README.md

---
license: mit
task_categories:
  - robotics
  - reinforcement-learning
tags:
  - LeRobot-v3
  - piper-robot
  - teleoperation
  - manipulation
  - imitation-learning
size_categories:
  - 1K<n<10K
language:
  - en
pretty_name: PiPER Robot Teaching Episodes
---

# PiPER Robot Teaching Episodes Dataset

**13 teleoperation demonstrations** for robot manipulation using a 7-DOF PiPER arm. Fully compatible with LeRobot v3 format.

## Quick Info

- **Episodes**: 13 | **Tasks**: 12 | **Size**: ~6.2 GB | **FPS**: 30
- **Robot**: PiPER 7-DOF arm | **Cameras**: Table (800×720) + Wrist | **Version**: v0.3
- **Format**: HDF5 + PNG images | **Compatible**: LeRobot v3, ACT, Diffusion Policy, SmolVLA

## Tasks

`cleaningcloth` `fillamentroll` `gamecontroller` `hexwrench` `pencil` `scissors` `scissors_hidden` `screwdriver` `smallkey` `smallpaper` `smallwoodenstick` `thinmetaldisk`

## Dataset Structure

```
{episode_name}_{timestamp}.hdf5          # Robot state, actions, compressed images
{episode_name}_{timestamp}.json          # Episode metadata (frames, fps, stats)
{episode_name}_images/
  ├── observation.images.table_cam/      # 800×720 PNG frames
  └── observation.images.wrist_cam/      # PNG frames (vertically flipped)
meta_data/
  ├── info.json                          # Dataset config, encoding, shapes
  ├── tasks.jsonl                        # Task definitions
  └── episodes.jsonl                     # Episode-task mapping
info.json                                # Root metadata (LeRobot v3)
```

### HDF5 Structure

- `observations/state`: 7-DOF joint angles (degrees)
- `observations/images/table_cam`: Compressed JPEG images
- `observations/images/wrist_cam`: Compressed JPEG images
- `actions`: 7-DOF commands
- `timestamps`: Frame timestamps

### Metadata (JSON)

Each episode JSON contains: `episode_name`, `n_frames`, `duration_seconds`, `fps`, `state_dim`, `cameras`, `state_stats` (mean/std/min/max), `recording_date`

## Usage

### LeRobot Library (Recommended)

```python
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("charithmunasinghe/piper_picking_tests", version="v0.3")

for batch in dataset:
    state = batch['observation']['observation.state']
    action = batch['action']
    table_img = batch['observation']['observation.images.table_cam']
    wrist_img = batch['observation']['observation.images.wrist_cam']
```

### Visualize

```python
from lerobot.scripts.visualize_dataset import visualize_dataset

visualize_dataset(
    repo_id="charithmunasinghe/piper_picking_tests",
    episode_index=0,
    version="v0.3"
)
```

### Direct HDF5 Access

```python
import h5py
from PIL import Image
from pathlib import Path

with h5py.File("screwdriver_20251104_203022.hdf5", 'r') as f:
    states = f['observations/state'][:]
    actions = f['actions'][:]

img = Image.open("screwdriver_images/observation.images.table_cam/frame_000000.png")
```

## Dataset Info

- **Collection**: Human teleoperation in lab environment
- **Preprocessing**: Table camera cropped (+300,0 offset), wrist camera flipped vertically
- **Split**: Single train split (13 episodes)
- **License**: MIT

## Citation

```bibtex
@dataset{piper_teaching_episodes_2025,
  title={PiPER Robot Teaching Episodes Dataset},
  author={Munasinghe, Charith and Toffetti, Giovanni},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/charithmunasinghe/piper_picking_tests}}
}
```

**Contact**: charithmunasinghe (Hugging Face)