Hugging Face's logo

Datasets:
tklee-3012
/
cross-embodiment

Tasks:
Robotics
Size:
1K<n<10K
Tags:
maniskill
cross-embodiment
imitation-learning
vision-language-action
Dataset card Files
xet
 Community
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
env_info: struct<env_id: string, env_kwargs: struct<robot_uids: string, obs_mode: string, render_mode: string, (... 147 chars omitted)
  child 0, env_id: string
  child 1, env_kwargs: struct<robot_uids: string, obs_mode: string, render_mode: string, control_mode: string, sim_backend: (... 85 chars omitted)
      child 0, robot_uids: string
      child 1, obs_mode: string
      child 2, render_mode: string
      child 3, control_mode: string
      child 4, sim_backend: string
      child 5, render_backend: string
      child 6, sensor_configs: struct<width: int64, height: int64>
          child 0, width: int64
          child 1, height: int64
  child 2, max_episode_steps: int64
commit_info: null
episodes: list<item: struct<episode_id: int64, episode_seed: int64, control_mode: string, elapsed_steps: int64 (... 67 chars omitted)
  child 0, item: struct<episode_id: int64, episode_seed: int64, control_mode: string, elapsed_steps: int64, reset_kwa (... 55 chars omitted)
      child 0, episode_id: int64
      child 1, episode_seed: int64
      child 2, control_mode: string
      child 3, elapsed_steps: int64
      child 4, reset_kwargs: struct<options: null, seed: int64>
          child 0, options: null
          child 1, seed: int64
      child 5, success: bool
source_type: string
source_desc: string
kept: int64
task: string
per_seed: list<item: struct<seed: int64, kept: bool, success: string, steps: int64>>
  child 0, item: struct<seed: int64, kept: bool, success: string, steps: int64>
      child 0, seed: int64
      child 1, kept: bool
      child 2, success: string
      child 3, steps: int64
minutes: double
max_len: int64
robot: string
attempts: int64
mean_len_kept: double
to
{'robot': Value('string'), 'task': Value('string'), 'kept': Value('int64'), 'attempts': Value('int64'), 'max_len': Value('int64'), 'mean_len_kept': Value('float64'), 'minutes': Value('float64'), 'per_seed': List({'seed': Value('int64'), 'kept': Value('bool'), 'success': Value('string'), 'steps': Value('int64')})}
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
              env_info: struct<env_id: string, env_kwargs: struct<robot_uids: string, obs_mode: string, render_mode: string, (... 147 chars omitted)
                child 0, env_id: string
                child 1, env_kwargs: struct<robot_uids: string, obs_mode: string, render_mode: string, control_mode: string, sim_backend: (... 85 chars omitted)
                    child 0, robot_uids: string
                    child 1, obs_mode: string
                    child 2, render_mode: string
                    child 3, control_mode: string
                    child 4, sim_backend: string
                    child 5, render_backend: string
                    child 6, sensor_configs: struct<width: int64, height: int64>
                        child 0, width: int64
                        child 1, height: int64
                child 2, max_episode_steps: int64
              commit_info: null
              episodes: list<item: struct<episode_id: int64, episode_seed: int64, control_mode: string, elapsed_steps: int64 (... 67 chars omitted)
                child 0, item: struct<episode_id: int64, episode_seed: int64, control_mode: string, elapsed_steps: int64, reset_kwa (... 55 chars omitted)
                    child 0, episode_id: int64
                    child 1, episode_seed: int64
                    child 2, control_mode: string
                    child 3, elapsed_steps: int64
                    child 4, reset_kwargs: struct<options: null, seed: int64>
                        child 0, options: null
                        child 1, seed: int64
                    child 5, success: bool
              source_type: string
              source_desc: string
              kept: int64
              task: string
              per_seed: list<item: struct<seed: int64, kept: bool, success: string, steps: int64>>
                child 0, item: struct<seed: int64, kept: bool, success: string, steps: int64>
                    child 0, seed: int64
                    child 1, kept: bool
                    child 2, success: string
                    child 3, steps: int64
              minutes: double
              max_len: int64
              robot: string
              attempts: int64
              mean_len_kept: double
              to
              {'robot': Value('string'), 'task': Value('string'), 'kept': Value('int64'), 'attempts': Value('int64'), 'max_len': Value('int64'), 'mean_len_kept': Value('float64'), 'minutes': Value('float64'), 'per_seed': List({'seed': Value('int64'), 'kept': Value('bool'), 'success': Value('string'), 'steps': Value('int64')})}
              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.

ManiSkill Cross-Embodiment 3×5

Demonstrations for a cross-embodiment transfer benchmark built on ManiSkill 3: three robot arms × five table-top tasks, with one (robot, task) cell per robot held out (no data). A policy trained on the 12 training cells is tested on the 3 held-out cells, where the task was only ever demonstrated by the other robots.

