README.md

---
license: apache-2.0
task_categories:
- robotics
tags:
- robotics
- manipulation
- bimanual
- dexterous-hand
- imitation-learning
- lerobot
size_categories:
- 100K<n<1M
---

# PetalDex — Bimanual Dexterous "Arrange Flowers" Dataset

**Project website:** [https://petaldex.github.io/](https://petaldex.github.io/)

PetalDex is a bimanual dexterous manipulation dataset for the **"arrange flowers"** task,
collected on a `wuji_bimanual` robot (two 7‑DoF arms + two 20‑DoF dexterous hands).
Each frame provides two RGB camera views (head + right wrist), a 54‑D proprioceptive
state, and a 54‑D action, recorded at **30 fps**.

The dataset is released in **two formats** (identical content), so you can use whichever
fits your pipeline:

| Format | Path | Images | Loader |
|--------|------|--------|--------|
| **LeRobot v3.0** | `lerobot/<subset>/` | AV1 video (`.mp4`) | `lerobot` / parquet |
| **HDF5** | `hdf5/<subset>/` | per‑frame JPEG in `.h5` | `h5py` + `cv2` |

## Subsets

| Subset | Episodes | Frames | Description |
|--------|---------:|-------:|-------------|
| `robot_auto` | 525 | 956,892 | Robot‑collected "arrange flowers" episodes (merged). |
| `robot_human_co-creation` | 200 | 268,322 | Human–robot co‑creation episodes. |

## Data schema

- `observation.state` — `float32[54]` = `left_arm(7) + right_arm(7) + left_hand(20) + right_hand(20)`
- `action` — `float32[54]` (same layout as state)
- `observation.images.head` — `224×224×3` RGB
- `observation.images.right_wrist` — `224×224×3` RGB
- `fps` — 30 · `robot_type` — `wuji_bimanual` · `task` — `"arrange flowers"`

## Repository layout

```
PetalDex/
├── lerobot/
│   ├── robot_auto/                     # LeRobot v3.0 dataset (data/ + videos/ + meta/)
│   └── robot_human_co-creation/
└── hdf5/
    ├── robot_auto/                     # episode_000000.h5 ... + dataset_meta.json
    └── robot_human_co-creation/
```

> Note: this repo hosts **four** sub‑datasets, so `LeRobotDataset("jasonGUself/PetalDex")`
> at the root will not work — load a specific subset folder instead (see below).

## Usage

### LeRobot v3.0

Download a subset and point `LeRobotDataset` at its local root:

```python
from huggingface_hub import snapshot_download
from lerobot.datasets.lerobot_dataset import LeRobotDataset

local = snapshot_download(
    repo_id="jasonGUself/PetalDex", repo_type="dataset",
    allow_patterns="lerobot/robot_auto/*",
)
ds = LeRobotDataset("jasonGUself/PetalDex", root=f"{local}/lerobot/robot_auto")
print(ds[0].keys())
```

### HDF5

Each episode is one `.h5` file. Images are stored as per‑frame JPEG bytes
(variable‑length `uint8`), decode with OpenCV (returns BGR by cv2 convention):

```python
import h5py, cv2, numpy as np

with h5py.File("hdf5/robot_auto/episode_000000.h5", "r") as f:
    T      = int(f.attrs["num_frames"])            # attrs: fps, task, robot_type, ...
    state  = f["observations/state"][:]            # (T, 54) float32
    action = f["action"][:]                        # (T, 54) float32
    head   = cv2.imdecode(f["observations/images/head"][0],       cv2.IMREAD_COLOR)  # (224,224,3)
    wrist  = cv2.imdecode(f["observations/images/right_wrist"][0], cv2.IMREAD_COLOR)
```

HDF5 layout per file:

```
attrs: robot_type, fps, task, episode_index, num_frames,
       image_encoding="jpeg", image_shape=[224,224,3],
       state_dim=54, action_dim=54, state_layout
/observations/images/head          vlen uint8 (T,)   # JPEG bytes per frame
/observations/images/right_wrist   vlen uint8 (T,)
/observations/state                float32 (T, 54)
/action                            float32 (T, 54)
/timestamp                         float32 (T,)
/frame_index                       int64   (T,)
```

## Notes

- HDF5 images are re‑encoded to JPEG (quality 95) from the source AV1 video — visually
  lossless but not bit‑identical to the LeRobot video frames.
- The two camera streams are packed into different numbers of video files in the LeRobot
  format; frame↔episode alignment is handled by `meta/episodes/*.parquet`.