---
pretty_name: RoboSteer Preprocessing
language:
- en
tags:
- robosteer
- preprocessing
- motion-generation
- intermediate-assets
configs:
- config_name: level1_txt_gen
  data_files:
  - split: data
    path: level1/text/txt_imit/txt_gen/data.parquet
- config_name: level1_txt_comp_hands
  data_files:
  - split: data
    path: level1/text/txt_imit/txt_comp_hands/data.parquet
- config_name: level1_txt_comp_legs
  data_files:
  - split: data
    path: level1/text/txt_imit/txt_comp_legs/data.parquet
- config_name: level1_txt_fore
  data_files:
  - split: data
    path: level1/text/txt_pred/txt_fore/data.parquet
- config_name: level1_txt_retro
  data_files:
  - split: data
    path: level1/text/txt_pred/txt_retro/data.parquet
- config_name: level1_txt_inter
  data_files:
  - split: data
    path: level1/text/txt_pred/txt_inter/data.parquet
- config_name: level1_img_txt
  data_files:
  - split: data
    path: level1/text/img_txt/data.parquet
- config_name: level1_mul_bal_txt
  data_files:
  - split: data
    path: level1/text/mul_bal_txt/data.parquet
- config_name: level1_audio_motion_instructions
  data_files:
  - split: data
    path: level1/audio/motion_instructions/data.parquet
- config_name: level1_img_txt_human_index
  data_files:
  - split: data
    path: level1/image/img_txt_human/index.parquet
- config_name: level1_img_txt_skel_index
  data_files:
  - split: data
    path: level1/image/img_txt_skel/index.parquet
---

# RoboSteer Preprocessing

Reusable intermediate preprocessing assets for RoboSteer. This initial release contains Level 1 processed text instructions and image-conditioned static videos. Other levels and preprocessing stages can be added under their own directories.

## Release v1.0.0

- **213,044 text records** in nine lossless UTF-8 Parquet tables.
- **28,666 original MP4 files**: 14,333 IMG_TXT_HUMAN and 14,333 IMG_TXT_SKEL.
- Independent, uncompressed tar shards contain MP4 files and their normalized JSON metadata.
- Per-sample indexes, release inventory, SHA-256 checksums, and validation report.

These are preprocessing assets, not a declared train/validation/test split. The `data` split means the complete asset collection for a configuration. Audio motion instructions are text derived from audio; audio recordings and raw transcripts are not included. MUL_BAL is included as processed text; this release does not contain MUL_BAL videos. Prediction-video manifests are outside this release scope.

## Layout

```text
level1/
  text/txt_imit/{txt_gen,txt_comp_hands,txt_comp_legs}/data.parquet
  text/txt_pred/{txt_fore,txt_retro,txt_inter}/data.parquet
  text/{img_txt,mul_bal_txt}/data.parquet
  audio/motion_instructions/data.parquet
  image/{img_txt_human,img_txt_skel}/
    videos-00000-of-NNNNN.tar
    videos-00000-of-NNNNN.index.parquet
    index.parquet
  manifests/{assets,validation,provenance}.json
scripts/restore_text.py
scripts/build_dataset.py
SHA256SUMS
CHANGELOG.md
```

## Load text or an index

```python
from datasets import load_dataset
ds = load_dataset("PhoebeCC/RoboSteer-Preprocessing", "level1_txt_gen", split="data", revision="v1.0.0")
video_index = load_dataset("PhoebeCC/RoboSteer-Preprocessing", "level1_img_txt_human_index", split="data", revision="v1.0.0")
```

Text columns: `sample_id`, `source_sample_id`, `task`, `level`, `source_modality`, `text`, `original_relpath`, `size_bytes`, `sha256`. Original UTF-8 bytes, including line endings, can be reconstructed with `text.encode('utf-8')`. Use `scripts/restore_text.py` to restore the original relative file layout.

## Download video shards

```python
from huggingface_hub import snapshot_download
snapshot_download("PhoebeCC/RoboSteer-Preprocessing", repo_type="dataset",
    revision="v1.0.0", local_dir="robosteer-preprocessing",
    allow_patterns=["level1/image/img_txt_human/*", "level1/manifests/*", "README.md", "SHA256SUMS"])
```

Each tar is independently extractable. The index identifies its repository-relative `shard`, MP4 `member_path`, JSON `metadata_member_path`, and checksums. Video indexes are searchable tables; this release does not configure an embedded-video viewer. Extract each task into a separate folder. MP4 files are preserved byte for byte, without re-encoding.

## Provenance and alignment

`sample_id` preserves the source filename stem. `source_sample_id` removes only the known task prefix (or audio suffix). Use `(task, sample_id)` as an asset key; use `source_sample_id` to join related tasks, without assuming all tasks cover the same samples. Coverage results are recorded in the validation report.

Video metadata originates from existing sidecar JSON files. Metadata text is the original sidecar description, not necessarily the rewritten `text/img_txt` instruction. Image paths are converted to paths relative to the external public-data root. Source duration and generated duration are metadata values, not newly measured properties. The source generator specifies 25 FPS; this packaging run does not probe actual MP4 frame rates. Source images, external task JSONs, model weights, raw audio, audit logs, and organization backups are not bundled.

## Rights and citation

The repository owner has not yet supplied a release license or complete upstream attribution. No blanket open-source license is asserted for these derived assets. Upstream restrictions remain applicable; consult the owner about permitted redistribution and use. Formal benchmark citation and upstream dataset attribution will be added when supplied.

## Reproducibility

See `level1/manifests/provenance.json` for packaging rules and generator fingerprints, `assets.json` for counts, and `validation.json` for checks. `SHA256SUMS` covers all payload and documentation files except itself. SHA-256 verification can be performed with `sha256sum -c SHA256SUMS` after a full download. Release tags identify immutable intended snapshots; pin the tag or commit for experiments. Later assets should extend task-specific paths and receive a new version and changelog entry.