Grasp Pose Generation and Visualization
Installation
pip install torch numba numpy "open3d>=0.16.0" tqdm
Part 1: Grasp Pose Generation
Grasp poses are generated based on mesh geometry and force closure analysis. The annotation pipeline follows the process from AnyGrasp SDK and supports GPU acceleration. The output includes both dense annotations and filtered sparse annotations.
We use the sparse annotation results for grasping. The output is an N x 17 array with the following structure:
[score, width, height, depth, view_angles (9), points (3), obj_ids (1)]
| Field | Dimensions | Description |
|---|---|---|
score |
1 | Grasp quality score (0.1-1.0, lower is better) |
width |
1 | Grasp width |
height |
1 | Grasp height |
depth |
1 | Grasp depth |
view_angles |
9 | Can be reshaped to a 3x3 rotation matrix |
points |
3 | 3D position of grasp center |
obj_ids |
1 | Object identifier |
The definitions of width, height, and depth follow the conventions in the GraspNet and AnyGrasp papers. The view_angles combined with points and depth form the transformation T_obj_tcp. The gripper frame definition is consistent with the GraspNet API:
Note: The generated grasp poses require coordinate transformation before use. Refer to pose_post_process_fn() in template_robot.py for details.
References:
Usage
python gen_sparse_label.py --obj_path xxx.obj --unit mm --sparse_num 3000 --max_widths 0.1
Arguments:
--obj_path: Path to the object mesh file (.obj format)--unit: Unit of the mesh, eithermm(millimeters) orm(meters)--sparse_num: Number of sparse grasp poses to filter (recommended: 3000)--max_widths: Maximum gripper width (default: 0.1)
Part 2: Grasp Pose Visualization
Visualize generated grasp poses on the object mesh.
Usage
python vis_grasp.py --obj_path xxx.obj --N 100 --unit mm
Arguments:
--obj_path: Path to the object mesh file (.obj format)-N: Number of grasp poses to visualize--unit: Unit of the mesh, eithermm(millimeters) orm(meters)