Klairy/Grasp-Anything

---
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