 1, satisfied: list<item: null>
child 0, item: null
child 2, unsatisfied: list<item: string>
child 0, item: string
held: list<item: struct<step: int64, left: string, right: string>>
child 0, item: struct<step: int64, left: string, right: string>
child 0, step: int64
child 1, left: string
child 2, right: string
distances: list<item: struct<step: int64, candle.n.01_1: double, butter_cookie.n.01_1: double, candle.n.01_3: d (... 503 chars omitted)
child 0, item: struct<step: int64, candle.n.01_1: double, butter_cookie.n.01_1: double, candle.n.01_3: double, bow. (... 491 chars omitted)
child 0, step: int64
child 1, candle.n.01_1: double
child 2, butter_cookie.n.01_1: double
child 3, candle.n.01_3: double
child 4, bow.n.08_1: double
child 5, swiss_cheese.n.01_3: double
child 6, wicker_basket.n.01_4: double
child 7, butter_cookie.n.01_2: double
child 8, swiss_cheese.n.01_4: double
child 9, wicker_basket.n.01_1: double
child 10, swiss_cheese.n.01_2: double
child 11, floor.n.01_1: double
child 12, swiss_cheese.n.01_1: double
child 13, candle.n.01_4: double
child 14, butter_cookie.n.01_3: double
child 15, bow.n.08_3: double
child 16, bow.n.08_4: double
child 17, bow.n.08_2: double
child 18, wicker_basket.n.01_3: double
child 19, butter_cookie.n.01_4: double
child 20, wicker_basket.n.01_2: double
child 21, candle.n.01_2: double
child 22, table.n.02_1: double
commands: list<item: struct<step: int64, base: list<item: double>>>
child 0, item: struct<step: int64, base: list<item: double>>
child 0, step: int64
child 1, base: list<item: double>
child 0, item: double
unachieved: list<item: null>
child 0, item: null
success: bool
time: struct<simulator_steps: int64, simulator_time: double, normalized_time: double>
child 0, simulator_steps: int64
child 1, simulator_time: double
child 2, normalized_time: double
normalized_agent_distance: struct<base: double, left: double, right: double>
child 0, base: double
child 1, left: double
child 2, right: double
agent_distance: struct<base: double, left: double, right: double>
child 0, base: double
child 1, left: double
child 2, right: double
q_score: struct<final: double>
child 0, final: double
to
{'task': Value('string'), 'instance_id': Value('int64'), 'rollout_id': Value('int64'), 'steps': Value('int64'), 'success': Value('bool'), 'agent_distance': {'base': Value('float64'), 'left': Value('float64'), 'right': Value('float64')}, 'normalized_agent_distance': {'base': Value('float64'), 'left': Value('float64'), 'right': Value('float64')}, 'q_score': {'final': Value('float64')}, 'time': {'simulator_steps': Value('int64'), 'simulator_time': Value('float64'), 'normalized_time': Value('float64')}}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
BEHAVIOR-1K 2026 Challenge — evaluation README and self-evaluation results
This repository is the stable link for our 2026 BEHAVIOR Challenge submission. It holds the self-evaluation rollouts and the instructions the organizers need to run our policy.
Policy image
| URI | docker.io/kimgeuneon/behavior-2026-policy:v2-2026-260904 |
| Digest | sha256:1d9dfa41dc4a043eb918e0f63a151036af89bbab776ed644081480e980ade678 |
| Registry | Docker Hub, public |
| Authentication | None. Anonymous pull works; no token or account is needed. |
| Size | ~164 GB as reported by docker images; ~75 GB of image content. Allow time and disk for the pull. |
| Architecture | linux/amd64 |
| GPU | Required. One GPU; the policy keeps one checkpoint resident. |
The digest is the same image the self-evaluation below was produced with, so pinning it reproduces exactly what we measured.
docker pull docker.io/kimgeuneon/behavior-2026-policy@sha256:1d9dfa41dc4a043eb918e0f63a151036af89bbab776ed644081480e980ade678
Running the policy server
The image has an entrypoint; no command is needed.
docker run --rm --gpus device=0 -p 127.0.0.1:8000:8000 \
docker.io/kimgeuneon/behavior-2026-policy:v2-2026-260904
- Serves the websocket policy protocol on port 8000 (
omnigibson.eval'sWebsocketPolicyServer): a metadata frame on connect, one msgpack request per step,{"reset": true}between rollouts. GET /healthzon the same port returns 200 once the checkpoint is loaded. Wait for it before opening the socket — the first load takes on the order of a minute.- One container serves one evaluator. For parallel evaluation, start one container per concurrent rollout on its own GPU and port.
Capacity
- Peak GPU memory measured over 8,839 samples of a 1000-rollout sweep: median 7.3 GiB, maximum 13.6 GiB, inside the 24 GB budget. Measured on a 96 GB card; we have not yet re-measured against a hard 24 GB cap.
- CPU use by the policy is negligible (~0.04 cores per lane). The simulator, which you run, is the CPU-heavy side.
- The image carries five checkpoints and loads one at a time, switching on the
task_idin the observation. Only one is ever resident.
Self-evaluation results
Produced with the image digest above.
| Run | 2025-1st-full-260904 |
| Rollouts | 1000 = 100 tasks x 10 instances x 1 rollout |
| Mean q-score | 0.018830 |
| Successes | 0 |
| Missing rollouts | none |
Layout
<task>/json/<task>_<instance>_<rollout>.json metrics, 1000 files
<task>/videos/<task>_<instance>_<rollout>.mp4 recordings, 1000 files
<task>/timeline/... per-step predicate timeline (ours, not required)
run.json manifest: images by digest, lanes, sweep
summary.json aggregate and per-task q-scores
Every path in the submitted bundle's videos.txt resolves directly against this
repository:
https://huggingface.co/datasets/geuneon/behavior-2026-selfeval/resolve/main/<path from videos.txt>
Evaluation setup
The sweep used the official harness with default settings: BEHAVIOR-1K v3.9.2,
the public test instances, the default episode timeout (1.5x the mean human demo
length, in simulator steps), and --write-video. The policy received RGB, depth
and proprioception only.
The env wrapper (wrapper.py in the bundle) records a predicate timeline for our
own debugging. It observes; it does not change the environment, the observation
the policy receives, or the score.
python -m omnigibson.eval.eval \
--task-name <task> --instance-indices 0 1 2 3 4 5 6 7 8 9 \
--num-rollouts 1 --host 127.0.0.1 --port 8000 \
--env-wrapper ludo_simwrap.predicate_timeline.PredicateTimelineWrapper \
--robot-config r1pro.yaml \
--output-dir <out> --write-video
Contact
Questions about the image, the registry, or these results: see the contact email on the submission form.
- Downloads last month
- 374