PickCube PushCube PullCube StackCube PlaceSphere
Panda train train train held-out train
XArm6-Robotiq train held-out train train train
WidowX-AI held-out train train train train

All three robots share the same scene, the same policy camera and the same 7-D action (end-effector delta position + delta rotation + gripper). This dataset contains the 12 training cells only.

Contents 

demos/<robot>/<task>/scripted200/
    trajectory.h5     200 successful episodes (traj_0 ... traj_199)
    trajectory.json   ManiSkill metadata: env kwargs, per-episode seed, length, success
    summary.json      generation log: every attempted seed, kept or not, episode length
matrix3x5_spec.json   training spec (path, robot, instruction per cell)

2,400 episodes, 147,817 transitions, about 2.5 GB.

Robot Task Episodes Attempts Mean length (steps)
panda PickCube-v1 200 200 35.1
panda PushCube-v1 200 204 55.5
panda PullCube-v1 200 200 54.1
panda PlaceSphere-v1 200 200 72.2
xarm6_robotiq PickCube-v1 200 200 35.8
xarm6_robotiq PullCube-v1 200 200 54.6
xarm6_robotiq StackCube-v1 200 203 81.3
xarm6_robotiq PlaceSphere-v1 200 200 74.4
widowxai PushCube-v1 200 252 53.9
widowxai PullCube-v1 200 248 68.3
widowxai StackCube-v1 200 319 81.0
widowxai PlaceSphere-v1 200 200 72.9

Per-episode fields (HDF5, ManiSkill RecordEpisode format)

Key Shape Content
obs/sensor_data/base_camera/rgb (T+1, 256, 256, 3) uint8 policy camera (gzip)
obs/agent/qpos, obs/agent/qvel (T+1, dof) robot joint state
obs/extra/tcp_pose (T+1, 7) gripper TCP pose (xyz + wxyz quaternion)
actions (T, 7) pd_ee_delta_pose: xyz delta, rotation delta (always 0 here), gripper
env_states/actors/<name> (T+1, 13) object poses and velocities (cube, cubeA/cubeB, sphere, bin, goal_region, goal_site)
success, rewards, terminated, truncated (T,) per-step flags

Gripper command convention: Panda and WidowX-AI use +1 open / −1 closed; XArm6-Robotiq uses its raw knuckle target (0 open, 0.81 closed). Instructions per task are in matrix3x5_spec.json.

How the data was generated

  • Simulator: ManiSkill 3.0.1, SAPIEN 3.0.3, CPU physics, software rendering, 256 px RGB, control at 20 Hz.
  • Policy: one privileged scripted policy per task, identical for all robots. It reads object poses from the simulator and commands world-frame TCP deltas (≤ 4 cm per step) and gripper open/close: approach above the object, descend, grasp, lift, carry, release; Push/Pull drive the gripper from behind the cube with closed-loop re-aiming. The gripper is never rotated. Only successful episodes of ≤ 250 steps are kept.
  • Seeds: training seeds start at 0 (0–318). Evaluation scenes in our experiments use seeds 1000+.
  • Success: ManiSkill's own success for Push, Pull, Stack and PlaceSphere. For PickCube: cube grasped and lifted ≥ 5 cm (PickCube's goal sphere is hidden from the cameras, so an image-based policy cannot know it).
  • WidowX-AI needs three changes that the code applies at runtime (no ManiSkill files edited): an end-effector delta-pose controller (ManiSkill ships joint control only), both gripper carriages driven symmetrically, and a table mount at x = −0.35 m (at ManiSkill's default −0.615 m it cannot reach the shared object region). PickCube uses Panda's configuration (cube size, spawn region, camera) for all robots.

To reproduce the WidowX-AI setup when replaying or evaluating, use scripts/embodiments.py from the code below.

Code

Generation, training (UniVLA and object-centric UniVLA) and evaluation code: branch object-centric of https://github.com/anh147/lasct (scripts/gen_scripted_demos.py, scripts/slurm/matrix_train.sh, scripts/slurm/matrix_eval.sh).

from huggingface_hub import snapshot_download
path = snapshot_download("tklee-3012/cross-embodiment", repo_type="dataset")
# put <path>/demos where the code expects demos/ (or symlink), then train with configs/matrix3x5_spec.json

Licence

Rendered from ManiSkill (Apache-2.0) scenes and robot assets (Franka Panda, UFactory XArm6 with Robotiq 2F-85, Trossen WidowX-AI). Check the licences of the robot assets before redistribution.
Downloads last month
68
Total file size:
2.73 GB