NeuralGrasps Dataset

This repository hosts the code and utilities associated with pipeline for generating the training dataset of grasps, point clouds, contact maps and SDF values used in the paper NeuralGrasps: Learning Implicit Representations for Grasps of Multiple Robotic Hands.

It includes things like:

graspit interface python code
processing of the grasp information (xml files) to 3d models
rendering of the 3d models and obtaining the point clouds from a given camera viewpoint.
saving all results to an output directory.

Graspit grasp generation is done on a python-2 environment on a Ubunut 18.04 docker image. All the other steps make use of a python-3 env (see rendering-test/environment.yml)

Grasp Generation

Initial Setup (Docker container) for GraspIt Grasp Generation

I am sharing the container image ([CONTAINER-NAME].tar.gz -- need to extract it first) on ut-dallas box. Example container name = img-grasp-pyb.tar.gz

Extract the img-grasp-pyb.tar.gz file -- this will give you a .tar archive. You can load this as docker image using: sudo docker load --input img-grasp-pyb.tar. The docker image will now be loaded as img-grasp-pyb

To generate grasps using Graspit, we require support for GUI mode:

Run xhost + (required)

Start the docker container as follows (supports GUI mode for graspit):

sudo nvidia-docker run -it --name CONTAINER_NAME --privileged --gpus all \
-e NVIDIA_DRIVER_CAPABILITIES=all -e DISPLAY=$DISPLAY -v /tmp/.X11-unix:/tmp/.X11-unix \
--device=/dev/dri:/dev/dri -v PATH_TO_DOCKER_DATA_FOLDER:/root/data/ IMAGE_NAME`

PATH_TO_DOCKER_DATA_FOLDER is where you have extracted the docker_data folder or it can simply be a symlink to an appropriate folder which will be mounted on the docker container at the /root/data/ location.
IMAGE_NAME: docker image name (for us it's img-grasp-pyb)
CONTAINER_NAME : a descriptive name for the container
This will open an interactive bash session. tmux is already installed, so you can open multiple terminals. Note that the tmux "prefix" is Ctrl-b but I have changed it to Ctrl-a. So for example, if you need to create new window in a session, it is: Ctrl-a c. Using tmux is highly recommended as we need to have multiple terminals!
Additional notes (for future reference):
- Graspit Directory: /opt/graspit/
- The ycb model xml files are already added the graspit directory.
- For ycb models to graspit, see prepare-ycb-graspit/README.md

Grasp Generation Script

Python 2 env -- default one on the container (using this one). Make sure that you have mounted the docker_data dir on host (or any other host dir of your choice) as /root/data/ on the container. You can check this from inside the docker containter: ls ~/data/ -- should list the files from the host folder.

Run roscore in terminal
In a separate terminal, run:
- source ~/graspit_ros_ws/devel/setup.bash
- Inside the same terminal where you sourced graspit_ros_ws, run generate_grasps.py with appropriate flags.

This will output the initial set of grasps to ~/data/output_dataset/initial_grasps/ folder in the docker container. Note that since ~/data/ is actually a host mounted dir, you can also see the changes via file explorer on the host. Once the grasps are generated, we do not need the docker container. Also note that the output_dataset is referred to as the dataset directory in this readme.

Example:

python generate_grasps.py -e \
--path_out /root/data/output_dataset/initial_grasps/ \
--n_jobs 16 --max_steps 100000 \
-l ycb_models_file.txt --num_grasps 300 \
--gripper fetch_gripper`

-e makes the process headless (non-gui) making it much faster to run in practice.
-l option specifies the file which shows ycb models will be used.
Options for --gripper (case sensitive):
- fetch_gripper
- Barrett
- HumanHand
- panda_grasp
- Allegro
We might need to tune the --num_grasps and --gripper flags depending on how the process goes. Can remove -e flag to debug the vizualization.

Grasp Refinement

Run refinement code refine_grasps.py which will do farthest point sampling and output a fixed number of grasps (here 50) to ~/data/output_dataset/refined_grasps as refined-*.json files. Note: we do not modify the grasps (or refine them)!

