ViaCatalyst/robomimic-can-ph-lerobot-v3

RoboTurk Can PH HDF5 → verified LeRobot v3 episodes

Before → after: nested RoboTurk/robomimic HDF5 demonstrations become validated LeRobot v3.0 episodes while original actions and episode boundaries are retained. Convert robomimic HDF5 free →

Community conversion produced by ViaCatalyst BYOD. This repository is not an official upstream release and is not affiliated with the robomimic or RoboTurk authors.

This is a compact, provenance-complete conversion of the first 10 demonstrations from the pinned robomimic Can proficient-human (PH) low-dimensional v1.5 file. The source demonstrations were collected through RoboTurk; the output is a reproducible LeRobot v3 reference dataset and validation-quality example.

At a glance

Property	Value
Input format / size	robomimic HDF5 / 46,889,752 bytes
LeRobot format	v3.0
Robot	Franka Emika Panda
Task	Place the can in the target bin
Episodes / frames	10 / 1,160
Duration / FPS	58.0 seconds / 20
Observation modality	Low-dimensional state
Action / state dimension	7 / 57
Episode length range	98–134 frames
Official LeRobot reader	Passed with LeRobot 0.6.0

Features and source mapping

LeRobot feature	dtype	shape	Source mapping
action	float32	[7]	Direct per-frame copy from each demonstration's actions dataset, with a float32 cast only
observation.state	float32	[57]	Deterministic concatenation of numeric obs/* vectors; component names are retained in meta/info.json
episode_index	int64	[1]	Preserved data/demo_* HDF5 boundary
frame_index	int64	[1]	Zero-based frame position inside each demonstration
timestamp	float32	[1]	frame_index / 20 seconds
task_index	int64	[1]	Maps to Place the can in the target bin in meta/tasks.parquet

Original-action preservation

The conversion reads the original actions array for every selected demonstration and writes the corresponding row directly to action. It does not replay a learned policy, regenerate or interpolate actions, relabel success, or flatten episode boundaries. This matters for RoboTurk-derived benchmarks: the published values remain the demonstrations, not policy rollouts masquerading as source actions.

The selected low-dimensional file contains no aligned RGB, depth, audio, or video observations; none are claimed by this release.

Validation evidence

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

• Dataset metadata counts
• Parquet schema and frame count
• Ten episode boundaries and timestamp regularity
• Feature dimensions and finite values across 1,160 frames
• LeRobot v3 relational metadata
• License and provenance completeness
• Official LeRobot reader smoke test (10 episodes / 1,160 frames loaded)

The readiness score is 92/100 — Review recommended. One medium-confidence signal is disclosed rather than hidden: observation.state has a 76.12% robust-outlier frame rate. Because a frame is flagged when any one of 57 heterogeneous state dimensions crosses the six-MAD threshold, this is not proof of corruption. Review the named feature distributions and robot-specific units before training.

Why this release is different

The deliverable is more than a schema rewrite: it pins the upstream artifact, verifies its SHA-256, preserves source actions and episode boundaries, documents every state component, checks all numeric values, rebuilds LeRobot v3 relational metadata, and proves that the official LeRobotDataset reader can load the result. That evidence travels with the data.

Source, revision, and integrity

• Source dataset: robomimic/robomimic_datasets
• Pinned source revision: 74fa018461f479cd9fd15b924a16103012096203
• Source file: v1.5/can/ph/low_dim_v15.hdf5
• Source SHA-256: 3f2eb92e0a5025d0095e866ac16cc8092d6a762abe27dec90dbaff9027282962
• Converter adapter: robomimic 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 upstream dataset card declares MIT. This conversion retains that identifier and the original citation. Conversion does not transfer ownership, create affiliation, or replace upstream terms.

Intended use

• Validating LeRobot v3 readers and HDF5 conversion pipelines
• Small offline imitation-learning and educational experiments
• Auditing original-action and episode-boundary preservation
• Comparing provenance and validation tooling across robomimic tasks

Limitations

• This is a community conversion, not an official upstream release.
• It contains the first 10 episodes, not all 200 Can PH demonstrations.
• It is a low-dimensional state-action release and is not multimodal.
• Statistical quality signals require domain-aware review before training.
• Validation does not measure downstream policy performance or task success.

Load with LeRobot

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("ViaCatalyst/robomimic-can-ph-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 data, contact ViaCatalyst support through the platform.

Citation

Please cite the original robomimic work:

@inproceedings{robomimic2021,
  title={What Matters in Learning from Offline Human Demonstrations for Robot Manipulation},
  author={Ajay Mandlekar and Danfei Xu and Josiah Wong and Soroush Nasiriany and Chen Wang and Rohun Kulkarni and Li Fei-Fei and Silvio Savarese and Yuke Zhu and Roberto Mart\'in-Mart\'in},
  booktitle={Conference on Robot Learning (CoRL)},
  year={2021}
}