```{eval-rst}
:orphan:
```


# GraspNet - Baseline

Robotic grasping is a fundamental challenge in robotics, requiring the generation of stable and feasible grasps for a wide variety of objects. Existing methods often struggle with generalization across diverse objects and lack large-scale datasets for training and evaluation. The GraspNet project introduced a **GraspNet-1Billion Dataset** and a baseline **Grasp Generation Model**:

- GraspNet-1Billion Dataset:

  - A massive dataset containing 1 billion grasp poses for 88,000 object models.
  - Each object is associated with multiple grasp poses, annotated with stability labels and quality scores.
  - The dataset is designed to cover a wide range of object shapes, sizes, and materials, enabling robust training and evaluation of grasp generation models.

- Grasp Generation Model:

  - A deep learning-based model that predicts 6-DoF grasp poses (position and orientation) for objects in a scene.
  - The model takes as input a point cloud representation of the scene and outputs a set of candidate grasps, ranked by their predicted stability and quality.

**Grasp Representation:**
- Grasps are represented as 6-DoF poses of a robotic gripper, defined by:

> - A 3D position (where the gripper should be placed).
> - A 3D orientation (how the gripper should be aligned).
> - Each grasp is also associated with a quality score that indicates its stability and feasibility.

![graspnet](../../developer_tools/assets/images/graspnet.png) ![Diffusion Policy]

**Model Architecture:**

- The grasp generation model consists of:

  - Point Cloud Encoder:

    - A neural network (e.g., PointNet or PointNet++) processes the input point cloud to extract features.

  - Grasp Proposal Network:

    - Generates candidate grasp poses based on the extracted features.

  - Grasp Evaluation Network:

    - Scores and ranks the candidate grasps based on their predicted stability and quality.

- The model is trained end-to-end using the GraspNet-1Billion dataset.

**More Information:**

- Full paper: <https://openaccess.thecvf.com/content_CVPR_2020/papers/Fang_GraspNet-1Billion_A_Large-Scale_Benchmark_for_General_Object_Grasping_CVPR_2020_paper.pdf>
- Homepage: <https://graspnet.net/>
- Github link: <https://github.com/graspnet/graspnet-baseline?tab=readme-ov-file>

## Model Conversion