ViaCatalyst/mobile-aloha-cabinet-episode0-lerobot-v3

79.5 MB Mobile ALOHA HDF5 → 3-camera LeRobot v3

Before → after: a native Mobile ALOHA episode with nested robot signals and three padded-JPEG camera arrays becomes a validated, multimodal LeRobot v3.0 dataset. Convert ALOHA HDF5 free →

Community conversion produced by ViaCatalyst BYOD. This repository is not an official upstream release and is not affiliated with the Mobile ALOHA authors or the source-mirror maintainer.

This is a provenance-complete conversion of one pinned Mobile ALOHA Cabinet episode. It demonstrates the awkward parts that generic HDF5 exporters miss: a root-level episode layout, nested joint telemetry, separately logged arm and mobile-base actions, and camera frames stored as padded JPEG byte arrays.

At a glance

Property	Value
Input format / size	Native HDF5 / 79,478,352 bytes
LeRobot format	v3.0
Robot	Mobile ALOHA
Task	Open the top cabinet, store the pot inside it, then close the cabinet
Episodes / frames	1 / 1,500
Duration / FPS	30.0 seconds / 50
Observation modalities	42-D state + 3 RGB camera streams
Action	14-D arm action + 2-D base action
Image resolution	480 × 640 × 3, H.264
Official LeRobot reader	Passed with LeRobot 0.6.0

What this conversion proves

• Native ALOHA episode files do not need to be manually reshaped into robomimic-style data/demo_* groups.
• Padded JPEG arrays are decoded frame by frame, checked for complete alignment, and emitted as three standard LeRobot video features.
• Arm and mobile-base commands remain distinguishable in the 16-D action names while being presented as one training-ready action vector.
• Every published claim is backed by a pinned input revision, SHA-256, feature mapping, machine-readable validation report, and official-reader test.

Features and source mapping

LeRobot feature	dtype	shape	Source mapping
action	float32	[16]	Direct concatenation of root action [14] followed by base_action [2]
observation.state	float32	[42]	observations/qpos, observations/qvel, and observations/effort, 14 values each
observation.images.cam_high	video	[480, 640, 3]	JPEG frames decoded from observations/images/cam_high
observation.images.cam_left_wrist	video	[480, 640, 3]	JPEG frames decoded from observations/images/cam_left_wrist
observation.images.cam_right_wrist	video	[480, 640, 3]	JPEG frames decoded from observations/images/cam_right_wrist
episode_index / frame_index	int64	[1]	One source HDF5 file becomes one episode; frame order is preserved
timestamp	float32	[1]	frame_index / 50 seconds
task_index	int64	[1]	Maps to the Cabinet instruction in meta/tasks.parquet

Original-action preservation

The converter copies the source arm and base command rows directly, with a float32 cast only, and concatenates them in the documented order. It does not replay a policy, regenerate commands, interpolate action values, or infer base motion from images. Camera frames are decoded from source JPEG payloads and re-encoded as H.264, so pixels are training-aligned but not byte-identical to the HDF5 payload.

Validation evidence

The complete evidence is in validation-report.json. All eight critical automated checks passed:

• Dataset metadata counts
• Parquet schema and frame count
• Episode boundaries and 50 Hz timestamp regularity
• Feature dimensions and finite values across 1,500 frames
• LeRobot v3 relational metadata
• Three-stream video decoding and frame alignment
• License and provenance completeness
• Official LeRobot reader smoke test (1 episode / 1,500 frames loaded)

The readiness score is 84/100 — Review recommended. The report discloses two medium-confidence robust-outlier signals: 78.00% of frames for observation.state and 53.13% for action. Because each frame is flagged when any one of 42 state or 16 action dimensions exceeds six median absolute deviations, these rates are not proof of corruption; they are a prompt to inspect robot-specific units, low-variance channels, gripper transitions, and base-motion events before training.

Source, revision, and integrity

• Source mirror: sumo43/mobile-aloha
• Pinned source revision: 855619e6f1a0aa3effca74fdcdfd651bc80bab9a
• Source file: public_aloha_mobile_dataset/aloha_mobile_cabinet/episode_0.hdf5
• Source SHA-256: d82fedbf16a054c126db6e5f7c7dbc4732398ce6d1ffcd242adfbe1d979cfd0e
• Converter adapter: aloha adapter 1.1.0
• Adapter implementation SHA-256: 98ca4808d193a7afb5fe63bf33ca2731f560079addbba5070f07ef3634175111

Audit artifacts: provenance.json, bundle-manifest.json, bundle-manifest.external.json, and UPSTREAM_LICENSE.md.

License and attribution

The pinned source-mirror card declares MIT and links the Mobile ALOHA paper. This conversion retains that declared identifier, source-mirror identity, and paper citation. Conversion does not transfer ownership, create affiliation, or replace the source terms.

Intended use

• Validating native ALOHA/Mobile ALOHA HDF5 ingestion and LeRobot v3 readers
• Testing multimodal loaders with synchronized high, left-wrist, and right-wrist cameras
• Auditing arm-plus-base action mappings before imitation-learning experiments
• Reproducing conversion and validation from a pinned source artifact

Limitations

• This is a community conversion of one mirrored episode, not the complete Mobile ALOHA Cabinet dataset or an official upstream release.
• Users should independently review the mirror's terse license declaration before high-stakes commercial redistribution.
• The 42-D state includes qpos, qvel, and effort; policies may require a narrower state selection.
• JPEG decoding followed by H.264 encoding is not pixel-byte-preserving.
• Automated validation does not measure task success, demonstration quality, or downstream policy performance.

Load with LeRobot

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("ViaCatalyst/mobile-aloha-cabinet-episode0-lerobot-v3")
print(dataset.meta.total_episodes, dataset.meta.total_frames)

Conversion tooling

Converted and validated with the ViaCatalyst BYOD Processing Platform, a free web workflow for converting robotics datasets to LeRobot format. For high-volume ALOHA data, contact ViaCatalyst support through the platform.

Citation

Please cite the original Mobile ALOHA work:

@inproceedings{fu2024mobile,
  title={Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation},
  author={Fu, Zipeng and Zhao, Tony Z. and Finn, Chelsea},
  booktitle={Conference on Robot Learning},
  year={2024}
}