---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- lerobot
- robotics
- imitation-learning
- bimanual
- end-effector-control
configs:
- config_name: default
  data_files: data/*/*.parquet
---

This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).

<a class="flex" href="https://huggingface.co/spaces/lerobot/visualize_dataset?path=rainbowrobotics/box_tilting_all_v3">
<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl.svg"/>
<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl-dark.svg"/>
</a>

# RB-Y1 Box Tilting Dataset

A bimanual robot manipulation dataset collected using the Rainbow Robotics RB-Y1.

The robot rotates a box 90 degrees upward using both arms.

## Dataset Summary

- Robot: Rainbow Robotics RB-Y1
- LeRobot format: v3.0
- Episodes: 88
- Frames: 51,827
- Frequency: 30 FPS
- Duration: approximately 28 minutes and 48 seconds
- Number of tasks: 1
- Camera: Front RGB camera
- Image resolution: 640 x 480
- Task instruction: `Rotate the box 90 degrees upward`

## Action Representation

The action contains absolute end-effector poses for both robot arms.

Each arm is represented using:

    [x, y, z, wx, wy, wz]

- `x`, `y`, `z`: Cartesian end-effector position
- `wx`, `wy`, `wz`: rotation vector representation
- Right arm: 6 dimensions
- Left arm: 6 dimensions
- Total action dimension: 12

The rotation vector direction represents the rotation axis, and its magnitude
represents the rotation angle in radians.

## Observation Representation

The observation contains:

- Front RGB camera image
- Right arm joint positions
- Left arm joint positions
- Right end-effector pose
- Left end-effector pose

The robot state contains 26 dimensions:

- Right arm joints: 7
- Left arm joints: 7
- Right end-effector pose: 6
- Left end-effector pose: 6

## Dataset Structure

The dataset follows the LeRobotDataset v3.0 structure.

    meta/
    data/
    videos/

Metadata describing the dataset and its features is available in
[meta/info.json](meta/info.json).

Frame-level observations and actions are stored as Parquet files.

Camera observations are stored as MP4 video files.

## Intended Use

This dataset is intended for research involving:

- Imitation learning
- Vision-based robot manipulation
- Bimanual manipulation
- End-effector pose control
- Action chunking policies
- Vision-Language-Action models

## License

Apache License 2.0