Datasets:

ViaCatalyst
/

mimicgen-square-d0-lerobot-v3

license: cc-by-4.0
pretty_name: 1.62 GB MimicGen HDF5 → multimodal LeRobot v3
task_categories:
  - robotics
tags:
  - LeRobot
  - robotics
  - imitation-learning
  - offline-rl
  - multimodal
  - mimicgen
  - community-conversion
  - v3.0

1.62 GB MimicGen HDF5 → multimodal LeRobot v3

Before → after: HDF5 demonstrations with two RGB streams become a validated, multimodal LeRobot v3.0 subset. Convert HDF5 free →

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

This is a compact, provenance-complete conversion of the first 10 episodes from the pinned MimicGen Square D0 core HDF5 file. It provides a reproducible multimodal LeRobot v3 reference dataset and conversion-quality example.

At a glance

Property	Value
LeRobot format	`v3.0`
Robot	Panda
Task	Assemble the square nut on its matching peg
Episodes	10
Frames	1,444
Duration	72.20 seconds
FPS	20
Observation modalities	State + 2 RGB camera streams
RGB streams	`agentview_image`, `robot0_eye_in_hand_image`
Image resolution	84 × 84 × 3, H.264
Action / state dimension	7 / 59
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`	`[59]`	Deterministic concatenation of numeric `obs/*` vectors; names are documented in `meta/info.json`
`observation.images.agentview_image`	video	`[84, 84, 3]`	Aligned RGB frames from `obs/agentview_image`, encoded H.264 at 20 FPS
`observation.images.robot0_eye_in_hand_image`	video	`[84, 84, 3]`	Aligned RGB frames from `obs/robot0_eye_in_hand_image`, encoded H.264 at 20 FPS
`episode_index` / `frame_index`	`int64`	`[1]`	Preserved HDF5 episode boundaries and zero-based frame positions
`timestamp`	`float32`	`[1]`	`frame_index / 20` seconds
`task_index`	`int64`	`[1]`	Maps to the Square task in `meta/tasks.parquet`

Original-action preservation

The conversion reads the original actions array for every selected episode and writes each row directly to action. It does not replay a learned policy, regenerate actions, interpolate action values, or relabel demonstrations. The two camera streams are retained only when aligned to their episode frames; output timestamps are normalized to the declared 20 FPS.

Validation evidence

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

Dataset metadata counts
Parquet schema and frame count
Episode boundaries and timestamp regularity
Feature dimensions and finite values
LeRobot v3 relational metadata
Video-frame alignment
License and provenance completeness
Official LeRobot reader smoke test

The readiness score is 84/100 — Review recommended. Two medium-confidence statistical signals are disclosed: observation.state has an 84.76% robust-outlier frame rate and action has a 4.85% rate. These are robust distribution signals, not proof of corruption; review the feature-level evidence before training.

Source, revision, and integrity

Source dataset: amandlek/mimicgen_datasets
Pinned source revision: 33016f8a62c02334f929f2913af8fdd2a8a129e1
Source file: core/square_d0.hdf5
Source file size: 1,621,351,476 bytes
Source SHA-256: 41fc24bce0f88343099c0b1b5bf6eee08cbc35851e71276d4509a01c9b75481c
Converter adapter: mimicgen adapter 1.0.0
Converter code revision: 4230bc4b56abe85116b202398b1a625a81c1c55c

Additional audit files: provenance.json, bundle-manifest.json, bundle-manifest.external.json, and UPSTREAM_LICENSE.md.

License and attribution

The upstream MimicGen project states that its datasets are released under CC-BY 4.0. This conversion retains that identifier, attribution, and citation. Conversion does not transfer ownership, create affiliation, or replace upstream terms.

Intended use

Testing multimodal LeRobot v3 readers, video handling, and data pipelines
Small offline imitation-learning or educational experiments
Reproducible HDF5-to-LeRobot conversion evaluation
Comparing schema, provenance, and validation tooling across state and camera streams

Limitations

This is a community conversion, not an official upstream release.
It contains only the first 10 episodes from the pinned Square D0 core file, not the complete MimicGen corpus or full Square D0 split.
It retains only adapter-recognized state, action, and RGB streams; inspect meta/info.json before training.
Statistical quality signals require domain-aware review before training.
Automated validation does not measure policy performance, simulation transfer, or task success.

Load with LeRobot

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("ViaCatalyst/mimicgen-square-d0-lerobot-v3")
print(dataset.meta.total_episodes, dataset.meta.total_frames)

Conversion tooling

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

Citation

Please cite the original MimicGen work:

@inproceedings{mandlekar2023mimicgen,
  title={MimicGen: A Data Generation System for Scalable Robot Learning using Human Demonstrations},
  author={Mandlekar, Ajay and Nasiriany, Soroush and Wen, Bowen and Akinola, Iretiayo and Narang, Yashraj and Fan, Linxi and Zhu, Yuke and Fox, Dieter},
  booktitle={7th Annual Conference on Robot Learning},
  year={2023}
}