Datasets:

chiyao
/

panda-reach-cube

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
status: string
path: string
bytes: int64
episodes: int64
frames: int64
changed_frames: int64
output: string
green_pixels_removed: int64
max_restored_eef_error_m: double
original_green_pixels: int64
task: string
preserved_columns: list<item: string>
  child 0, item: string
source: string
all_nonpixel_columns_equal: bool
elapsed_seconds: double
renderer_fingerprint: string
method: string
target_marker_visible: bool
new_green_pixels: int64
source_reconstruction: list<item: struct<row: int64, visible_reconstruction_mae: double>>
  child 0, item: struct<row: int64, visible_reconstruction_mae: double>
      child 0, row: int64
      child 1, visible_reconstruction_mae: double
to
{'all_nonpixel_columns_equal': Value('bool'), 'changed_frames': Value('int64'), 'new_green_pixels': Value('int64'), 'original_green_pixels': Value('int64'), 'target_marker_visible': Value('bool'), 'method': Value('string'), 'max_restored_eef_error_m': Value('float64'), 'source_reconstruction': List({'row': Value('int64'), 'visible_reconstruction_mae': Value('float64')}), 'preserved_columns': List(Value('string')), 'green_pixels_removed': Value('int64'), 'task': Value('string'), 'status': Value('string'), 'source': Value('string'), 'output': Value('string'), 'bytes': Value('int64'), 'elapsed_seconds': Value('float64'), 'renderer_fingerprint': 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
              status: string
              path: string
              bytes: int64
              episodes: int64
              frames: int64
              changed_frames: int64
              output: string
              green_pixels_removed: int64
              max_restored_eef_error_m: double
              original_green_pixels: int64
              task: string
              preserved_columns: list<item: string>
                child 0, item: string
              source: string
              all_nonpixel_columns_equal: bool
              elapsed_seconds: double
              renderer_fingerprint: string
              method: string
              target_marker_visible: bool
              new_green_pixels: int64
              source_reconstruction: list<item: struct<row: int64, visible_reconstruction_mae: double>>
                child 0, item: struct<row: int64, visible_reconstruction_mae: double>
                    child 0, row: int64
                    child 1, visible_reconstruction_mae: double
              to
              {'all_nonpixel_columns_equal': Value('bool'), 'changed_frames': Value('int64'), 'new_green_pixels': Value('int64'), 'original_green_pixels': Value('int64'), 'target_marker_visible': Value('bool'), 'method': Value('string'), 'max_restored_eef_error_m': Value('float64'), 'source_reconstruction': List({'row': Value('int64'), 'visible_reconstruction_mae': Value('float64')}), 'preserved_columns': List(Value('string')), 'green_pixels_removed': Value('int64'), 'task': Value('string'), 'status': Value('string'), 'source': Value('string'), 'output': Value('string'), 'bytes': Value('int64'), 'elapsed_seconds': Value('float64'), 'renderer_fingerprint': 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.

Panda Reach and single-cube demonstrations

Two simulated Franka Panda datasets using the policies and simulator fork from cross_embodiment_transfer, with recording settings aligned to the OGBench single-cube training pipeline. The HDF5 files contain actual rendered RGB observations and simulator states.

Dataset	File	Bytes	Episodes	Frames
Reaching	panda_reach_expert.h5	1,513,811,412	85	17,085
Single-cube pick-and-place	panda_cube_single_expert.h5	1,501,609,098	92	18,492

Each episode has 200 actions and 201 observations at 20 Hz. Only demonstrations passing the collector's task-quality checks are retained. Cube demonstrations must reach the sampled placement goal, lift the cube at least 7 cm, grasp it, and reach the retreat stage. Reaching must finish within 3 cm of its target and move the end effector at least 8 cm. The cube success threshold is 4 cm.

Images and previews

pixels contains all simulator-rendered 224 x 224 RGB uint8 frames. The green goal marker is hidden in these images. The green goal remains visible in the diagnostic 3D trajectory videos. Midpoint PNGs are inspection previews; they are not used to construct the dataset.

Task	Rendered RGB	End-effector trajectory
Reaching	RGB video	3D trajectory
Pick-and-place	RGB video	3D trajectory

Download and read

These are raw HDF5 files. Use h5py and import hdf5plugin before accessing the lossless Blosc LZ4-compressed pixels:

from huggingface_hub import snapshot_download
import hdf5plugin
import h5py

snapshot_download(
    repo_id="chiyao/panda-reach-cube",
    repo_type="dataset",
    allow_patterns=["data/*"],
    local_dir="panda_cube_dataset",
)
with h5py.File("panda_cube_dataset/data/panda_reach_expert.h5", "r") as dataset:
    rgb = dataset["pixels"][0]  # (224, 224, 3), uint8
    action = dataset["action"][0]  # (5,), normalized

data/SHA256SUMS lists hashes for both HDF5 files. Per-file manifests and validation reports describe the collection settings, episode quality, and preservation of non-pixel columns when removing the goal marker.

Schema and training settings

Field	Description
`pixels`	Rendered RGB, shape `(frames, 224, 224, 3)`
`action`	Normalized `[dx, dy, dz, dyaw, gripper]`, shape `(frames, 5)`
`observation`	Recorded robot/task state; 26 values for Reach and 30 for cube
`ep_idx`, `ep_len`, `ep_offset`, `step_idx`	Episode membership, lengths, offsets and local indices
`proprio_effector_pos`	World XYZ, meters
`proprio_effector_yaw`	Wrist yaw, radians
`proprio_gripper_opening`	Closing fraction: 0 open, 1 closed
`qpos`, `qvel`, `control`, `robot_action`	Simulator state and original seven-dimensional robot commands
`success`, `phase`, `seed`	Task outcome and collection diagnostics

Action scales are 0.05 m in XYZ and 0.3 rad in yaw. Positive gripper action closes; negative opens. Row t contains observation t and the outgoing action that produces observation t+1. The last action in each episode is NaN padding, matching the reference cube dataset's convention.

The accompanying training scripts use frameskip 5 and two recorded frames (one context and one target), separated by 0.25 seconds. The loader groups five actions into a 25-value input. The merged domain state is XYZ, yaw, and gripper closing fraction, with periodic yaw wrapping.

Reaching keeps wrist yaw fixed and the gripper open. Its goal metadata remains available for diagnostics. Cube metadata includes privileged object and target states. Goal markers are absent from both datasets' rendered image inputs.

Relationship to OGBench and limitations

The robot is a seven-joint Panda with PandaTouchGripper, using the linked robosuite fork and torque-based operational-space control. The camera and dynamics differ from OGBench's UR5e scene. Cube sampling bounds are translated by [-0.425, 0, 0.8] m into the Panda table frame. Reach uses the upstream Reach target bounds. These are successful deterministic feedback demonstrations; they do not reproduce OGBench's noisy Markov oracle or cover failed behavior.

Code, provenance and licenses

Collection code, pinned dependencies, HTC sbatch scripts, validation records and four DWM/LeWM training launchers are in the source branch.

cross_embodiment_transfer: 3988511598a39c8a8a6530055b72609bf1b8b373, Apache-2.0.
tianyudwang/robosuite fork: 3429d3b5ad589d50c4468dd6e78c0a2917f100be, MIT with its MuJoCo notice.
MuJoCo 2.3.7 and NumPy 1.26.4 were used for recording.

The dataset release uses the source project's MIT license, included in LICENSE. Upstream licenses are preserved under licenses/.

Downloads last month: 86

Total file size:

3.02 GB