level2/README.md

# RoboSteer Level 2 preprocessing assets

[English](README.md) | [中文](README.zh-CN.md)

This directory provides processed text instructions and video input manifests for seven Level 2 motion-control tasks. It does not contain generated model motions or video media.

## Text instructions

`text/L2_TXT_<category>/` preserves the original UTF-8 TXT contents and filenames. Order is split into `part-000/` and `part-001/` to stay below the Hub's per-directory file limit; read TXT files recursively. Part assignments follow sorted filenames in this release and have no semantic meaning.

| Category | Instructions |
|---|---:|
| Amplitude | 2,550 |
| BodyRestrain | 3,261 |
| Direction | 5,122 |
| Order | 15,884 |
| Speed | 2,404 |
| Times | 9,738 |
| Trajectory | 1,705 |

Total: **40,664 instructions**. Four existing `duration_seconds.json` files are included for Amplitude, BodyRestrain, Direction and Speed. Order, Times and Trajectory have no duration file in this release. No missing durations were inferred or fabricated.

## Video manifests

`video/manifests/<category>/` contains 11 JSON dictionaries. Each key is a task JSON filename (the task ID plus `.json`); its value is a list of relative video paths. Amplitude, BodyRestrain, Direction and Speed each have `Video_Human/` and `Video_Skel/` manifests. Order, Times and Trajectory use `Video/`.

Example structure:

```json
{"task_id.json": ["Data_Processed/path/to/video.mp4"]}
```

Resolve each video path relative to your separately obtained public data root, which must contain `Data_RAW/` and/or `Data_Processed/`. No server-specific absolute paths or video symlinks are included. Multiple tasks may reference the same video; task counts and unique video counts are not equivalent.

Order's `Video/video_paths.json` is derived from the source task-to-result mapping by retaining only `task_json` and `original_video`. The remaining ten manifests preserve the original task-to-video mappings with normalized JSON formatting. Historical result-copy summaries and model-result mappings are excluded.

| Manifest | Task entries |
|---|---:|
| `video/manifests/Amplitude/Video_Human/video_paths.json` | 2,550 |
| `video/manifests/Amplitude/Video_Skel/video_paths.json` | 2,550 |
| `video/manifests/BodyRestrain/Video_Human/video_paths.json` | 3,261 |
| `video/manifests/BodyRestrain/Video_Skel/video_paths.json` | 3,261 |
| `video/manifests/Direction/Video_Human/video_paths.json` | 5,122 |
| `video/manifests/Direction/Video_Skel/video_paths.json` | 5,122 |
| `video/manifests/Order/Video/video_paths.json` | 15,884 |
| `video/manifests/Speed/Video_Human/speed_video_human_paths.json` | 2,404 |
| `video/manifests/Speed/Video_Skel/speed_video_skel_paths.json` | 2,404 |
| `video/manifests/Times/Video/times_video_paths.json` | 9,738 |
| `video/manifests/Trajectory/Video/video_paths.json` | 1,705 |

## Loading

```python
from pathlib import Path
import json

root = Path("level2")
for txt in sorted((root / "text").rglob("*.txt")):
    sample_id = txt.stem
    instruction = txt.read_text(encoding="utf-8")

video_root = Path("/your/public/data/root")
manifest = json.loads((root / "video/manifests/Order/Video/video_paths.json").read_text())
for task_file, relative_paths in manifest.items():
    video_paths = [video_root / path for path in relative_paths]
```

## Scope and verification

This release contains processed text only, four existing duration files, eleven video manifests, and release documentation/checksums. Raw descriptions, scripts, credentials, logs, model outputs, video media and symlinks are excluded.

`SHA256SUMS` covers every other file under this `level2/` directory. Verify from that directory with `sha256sum -c SHA256SUMS`. `release_summary.json` records category and manifest counts. Validation checks file counts, nonempty UTF-8 instructions, manifest structure and relative paths; it does not certify motion quality or availability of externally hosted videos.

Prepared on 2026-09-17 from the reorganized RoboSteer Level 2 preprocessing workspace. This packaging does not grant additional rights to source datasets or external media; consult the relevant source terms when obtaining those assets.