# Krummholz Robotics - Jaco Play numeric trajectories

**PRELIMINARY • prepared 2026-10-22 • CC BY 4.0**

This cleaned research package contains **1,085 complete robot episodes, 77,965 ordered frames and 89 object-manipulation instructions** from the LeRobot conversion of CLVR Jaco Play. It covers the full numeric episode inventory of that pinned conversion. It does not include images or the full Open X-Embodiment collection.

## Intended use and limits

For preliminary offline AI experiments involving robot state/action sequences, language-conditioned tabletop manipulation, sequence modelling and data inspection. Without camera observations it cannot reproduce a vision-conditioned policy. This one-robot tabletop corpus does not establish general robotics performance. No real-robot execution or physical safety validation was performed. Coordinate units and mappings have not been independently verified, so state/action components keep generic source order; do not reinterpret them as verified motor commands.

## Files and use

- `frames.csv`: ordered numeric steps. Preserve `episode_id` boundaries and `frame_index` order.
- `tasks.csv`: 89 reviewed robot instructions; join `frames.task_id = tasks.task_id`.
- `episodes.csv`: episode lengths, task IDs and duplicate-detection fingerprints.
- `SCHEMA.json`: every field, source meaning and known constant components.
- `AUDIT.json`, `EXCLUSIONS.json`, `PROVENANCE.json`, `CHECKSUMS.sha256`: counts, decisions and reproducibility evidence.
- `build_release.py`: cleaning/validation script. Requirements are listed in its docstring.
- `LICENSE-CC-BY-4.0.txt` and `CITATION.cff`: source license and attribution.

The numeric CSV values retain the original float32 values as round-trippable decimal numbers. Four constant padding components remain to preserve source vector alignment. The all-episode LeRobot `train` designation is not an independent evaluation split; original TFDS train/test membership was not recovered. Hold out whole episodes rather than mixing frames from the same episode across evaluation and training.

## Cleaning and privacy

Retrieved only numeric data and metadata from pinned LeRobot commit `265773f4d24610e89040a9e11e1e39cbf4c8a5db`. Camera/video files were not downloaded or included. Removed global row numbers and storage, video and image-statistics bookkeeping. All 89 unique instruction strings were inspected and checked against explicit object/action vocabulary; they describe tabletop tasks and contain no names, contacts, addresses, user handles or numeric identifiers. Episode/task IDs identify robot data only. Public creator names remain in attribution documents as required by the source license.

Deduplication compares entire ordered trajectories using normalized instruction text plus timestamps, states, actions, reward and termination values. **0 duplicate episodes removed**. Repeated states within a trajectory are retained because dropping them changes timing; repeated task instructions label distinct demonstrations. No semantic near-duplicate removal is claimed. No out-of-scope task instructions or non-finite numeric trajectories were found. Source counts, per-episode bounds, min/max/count statistics, shapes and final-step flags were checked. See the audit for exact scope.

This supersedes the earlier two-episode, size-selected Jaco inspection sample for numeric research use. It is still preliminary and has not undergone independent robotics-expert validation.

## Sources, license and attribution

CLVR Jaco Play Dataset v1.0.0 (2023), by Shivin Dass, Jullian Yapeter, Jesse Zhang, Jiahui Zhang, Karl Pertsch, Stefanos Nikolaidis and Joseph J. Lim, USC/KAIST: https://github.com/clvrai/clvr_jaco_play_dataset

Numeric data supplied via the Hugging Face LeRobot conversion: https://huggingface.co/datasets/lerobot/jaco_play/tree/265773f4d24610e89040a9e11e1e39cbf4c8a5db

Jaco Play is one constituent of Open X-Embodiment: https://github.com/google-deepmind/open_x_embodiment

Data license: Creative Commons Attribution 4.0 International, https://creativecommons.org/licenses/by/4.0/ . Retain attribution, license and notices and indicate changes when redistributing. The included license contains the source warranty disclaimer. This Krummholz selection/CSV transformation is described above; source authors and LeRobot have not endorsed it.

## Rebuild

Download the four pinned files listed in `PROVENANCE.json`, naming them `source_data.parquet`, `source_episodes.parquet`, `source_tasks.parquet`, and `info.json` in a raw directory. Also retrieve the official `LICENSE` and `CITATION.cff`. Verify the listed SHA-256 values. Then run `python build_release.py RAW_DIRECTORY OUTPUT_DIRECTORY`. This uses the episode metadata for task strings because the pinned fastparquet reader does not decode the tasks-table large-string index. Every episode must supply exactly one instruction and map consistently to one source task ID; all 89 unique tasks are validated.