OpenArm Dataset

Quick start

Install

pip install openarm_dataset

Sample usage

Basic:

>>> import openarm_dataset
>>> dataset = openarm_dataset.Dataset("tests/fixture/dataset_0.3.0")
>>> dataset.meta.episodes
[{'id': '0', 'success': False, 'task_index': 0}, {'id': '3', 'success': True, 'task_index': 0}]
>>> dataset.meta.tasks
[{'prompt': 'Run test.', 'description': 'Longer task description if need.'}]
>>> dataset.num_episodes
2

Obs/Action:

>>> obs = dataset.load_obs(0)
>>> list(obs.keys())
['arms/right/qpos', 'arms/right/qvel', 'arms/right/qtorque', 'arms/left/qpos', 'arms/left/qvel', 'arms/left/qtorque', 'lifter/elevation']
>>> obs["arms/right/qpos"].shape
(746, 8)
>>> obs["arms/right/qpos"].head(2)
                                 joint1    joint2    joint3    joint4    joint5    joint6    joint7   gripper
timestamp
2026-02-25 09:04:11.614229214 -0.039352  0.989118 -0.051771  0.735691  0.077740 -0.070724  0.079488 -0.124674
2026-02-25 09:04:11.618732974 -0.039352  0.989118 -0.051771  0.735691  0.077740 -0.070724  0.079488 -0.124674

>>> action = dataset.load_action(0, use_unixtime=True)
>>> list(action.keys())
['arms/right/qpos', 'arms/left/qpos', 'lifter/elevation']
>>> action["arms/right/qpos"].shape
(90, 8)

The available keys reflect what was actually recorded: arm data may contain any of qpos/qvel/qtorque/pose, and embodiments whose files were not recorded (e.g. a declared but unused lifter) are simply absent. Datasets recorded with Cartesian teleoperation expose an 8-dim pose action (position + quaternion + gripper) instead of qpos:

>>> dataset = openarm_dataset.Dataset("tests/fixture/dataset_0.4.0_pose")
>>> action = dataset.load_action(0)
>>> list(action.keys())
['arms/right/pose', 'arms/left/pose', 'lifter/elevation']
>>> list(action["arms/right/pose"].columns)
['x', 'y', 'z', 'qw', 'qx', 'qy', 'qz', 'gripper']

The dataset itself only ever stores the raw recorded representation. To use a specific representation, pass state= to load_obs/load_action/sample: recorded data matching the request is returned as-is, and anything else is converted on the fly with openarm_control kinematics — qpos to pose via FK, pose to qpos via IK (seeded from the episode's recorded obs qpos and tracked along the pose trajectory; a warning reports samples where the solver failed or did not converge). rot6d (position + the first two rotation matrix columns + gripper, 10-dim) is derived from the pose:

>>> action = dataset.load_action(0, state="qpos")  # IK from the recorded pose
>>> list(action.keys())
['arms/right/qpos', 'arms/left/qpos', 'lifter/elevation']
>>> obs = dataset.load_obs(0, state="rot6d")       # FK, then 6D rotation
>>> list(obs["arms/right/rot6d"].columns)
['x', 'y', 'z', 'r11', 'r21', 'r31', 'r12', 'r22', 'r32', 'gripper']

Recorded pose data must be end-effector poses in the MuJoCo model's world frame (the openarm_control Cartesian-teleop convention); the kinematics use the model at its home keyframe, and time-varying lifter elevation is not folded in. Pass kinematics= to Dataset to override the engines built with openarm_dataset.kinematics.create_engines() (e.g. for a different scene or IK tolerances).

Camera:

>>> cameras = dataset.load_cameras(0)
>>> list(cameras.keys())
['wrist_left', 'wrist_right', 'ceiling', 'head']
>>> cam_head = cameras["head"]
>>> cam_head.num_frames
3
>>> cam_head.load_timestamps()
[1772010251.6187909, 1772010251.629775, 1772010251.6634612]
>>> frame = cam_head.get_frame(0)
>>> frame.timestamp
1772010251.6187909
>>> frame.path
PosixPath('.../head/1772010251618790832.jpeg')
>>> frame.load().shape
(600, 960, 3)
>>> for frame in cam_head.frames():
...     pass  # iterate over Frame objects

A camera's frames may be stored either as individual timestamped JPEG files in a directory (episodes/0/cameras/head/<timestamp>.jpeg) or packed into a single uncompressed tar archive (episodes/0/cameras/head.tar). Packing keeps the file count low enough for Hugging Face Hub's storage recommendations. Both layouts expose the same API shown above. For tar-backed cameras, frame.path is a synthetic .../head.tar/<timestamp>.jpeg path that locates the image inside the archive — it is not a real file, so use frame.load() or frame.read_bytes() to access the image data.

