README.md

---
license: mit
language:
- en
tags:
- vla
- robot-control
- understanding
---

# Demos25 - Dual-Arm Robot Manipulation Dataset

## Dataset Description

Demos25 is a comprehensive dual-arm robot manipulation dataset containing 25 episodes of supermarket packing tasks. The dataset follows the LeRobot format and includes multi-modal data with video observations, robot states, and actions.

### Dataset Summary

- **Total Episodes**: 25
- **Total Frames**: 44,492
- **FPS**: 30
- **Robot Type**: A2D (dual-arm robot)
- **Task**: Supermarket packing with item manipulation and bag organization

### Task Description

The robot is positioned in front of a cash register with four different types of items and a soft woven bag. The task involves altering the positions and orientations of the items and the soft woven bag on the cash register.

## Dataset Structure

The dataset is organized in LeRobot v2.1 format with the following structure:

```
demos25/
├── data/
│   └── chunk-000/
│       ├── episode_000000.parquet
│       ├── episode_000001.parquet
│       └── ...
├── videos/
│   └── chunk-000/
│       ├── observation.images.head/
│       ├── observation.images.hand_left/
│       └── observation.images.hand_right/
├── meta/
│   ├── info.json
│   ├── episodes.jsonl
│   ├── episodes_stats.jsonl
│   └── tasks.jsonl
└── interleaved_demo.jsonl
```

## Features

### Video Observations
- **observation.images.head**: Head camera view (480x640x3, 30fps)
- **observation.images.hand_left**: Left hand camera view (480x640x3, 30fps)  
- **observation.images.hand_right**: Right hand camera view (480x640x3, 30fps)

### Robot States
- **observation.states.joint.position**: 14 joint positions (7 per arm)
- **observation.states.joint.current_value**: 14 joint current values
- **observation.states.effector.position**: 2 gripper positions
- **observation.states.end.position**: 2 end-effector positions (3D)
- **observation.states.end.orientation**: 2 end-effector orientations (quaternion)
- **observation.states.head.position**: 2 head positions (yaw, pitch)
- **observation.states.robot.position**: 3 robot base positions
- **observation.states.robot.orientation**: 4 robot base orientations
- **observation.states.waist.position**: 2 waist positions (pitch, lift)

### Actions
- **actions.joint.position**: 14 joint position targets
- **actions.effector.position**: 2 gripper position targets
- **actions.end.position**: 2 end-effector position targets (3D)
- **actions.end.orientation**: 2 end-effector orientation targets (quaternion)
- **actions.head.position**: 2 head position targets
- **actions.robot.velocity**: 2 robot base velocities
- **actions.waist.position**: 2 waist position targets

## Usage

### Loading with LeRobot

```python
from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
from eo.data.lerobot_dataset import LeRobotDataset

# Load metadata
meta = LeRobotDatasetMetadata(
    repo_id="IPEC-COMMUNITY/demos25",
    root="./demos25",
)

# Load dataset
dataset = LeRobotDataset(
    repo_id="IPEC-COMMUNITY/demos25",
    root="./demos25",
    delta_timestamps={
        k: [i / meta.fps for i in range(0, 50)]
        for k in ["actions.joint.position", "actions.effector.position"]
    }
)
```

### Loading with EO-1 Framework

```python
from eo.data.lerobot_dataset import LeRobotDataset
from eo.data.schema import LerobotConfig

dataset = LeRobotDataset(
    repo_id="IPEC-COMMUNITY/demos25",
    root="./demos25",
    select_video_keys=[
        "observation.images.head", 
        "observation.images.hand_left", 
        "observation.images.hand_right"
    ],
    select_state_keys=[
        "observation.states.joint.position", 
        "observation.states.effector.position"
    ],
    select_action_keys=[
        "actions.joint.position", 
        "actions.effector.position"
    ],
    delta_timestamps={
        k: [i / 30 for i in range(0, 1500)] 
        for k in ["actions.joint.position", "actions.effector.position"]
    }
)
```

## Citation

If you use this dataset in your research, please cite:

```bibtex
@dataset{demos25_2024,
  title={Demos25: Dual-Arm Robot Manipulation Dataset for Supermarket Packing},
  author={IPEC-COMMUNITY},
  year={2024},
  url={https://huggingface.co/datasets/IPEC-COMMUNITY/demos25}
}
```

## License

This dataset is released under the MIT License.

## Contact

For questions or issues, please contact the IPEC-COMMUNITY team.