# Multi-GraspLLM: A Multimodal LLM for Multi-Hand Semantic-Guided Grasp Generation

**[Project Page](https://multi-graspllm.github.io/)** | **[arXiv](https://arxiv.org/abs/2412.08468)**

---
## Updates
- **2025.3**: We add the Jaco hand data which is our final version.
- **2025.1**: We add the Barrett hand data.
- **2024.12**: Released the Multi-GraspSet dataset with meshes of objects.
- **2024.12**: Released the Multi-GraspSet dataset with contact annotations.

---
## Overview

We introduce **Multi-GraspSet**, the first large-scale multi-hand grasp dataset enriched with automatic contact annotations.  


<table style="margin: auto; text-align: center;">
    <tr>
        <td>
            <img src="./assets/picture/dataset_construction_new.png" alt="Structure of Multi-GraspLLM" width="1000">
        </td>
    </tr>
    <tr>
        <td>
            <p>The Construction process of Multi-GraspSet</p>
        </td>
    </tr>
</table>


<table style="margin: auto; text-align: center;">
    <tr>
        <td>
            <img src="./assets/picture/vis_dataset_grasp_output_01_new.png" alt="Structure of Multi-GraspLLM" width="600">
        </td>
    </tr>
    <tr>
        <td>
            <p>Visualization of Multi-GraspSet with contact annotations</p>
        </td>
    </tr>
</table>


---




## Installation

Follow these steps to set up the evaluation environment:

1. **Create the Environment**
    ```bash
    conda create --name eval python=3.9
    ```

2. **Install PyTorch and Dependencies**
    ```bash
    pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2
    ```
    > ⚠️ Ensure the CUDA toolkit version matches your installed PyTorch version.

3. **Install Pytorch Kinematics**
    ```bash
    cd ./pytorch_kinematics
    pip install -e .
    ```

4. **Install Remaining Requirements**
    ```bash
    pip install -r requirements_eval.txt
    ```

---

## Visualize the Dataset



1. **Run the Visualization Code** 

    Open and execute the `vis_mid_dataset.ipynb` file to visualize the dataset.