Sampling:

>>> samples = dataset.sample(hz=30, episode_index=0)
>>> samples
[Sample(timestamp=1772010251.6202147), Sample(timestamp=1772010251.653548)]
>>> samples[0].timestamp
1772010251.6202147
>>> samples[0].obs["arms/right/qpos"]
array([-0.0393523 ,  0.9891182 , -0.05177076,  0.7356907 ,  0.07774002,
       -0.07072392,  0.07948788, -0.1246737 ], dtype=float32)
>>> samples[0].action["arms/right/qpos"]
array([ 0.03098021,  0.991799  , -0.16657865,  0.96951085,  0.01440866,
        0.14349142, -0.18980259,  0.08221525], dtype=float32)
>>> {name: frame.load().shape for name, frame in samples[0].cameras.items()}
{'wrist_left': (600, 960, 3), 'wrist_right': (600, 960, 3), 'ceiling': (600, 960, 3), 'head': (600, 960, 3)}

Dataset format (v0.4.0)

The current on-disk format written by Dataset.write() (and by the dora-openarm-dataset-recorder):

<root>/
  metadata.yaml                      # version: "0.4.0"
  episodes/<id>/
    obs/arms/<side>/state.parquet    # timestamp + any of qpos/qvel/qtorque/pose
    action/arms/<side>/state.parquet # timestamp + qpos or pose
    obs/lifter/elevation.parquet     # timestamp + value (only if recorded)
    action/lifter/elevation.parquet  # timestamp + value (only if recorded)
    cameras/<name>/<timestamp>.jpeg  # or cameras/<name>.tar

state.parquet is self-describing: its non-timestamp columns (each a list of floats per frame) define the available attributes. qpos/qvel/qtorque have one entry per joint; pose is 8-dim (x, y, z, qw, qx, qy, qz, gripper). Older formats (0.1.0–0.3.0 and unversioned) are still read transparently, and Dataset.write() always produces v0.4.0.

Command-line tools

Validate a dataset:

openarm-dataset-validate <input>

Exits with status 1 if any errors are reported.

Repair a dataset:

openarm-dataset-repair <input> \
    [-o <output>]    # write the repaired dataset here; repairs in place if omitted

Fills isolated single-frame gaps (a null or NaN in a qpos/qvel/ qtorque/pose/value array) by averaging the immediately preceding and following frame values, per array element. Gaps spanning two or more consecutive frames, and gaps at the first or last frame, cannot be averaged and are left untouched with a warning on stderr. The command always exits with status 0; run openarm-dataset-validate afterwards to confirm the result.

Merge multiple datasets:

openarm-dataset-merge <input1> <input2> [<input3> ...] \
    -o <output>    \
    [--symlink]    # create symlinks instead of copying episode data

All input datasets must have the same version, equipment, and frequencies. Tasks are deduplicated by prompt: identical prompts are treated as the same task. Episodes are renumbered sequentially starting from 0.

Convert a dataset:

openarm-dataset-convert <input> <output> \
    [--format {openarm,lerobot_v2.1,lerobot_v3.0,gr00t}] \
    [--camera-format {dir,tar}] # default dir (openarm only); tar packs each \
                                # camera into one .tar archive \
    [--fps INT]                # default 30 (lerobot/gr00t only) \
    [--smoothing-cutoff FLOAT] # default 1.0 (lerobot/gr00t only) \
    [--train-split FLOAT]      # default 0.8 (lerobot/gr00t only) \
    [--success-only]           # lerobot/gr00t only \
    [--state {qpos,pose,rot6d}] # default qpos (lerobot/gr00t only)

The --fps, --smoothing-cutoff, --train-split, --success-only, and --state flags apply only when --format lerobot_v2.1, --format lerobot_v3.0, or --format gr00t. The gr00t format produces a LeRobot v2.1 dataset plus a GR00T-compatible meta/modality.json (see Isaac-GR00T data preparation).

--state selects the arm state representation of the LeRobot output (default: qpos), converting on the fly (see the state= API above) when the dataset stores the other representation. Every arm stream is exported in this one representation, regardless of what was recorded. The converted values are only written to the LeRobot output; the OpenArm dataset itself always keeps the raw recorded data.

Upload a dataset to the Hugging Face Hub:

openarm-dataset-upload <input> \
    --repo-id <user>/<dataset> \
    [--private]                # create the repo as private if it does not exist

The whole dataset directory is uploaded to a dataset repository, creating it if it does not already exist, and tagged with the dataset version. Cameras stored as directories of JPEG files are repacked in place into one .tar archive per camera before uploading, to stay within Hugging Face Hub's file-count recommendations. Repacking is lossless and reversible (openarm-dataset-convert --camera-format dir restores the JPEG-directory layout).

Development

Test

uv sync
uv run pytest

Related links

• 💬 Join the community on Discord
• 📬 Contact us through openarm@enactic.ai

License

Licensed under the Apache License 2.0. See LICENSE.txt for details.

Copyright 2026 Enactic, Inc.

Code of Conduct

All participation in the OpenArm project is governed by our Code of Conduct.