The command for refine_grasps.py will be:

python refine_grasps.py \
--path_out ~/data/output_dataset/ \
--num_samples 50 \
--gripper fetch_gripper

Visualizing the Refined Grasps

If you have followed the docker commands, then GUI mode of applications is supported.
From a tmux terminal:
- Activate the pybullet env: conda activate pybullet
- Launch jupyter notebook: jupyter notebook --allow-root
- It should
Check out the rendering_test/test_grasp_viz-gripper.ipynb notebook:
- The pybullet connnect GUI command should launch the pybullet gui window. IF it does not launch, you can also do this viz outside the docker env by simply making a new pybullet on your host env via the rendering_test/environment.yml file.
  - Make sure to change the dataset_dir (since now you are accessing from the host).
- It assumes that the refined grasps are stored in output_dataset/refined_grasps/
- Change the object_model_name as needed
- Change the gripper_name as needed

Dataset Generation

Preparing YCB models for Pybullet

Use the download script in ycb-tools repo to download the dataset. We are using the google-16k scans. Download (clone) the ycb-tools in some appropriate location like ~/external_datasets/ycb_dataset/

Objects will be downloaded to ycb-tools/models/ycb/

Then check out the prepare-ycb-pybullet folder. Steps copied below:

Note: Use these generated urdf files to do the refined grasp vizualization

Generating the URDFs

Need to do this from outside the docker environment since the ycb data is not present inside the docker-data folder.

Simply go through the script ycb_prepare_urdf.py with appropriate command line flags:

You will only need to pass the -f (--ycb_dir) flag to point to the location of ycb models. For example it can then be: python ycb_prepare_urdf.py -f /mnt/Data/ycb-objects/ycb-tools/models/ycb/

The /mnt/Data/ycb-objects/ycb-tools/models/ycb/ contains the ycb object folders like 001_chips_can/, 003_cracker_box/ etc..

Default Params

-m (--models_file) is already the one that is default specified
-o can be anywhere on your system so long as you remember it for the point cloud generation process. Or to make matters easy, copy the generate urdfs to ../rendering-test/ycb-urdfs/
-t is the mesh type. No need to change it.

Point Cloud Rendering, SDF and Contact Maps

For each of the steps below, you can do a --help with the python script to know more about the command line arguments required. Most of the defaults are chosen carefully so they should work as is and they are straight forward to understand.

Point Cloud Generation via pybullet: See rendering-test/grasp_pc.py for generating point clouds of each grasp.

Ground Truth SDF Sampling from point clouds: See rendering-test/gen_sdf.py for generating the sdf samples (training data)

Contact Maps: See rendering-test/gen_obj_contact_map.py for generating the contact map of gripper-object grasping scene (across all the grasps for a gripper-object pair).

Visualizing the Grasps as RGB Images: There are two files to generate rgb images of each grasping scene.

rendering-test/gen_images.py
rendering-test/gen_high_res_images.py

You can tweak the params of either to get high or low-res images in order to quickly visualize how the grasps look like. Example run:

python gen_high_res_images.py \
-l ../grasp-generation/ycb_models_file_bkp.txt \
-o ../../../dataset_train \
--gripper Barrett

Assumes that the dataset dir (e.g. dataset_train) is present one level above the repository root (OR at least its symlink is present).

Dataset Folder Structure

Here object_003 is an example name. Each object will have similar sub-folders as shown. Each of the above scripts will populate the sub-folders on execution. The object_point_cloud.ply can be copied from the corresponding ycb directory.

|---initial_grasps/*.json
|---refined_grasps/*.json
|---object_003/
    |---point_cloud/
        |---Allegro/
        |---Barrett/
        |---HumanHand/
        |---fetch_gripper/
        |---panda_grasp/
    |---sdf/
    |---contactmap/
    |---images/
    |---object_point_cloud.ply
    |---norm_params.npz

Citation

Please consider citing the paper if it helps in your work.

@misc{https://doi.org/10.48550/arxiv.2207.02959,
      doi = {10.48550/ARXIV.2207.02959},
      url = {https://arxiv.org/abs/2207.02959},
      author = {Khargonkar, Ninad and Song, Neil and Xu, Zesheng and Prabhakaran, Balakrishnan and Xiang, Yu},
      title = {NeuralGrasps: Learning Implicit Representations for Grasps of Multiple Robotic Hands},       
      publisher = {arXiv},
      year = {2022},          
      copyright = {Creative Commons Attribution 4.0 International}
      }

Acknowledgements

Thanks to the authors of the following repositories for open-sourcing their code and from which major parts of our codebase were inspired from:

junjungoal/neuralgrasps-dataset-generation