README.md

---
license: apache-2.0
task_categories:
  - robotics
tags:
  - robotics
  - manipulation
  - cobot-magic
  - dual-arm
  - lerobot
  - atombench
pretty_name: AtomBench CobotMagic
size_categories:
  - 1K<n<10K
---

# AtomBench CobotMagic Dataset

A dual-arm robot manipulation dataset collected on a **Cobot Magic** robot platform, designed for evaluating robotic learning policies. The dataset contains 15 tasks (8 manipulation tasks + 7 instruction-following tasks), each with 100 expert demonstration episodes.

## Dataset Summary

| Item | Value |
|------|-------|
| Robot | Cobot Magic (Dual-arm) |
| Total Tasks | 15 |
| Episodes per Task | 100 |
| Total Episodes | 1,500 |
| FPS | 30 |
| Camera Views | 3 (top, left, right) |
| Video Resolution | 640×480 |
| Video Codec | H.264 (libx264) |
| Data Format | LeRobot v2.1 |

## Tasks

### Manipulation Tasks (dm1–dm8)

| ID | Task Prompt |
|----|-------------|
| dm1 | Pick up the basket with the left hand and place the cube into the basket with the right hand. |
| dm2 | Pick up the basket with the left hand and place the ball into the basket with the right hand. |
| dm3 | Place the bottles upright on the tray. |
| dm4 | Push the two cubes into the marked area. |
| dm5 | Pick up the bowl with the left hand and pour the coffee beans from the cup into the bowl with the right hand. |
| dm6 | Open the drawer with the left hand, place the block into the drawer with the right hand, then close the drawer with the left hand. |
| dm7 | Pick up one cube with the left hand and another cube with the right hand. First place the cube in the right hand at the target position, then stack the cube in the left hand on top. |
| dm8 | Open the box lid with the left hand, place the block into the box with the right hand, then close the box lid with the left hand. |

### Instruction-Following Tasks (di1–di7)

| ID | Task Prompt |
|----|-------------|
| di1 | Pick up the basket with the left hand and place the blue cube into the basket with the right hand. |
| di2 | Pick up the basket with the left hand and place the triangular block into the basket with the right hand. |
| di3 | Pick up the basket with the left hand and place the largest block into the basket with the right hand. |
| di4 | Pick up the basket with the left hand and place the object to the left of the red cylinder into the basket with the right hand. |
| di5 | Pick up one cube with each hand and place exactly two cubes into the same basket, leaving the remaining cube outside. |
| di6 | Pick up one cube with the left hand and place it into the upper basket. Pick up one cube with the right hand and place it into the lower basket. |
| di7 | Place all geometric objects except the red cube into the basket. |

## Dataset Structure

```
agilex/
├── README.md
├── {task_name}/
│   ├── meta/
│   │   ├── info.json          # Dataset metadata
│   │   ├── episodes.jsonl     # Episode metadata
│   │   └── tasks.jsonl        # Task descriptions
│   ├── data/
│   │   └── chunk-000/
│   │       ├── episode_000000.parquet
│   │       ├── episode_000001.parquet
│   │       └── ...
│   └── videos/
│       └── chunk-000/
│           ├── image_top/
│           │   ├── episode_000000.mp4
│           │   └── ...
│           ├── image_left/
│           │   ├── episode_000000.mp4
│           │   └── ...
│           └── image_right/
│               ├── episode_000000.mp4
│               └── ...
└── ...
```

## Features

Each episode frame contains:

| Feature | Type | Shape | Description |
|---------|------|-------|-------------|
| `action` | float32 | (14,) | Leader joint positions (right arm + left arm) |
| `observation.state` | float32 | (26,) | Follower joint positions + end-effector poses (both arms) |
| `observation.images.image_top` | video | (480, 640, 3) | Top camera view |
| `observation.images.image_left` | video | (480, 640, 3) | Left camera view |
| `observation.images.image_right` | video | (480, 640, 3) | Right camera view |
| `timestamp` | float32 | (1,) | Frame timestamp |
| `frame_index` | int64 | (1,) | Frame index within episode |
| `episode_index` | int64 | (1,) | Episode index |
| `index` | int64 | (1,) | Global frame index |
| `task_index` | int64 | (1,) | Task index |

### Action Space (14-dim)

| Dim | Name |
|-----|------|
| 0–5 | Right arm joint positions (6 DoF) |
| 6 | Right gripper position |
| 7–12 | Left arm joint positions (6 DoF) |
| 13 | Left gripper position |

### Observation State Space (26-dim)

| Dim | Name |
|-----|------|
| 0–5 | Right arm joint positions (6 DoF) |
| 6 | Right gripper position |
| 7–12 | Right end-effector pose (x, y, z, rx, ry, rz) |
| 13–18 | Left arm joint positions (6 DoF) |
| 19 | Left gripper position |
| 20–25 | Left end-effector pose (x, y, z, rx, ry, rz) |

## Usage

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

dataset = LeRobotDataset(
    repo_id="AtomBench/CobotMagic",
    root="path/to/local/cobotmagic",
)
```

## Citation

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

```bibtex
@misc{atombench,
  author = {AtomBench Team},
  title = {AtomBench CobotMagic: A Dual-Arm Robot Manipulation Dataset},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/AtomBench}}
}
```

## License

This dataset is released under the [Apache 2.0 License](https://www.apache.org/licenses/LICENSE-2.0).