| --- |
| license: mit |
| task_categories: |
| - robotics |
| language: |
| - en |
| tags: |
| - grasping, |
| - robotics, |
| - manipulation |
| size_categories: |
| - 1M<n<10M |
| --- |
| # Grasp-Anything Dataset |
|
|
| Welcome to the Grasp-Anything dataset! |
|
|
| We are excited to release the Grasp-Anything dataset, comprising 1 million samples and featuring 3 million objects. This dataset aims to facilitate research and development in robotic grasping tasks. |
|
|
| Please fill out the [download form](https://airvlab.github.io/grasp-anything/docs/download/) and agree to our license prior to downloading the dataset. |
|
|
| The Grasp-Anything++ dataset with grasping prompts is available [here](https://huggingface.co/datasets/airvlab/Grasp-Anything-pp). |
|
|
| ## Dataset Structure: |
|
|
| The dataset is structured as follows: |
|
|
| - **image**: contains .jpg images of each scene. Each scene is identified by a SHA-256 string, for instance, `0a5bd779e492513880bef534543ff031b169a045ed7ac809c5600c3268038f4d`. The size of each image is 416x416. |
| - **scene_description**: contains .pkl scene descriptions of each scene. Each scene description is correlated to the image. |
| - **grasp_label_positive**: contains .pt positive grasp labels of samples. |
| - **grasp_label_negative**: contains .pt negative grasp labels of samples. |
| - **mask**: contains .npy masks of objects residing in the image. |
| |
| ## Usage Instructions: |
| |
| To access the dataset, follow these steps: |
| |
| 1. **Download the dataset**: Obtain the dataset files by extracting the provided zip file. |
| |
| 2. **Merge Image Files**: For the `image` directory, merge the split files using the following command: |
| |
| ```sh |
| cat image_part_aa image_part_ab > image.zip |
| |
| For more information, please visit our [webpage](https://airvlab.github.io/grasp-anything/). |
| |
| For inquiries or assistance, please contact the authors at [this link](https://andvg3.github.io/). |
| |
| Best regards, |
| |
| Authors |
