---
pretty_name: RealDatasets
---

# RealDatasets

This repository contains the original robot demonstrations and three converted
LeRobot v3 datasets.

## Layout

```text
raw/
├── 抓娃娃/
├── 对折毛巾/
└── 放置绳子/

lerobot/
├── pack_sloth/
├── fold_cloth/
└── insert_rope/
```

Each `lerobot/<task>/` directory is an independent LeRobot v3 dataset root.
Because the datasets are stored below repository subdirectories, download the
selected subdirectory and pass its local path as `root`.

```python
from pathlib import Path

from huggingface_hub import snapshot_download
from lerobot.datasets import LeRobotDataset

repo_root = Path(
    snapshot_download(
        repo_id="Zili2002/RealDatasets",
        repo_type="dataset",
        allow_patterns=["lerobot/pack_sloth/**"],
    )
)

dataset = LeRobotDataset(
    repo_id="Zili2002/RealDatasets",
    root=repo_root / "lerobot" / "pack_sloth",
    video_backend="pyav",
)
```

## Converted data

- Nominal timeline: 30 Hz
- Video resolution: 848x480
- `camera_2`: `observation.images.wrist`
- `camera_4`: `observation.images.front`
- Original 1920x1080 frames are center-cropped to 1908x1080 and resized
  without changing the aspect ratio.

The source recordings contain timing gaps. Converted datasets preserve every
recorded image and expand each episode onto a strict 30 Hz timeline:

- real source frame: original image and `*_valid=true`
- missing slot: black image and `*_valid=false`
- missing slot state/action: NaN with their validity masks set to false
- `source.frame_index=-1` identifies an inserted slot

Consumers must filter invalid state/action rows before normalization, model
input, and loss computation. Multiplying NaN values by a zero mask is not
sufficient.

Each converted dataset also contains:

```text
raw_video/record_NNN/
├── camera_2_*.mp4
├── camera_4_*.mp4
├── frame_timestamps.json
└── timeline_30hz.json
```

`timeline_30hz.json` records the source-to-target frame mapping and the complete
image-validity mask. LeRobot normalization statistics are computed from valid
source rows only.

## Features

- `observation.images.wrist`: video `[480, 848, 3]`
- `observation.images.front`: video `[480, 848, 3]`
- `observation.state`: float32 `[8]`
- `action`: float32 `[7]`
- `source.timestamp`: nominal 30 Hz Unix timestamp
- `source.timestamp_relative`: nominal 30 Hz episode-relative timestamp
- `source.original_timestamp`: real source timestamp, NaN for inserted rows
- `source.original_timestamp_relative`: real relative timestamp, NaN for inserted rows
- `source.frame_index`: source frame index, `-1` for inserted rows
- `observation.images.wrist_valid`
- `observation.images.front_valid`
- `observation.state_valid`
- `action_valid`