Datasets:

TingtingDu
/

maniskill3_heldout

Tasks:

Robotics

Tags:

maniskill3

manipulation

held-out-tasks

License:

Dataset card Files Files and versions

xet

Community

Dataset Viewer

Duplicate

Split (1)

train

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
seed: int64
steps: int64
desc: list<item: string>
  child 0, item: string
env_id: string
source: string
success: bool
cameras: struct<view1: struct<radius: double, elev_deg: double, azim_deg: double>, view2: struct<radius: doub (... 174 chars omitted)
  child 0, view1: struct<radius: double, elev_deg: double, azim_deg: double>
      child 0, radius: double
      child 1, elev_deg: double
      child 2, azim_deg: double
  child 1, view2: struct<radius: double, elev_deg: double, azim_deg: double>
      child 0, radius: double
      child 1, elev_deg: double
      child 2, azim_deg: double
  child 2, view3: struct<radius: double, elev_deg: double, azim_deg: double>
      child 0, radius: double
      child 1, elev_deg: double
      child 2, azim_deg: double
  child 3, view4: struct<radius: double, elev_deg: double, azim_deg: double>
      child 0, radius: double
      child 1, elev_deg: double
      child 2, azim_deg: double
geometry_check_px: double
depth_out_of_codec_frac: struct<view1: double, view2: double, view3: double, view4: double>
  child 0, view1: double
  child 1, view2: double
  child 2, view3: double
  child 3, view4: double
10: string
11: string
17: string
1: string
5: string
3: string
9: string
4: string
14: string
8: string
12: string
7: string
18: string
15: string
2: string
16: string
13: string
6: string
to
{'1': Value('string'), '2': Value('string'), '3': Value('string'), '4': Value('string'), '5': Value('string'), '6': Value('string'), '7': Value('string'), '8': Value('string'), '9': Value('string'), '10': Value('string'), '11': Value('string'), '12': Value('string'), '13': Value('string'), '14': Value('string'), '15': Value('string'), '16': Value('string'), '17': Value('string'), '18': Value('string')}
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
              seed: int64
              steps: int64
              desc: list<item: string>
                child 0, item: string
              env_id: string
              source: string
              success: bool
              cameras: struct<view1: struct<radius: double, elev_deg: double, azim_deg: double>, view2: struct<radius: doub (... 174 chars omitted)
                child 0, view1: struct<radius: double, elev_deg: double, azim_deg: double>
                    child 0, radius: double
                    child 1, elev_deg: double
                    child 2, azim_deg: double
                child 1, view2: struct<radius: double, elev_deg: double, azim_deg: double>
                    child 0, radius: double
                    child 1, elev_deg: double
                    child 2, azim_deg: double
                child 2, view3: struct<radius: double, elev_deg: double, azim_deg: double>
                    child 0, radius: double
                    child 1, elev_deg: double
                    child 2, azim_deg: double
                child 3, view4: struct<radius: double, elev_deg: double, azim_deg: double>
                    child 0, radius: double
                    child 1, elev_deg: double
                    child 2, azim_deg: double
              geometry_check_px: double
              depth_out_of_codec_frac: struct<view1: double, view2: double, view3: double, view4: double>
                child 0, view1: double
                child 1, view2: double
                child 2, view3: double
                child 3, view4: double
              10: string
              11: string
              17: string
              1: string
              5: string
              3: string
              9: string
              4: string
              14: string
              8: string
              12: string
              7: string
              18: string
              15: string
              2: string
              16: string
              13: string
              6: string
              to
              {'1': Value('string'), '2': Value('string'), '3': Value('string'), '4': Value('string'), '5': Value('string'), '6': Value('string'), '7': Value('string'), '8': Value('string'), '9': Value('string'), '10': Value('string'), '11': Value('string'), '12': Value('string'), '13': Value('string'), '14': Value('string'), '15': Value('string'), '16': Value('string'), '17': Value('string'), '18': 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.

ManiSkill3 held-out tasks

Three ManiSkill3 tasks that are absent from the training tree of TingtingDu/maniskill3, rendered into the byte-identical on-disk layout that tree uses, so one loader reads both.

task	episodes	source of the demonstrations
`pull_cube`	50	ManiSkill3 motion-planning solver
`place_sphere`	50	ManiSkill3 motion-planning solver
`lift_peg_upright`	50	ManiSkill3 motion-planning solver

Layout

<task>/variation0/episodes/episode<N>/
    view{1..4}/rgb/video.mp4       # 512x512, 20 fps
    view{1..4}/depth.npz           # float16, METRES
    view{1..4}/mask.npz            # uint16 segmentation ids
    view{1..4}/camera_params.json  # per-frame 4x4 cam2world + intrinsics, RLBench convention
    actions.npy                    # [T,8] xyz + quat(xyzw) + gripper openness
    meta.json                      # written LAST -- its presence means the episode is complete
    scene_segments.json, seg_targets.json, handles.json

Cameras are sampled per episode (spherical, look-at the workspace centre, 55 deg FOV) and the written parameters are converted to the RLBench convention -- camera-to-world extrinsics with NEGATIVE focal lengths -- so both sources feed a model the same camera encoding. Each episode passes a geometry self-check at generation time: rendered depth is unprojected through the written parameters and re-projected, and the episode is only written if the round trip agrees to under 0.51 px.

A caveat for closed-loop evaluation

lift_peg_upright is included for offline/perception use but cannot be scored closed-loop by a Cartesian end-effector-pose controller: its motion-planning solution stands the peg up by pivoting it against the table while the gripper itself rotates only ~35 deg, which depends on contact-force transients a pose tracker does not reproduce. Ground-truth replay of the stored demonstrations through such a controller scores 0/20 on it, against 20/20 for pull_cube and place_sphere. Treat it as a harness floor, not a model result.

Generation

scripts/maniskill3_gen.py in ActionImages-Cogen, replaying motion-planning demonstrations frame by frame through set_state_dict.

Downloads last month: 43

Total file size:

2.22 GB