qualia-robotics/qualia-dataset-real

Qualia Real-Robot Datasets

Release	release-2026-07-19.3 (changelog)
Domain	Real-robot recordings
Format	LeRobot v2.1
Access	Public
Contact	Start a discussion

Public access makes the recordings inspectable and reproducible; usage rights still follow each source dataset. See License and provenance.

Dataset description

This repository collects Qualia's real-robot datasets in one place. Each dataset contains one instruction and records the source revision used to prepare it.

The collection will grow as recordings are recovered from Hugging Face and other archives. New data is added as a separate dataset when the instruction or capture setup changes; compatible recordings may be appended to an existing dataset after validation.

Using the collection

The Hub viewer opens catalog.parquet, which is the quickest way to see what is available. The trajectory files are stored in the dataset directories linked below.

This repository is not a single LeRobot or OpenPI training root. Choose one dataset directory, or prepare a separate training dataset from the recordings you need. Pin the release or commit used for each experiment.

Versioning

main contains the newest accepted state of the collection. Dated release tags identify reviewed snapshots; training manifests should record both the release tag and full commit SHA. The collection release is separate from LeRobot v2.1, which describes the on-disk data format.

Included datasets

Dataset	Instruction	Embodiment	Episodes	Recording time	Source dataset
Ring on pole base	Put the ring on the pole base without touching it, then take it out	SO-101	74	25m 21s	Johbsn/so101_ring_on_peg_v3
Put cylinder in base	put cylinder in base	TRLC DK1	160	47m 19s	mkohegyi/mini-chair-base_put

Dataset structure

data/
  <embodiment>/
    <instruction>/
      dataset.md        # datasheet for people
      dataset.yaml      # source and capture metadata
      meta/             # LeRobot v2.1 metadata
      data/             # one Parquet file per episode
      videos/           # one video per episode and camera
      qualia/           # episode provenance, statistics and validation

Each directory under data/<embodiment>/<instruction>/ is a complete, one-instruction LeRobot v2.1 dataset. The root catalog configuration exposes only catalog.parquet; it deliberately does not combine frame tables from robots with different action spaces.

Dataset creation

Curation rationale

The collection exists to make Qualia's training data easier to find, inspect and reproduce. It keeps source history and capture metadata next to the trajectories instead of relying on links scattered through chat messages.

Source data and processing

This release converts two existing LeRobot v3.0 source datasets. The source repositories were not modified. Episodes were converted to one-instruction v2.1 datasets, source episode identities were preserved, and all statistics were recalculated from the converted Parquet and video files.

Validation

Every dataset is checked for task consistency, episode and frame indices, timestamps, feature shapes, camera metadata, missing files, non-finite values and stale statistics. Video frame counts, durations and frame rates were also checked for this release.

Collection statistics

Datasets	Embodiments	Episodes	Frames	Recording time	Storage
2	2	234	130,817	1h 12m	2.4 GiB

By embodiment

Embodiment	Datasets	Episodes	Frames	Recording time	Storage
SO-101	1	74	45,636	25m 21s	453.4 MiB
TRLC DK1	1	160	85,181	47m 19s	2.0 GiB

Action and state ranges are calculated separately for each dataset because robots and control setups may use different action spaces.

More charts

• Episodes by embodiment
• Frames by embodiment
• Recording time by embodiment
• Storage by embodiment
• Camera coverage
• Split coverage
• Review status
• Outcome labels

Known limitations

• The episodes have not yet been reviewed or labelled for success and termination.
• Several legacy setup details—including hardware revision, calibration, operator and site—were not available in the source datasets and are marked unknown rather than guessed.
• The two datasets use different robots and action spaces and should not be mixed without an explicit training-time transform.

Contributing

We welcome recovered datasets, additional recordings, corrected metadata and improved review labels. See CONTRIBUTING.md for the required files and checks.

License and provenance

Qualia Robotics releases this collection under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license. Reuse requires attribution, must be non-commercial, and adaptations must use the same license.

Original source-license metadata remains recorded in each dataset.yaml for provenance. The ring source did not publish a license at the pinned revision, so its source-license field remains unknown; this records the upstream metadata and does not replace Qualia's license for this release.