vcpd: A Python repository from caijunhao

Volumetric-based Contact Point Detection for 7-DoF Grasping

This repository contains the implementation of the work "Volumetric-based Contact Point Detection for 7-DoF Grasping", including data generation, network training, performance validation on simulator, and ros scripts to perform real-robot grasping.

Overview of the grasp pipeline.

Paper, Video

Data Generation

Step 1: generate URDF files from .obj files.

In order to load meshes into both Pybullet and Isaac Gym, we first use mesh_processing.py to convert the .obj mesh files into the URDF format.

python scripts/data_collection/mesh_processing.py \ 
--mesh_path $PATH_TO_OBJ_FILES \ 
--mesh_type $MESH_TYPE \ 
--w $GRIPPER_WIDTH \ 
--output $OUTPUT_PATH_TO_URDF

Each URDF folder contains four files. One example is listed below:

-cube#0.010#0.010#0.080  // the folder including obj and URDF files
    -cube#0.010#0.010#0.080.obj  // the mesh file that is vertex-densed, used for grasp analysis
    -cube#0.010#0.010#0.080_col.obj  // the mesh used for collision checking
    -cube#0.010#0.010#0.080_vis.obj  // the mesh used for visualization
    -cube#0.010#0.010#0.080.urdf  // the URDF description file

Some meshes and the corresponding URDFs can be found at here.

Step 2: antipodal analysis on single mesh.

With the mesh set in URDF format, we generate contact pairs with grasp labels for each mesh by running

python scripts/data_collection/grasp_analysis.py \ 
--mesh_path $PATH_TO_URDF_FOLDER \ 
--config config/config.json \ 
--output ${PATH_TO_GRASP_LABEL}/${MESH_TYPE}_grasp_info 
--gui 0

For example, the output files of cube#0.010#0.010#0.080 include

-primitive_grasp_info
    -cube#0.010#0.010#0.080_antipodal_mean.npy
    -cube#0.010#0.010#0.080_antipodal_min.npy
    -cube#0.010#0.010#0.080_antipodal_raw.npy
    -cube#0.010#0.010#0.080_centers.npy
    -cube#0.010#0.010#0.080_collisions.npy
    -cube#0.010#0.010#0.080_directions.npy
    -cube#0.010#0.010#0.080_info.json
    -cube#0.010#0.010#0.080_intersected_face_ids.npy
    -cube#0.010#0.010#0.080_intersects.npy
    -cube#0.010#0.010#0.080_quaternions.npy
    -cube#0.010#0.010#0.080_vertex_ids.npy
    -cube#0.010#0.010#0.080_widths.npy

Step 3: random scene construction

Given sets of the mesh, contact pair, and the grasp quality, we can build stacked scenes with labelled contact points

python scripts/data_collection/scene_construction.py \
--config config/config.json \ 
--mesh $PATH_TO_URDF_FOLDER \ 
--info ${PATH_TO_GRASP_LABEL}/${MESH_TYPE}_grasp_info \ 
--output ${$PATH_TO_SCENE_OUTPUT}/${MESH_TYPE}/${NUM_OBJ}_objs \

${NUM_OBJ} is determined according to the number of objects stacked in the tray, which is configurated by scene/obj tag in config.json.

Network Training

Step 1: put scenes together

Since the scenes generated with different number of objects are placed in different folders, we need to put them together to simplify the pre-processing. For example, the folder structure of scenes of primitive-shaped objects is

-train  // the folder containing the soft links to the scene files
-scene
    -primitives
        -5objs
            -000000
            -000001
            ......
        -10objs
            ......
        -15objs
            ......
        -20objs
            ......

Then we can use the following command to generate soft links of the scene files in the train folder

for f0 in ../scene/primitives/5objs/*; do f1=$(echo $f0 | cut -d '/' -f 4); f2=$(echo $f0 | cut -d '/' -f 5); ln -s $f0 primitive\-$f1\-$f2; done

The output links are

-train
    -primitive-5objs-000000
    -primitive-5objs-000001
    ......
    -primitive-10objs-000000
    ......

Step 2: train the network

python scripts/cpn/train_cpn.py \ 
--config config/config.json \ 
--log $PATH_TO_MODEL \ 
--train_dir $PATH_TO_TRAIN \ 
--test_dir $PATH_TO_TEST

Performance Evaluation

Evaluate antipodal score and collision-free rate

python scripts/cpn/test_cpn.py \ 
--config config/config.json \ 
--mesh $PATH_TO_URDF_FOLDER \ 
--model_path $PATH_TO_MODEL/cpn_xxx.pth

Data

Scene data

Setup for real-robot grasping

TODO

License

The codebase and dataset are under CC BY-NC-SA 3.0 license. You may only use the code and data for academic purposes.

Citation

If you find our work useful, please consider citing.

@inproceedings{cai2022volumetric,
    title     = {Volumetric-based Contact Point Detection for 7-DoF Grasping},
    author    = {Cai, Junhao and Su, Jingcheng and Zhou, Zida and Cheng, Hui and Chen, Qifeng and Wang, Michael Yu},
    booktitle={Conference on Robot Learning (CoRL)},
    year={2022},
    organization={PMLR}
}

Acknowledgement

The implementation of SDF module is inspired by Andy Zeng's tsdf-fusion-python and Jingwen Wang's KinectFusion.

The real-robot experiments are based on the franka_ros_interface.

caijunhao/vcpd