Dynamic-Resolution Model Learning for Object Pile Manipulation

Website | Paper

hello-world.mp4

Dynamic-Resolution Model Learning for Object Pile Manipulation
Yixuan Wang*, Yunzhu Li*, Katherine Driggs-Campbell, Li Fei-Fei, Jiajun Wu
Robotics: Science and Systems, 2023.

Citation

If you use this code for your research, please cite:

@INPROCEEDINGS{Wang-RSS-23, 
    AUTHOR    = {Yixuan Wang AND Yunzhu Li AND Katherine Driggs-Campbell AND Li Fei-Fei AND Jiajun Wu}, 
    TITLE     = {{Dynamic-Resolution Model Learning for Object Pile Manipulation}}, 
    BOOKTITLE = {Proceedings of Robotics: Science and Systems}, 
    YEAR      = {2023}, 
    ADDRESS   = {Daegu, Republic of Korea}, 
    MONTH     = {July}, 
    DOI       = {10.15607/RSS.2023.XIX.047} 
} 
@inproceedings{li2018learning,
    Title={Learning Particle Dynamics for Manipulating Rigid Bodies, Deformable Objects, and Fluids},
    Author={Li, Yunzhu and Wu, Jiajun and Tedrake, Russ and Tenenbaum, Joshua B and Torralba, Antonio},
    Booktitle = {ICLR},
    Year = {2019}
}

Installation

Prerequisite

• Install docker-ce
• Install nvidia-docker
• Install Anaconda or Miniconda

Create conda environment

conda env create -f env.yaml && conda activate dyn-res-pile-manip

Install PyFleX

Run bash scripts/install_pyflex.sh. You may need to source ~/.bashrc to import PyFleX.

What does this script do?

We built our simulation using PyFleX. The original repository is here. We modified it with additional features, such as depth rendering and headless rendering. We put the modified PyFleX in our repo. Please follow the commands below to install it.

docker pull xingyu/softgym
docker run \
    -v {PATH_TO_REPO}/PyFleX:/workspace/PyFleX \
    -v {PATH_TO_CONDA_ENV}:/workspace/anaconda \
    -v /tmp/.X11-unix:/tmp/.X11-unix \
    --gpus all \
    -e DISPLAY=$DISPLAY \
    -e QT_X11_NO_MITSHM=1 \
    -it xingyu/softgym:latest bash

For example, in my local machine, the command is

docker run \
    -v /home/yixuan/dyn-res-pile-manip/PyFleX:/workspace/PyFleX \
    -v /home/yixuan/miniconda3/envs/rss-release/:/workspace/anaconda \
    -v /tmp/.X11-unix:/tmp/.X11-unix \
    --gpus all \
    -e DISPLAY=$DISPLAY \
    -e QT_X11_NO_MITSHM=1 \
    -it xingyu/softgym:latest bash

After entering docker environment, run

export PATH="/workspace/anaconda/bin:$PATH"
cd /workspace/PyFleX
export PYFLEXROOT=${PWD}
export PYTHONPATH=${PYFLEXROOT}/bindings/build:$PYTHONPATH
export LD_LIBRARY_PATH=${PYFLEXROOT}/external/SDL2-2.0.4/lib/x64:$LD_LIBRARY_PATH
cd bindings; mkdir build; cd build; /usr/bin/cmake ..; make -j

Add the following to ~/.bashrc

export PYFLEXROOT=/home/yixuan/dyn-res-pile-manip/PyFleX # replace with your own path
export PYTHONPATH=${PYFLEXROOT}/bindings/build:$PYTHONPATH
export LD_LIBRARY_PATH=${PYFLEXROOT}/external/SDL2-2.0.4/lib/x64:$LD_LIBRARY_PATH

Load custom pybullet_data

Run bash scripts/custom_robot_model.sh

What does this script do?

Since we use custom end-effector for our robot, please add our custom kinova and franka_panda to pybullet_data folder.

Download data and model

If you want to visualize awesome object pile manipulation: download model by running bash scripts/download_model.sh
If you want to train my own dynamics model and resolution regressor: download data by running bash scripts/download_data.sh
All dataset and model will be stored under data folder.

More about data downloading

It will download three datasets. You could choose to download only partial of them according to your needs.

• data/res_rgr_data: data for training resolution regressor
• data/res_rgr_data_small: data for training resolution regressor, but with only 30 data points. It is mainly to sanity check the code
• data/gnn_dyn_data: data for training dynamics

Task in sim

Run python visualize_mpc.py. You will see our robot system push the spreaded object pile into an I-shape pile.

Change initial configuration and task

Initial configurations and task are specified in config/mpc/config.yaml. init_pos provides some options of object pile initial states. Task specification can be changed in task.

Train GNN dynamics model

Run python -m train.train_gnn_dyn. It will save the trained model in data/gnn_dyn_model.

Train resolution regressor

Run python -m train.train_res_rgr. It will save the trained model in data/res_rgr_model.

Data generation

Generate data for dynamics model

Run python -m data_gen.gnn_dyn_data. It will save the generated data in data/gnn_dyn_data_custom.

Generate data for resolution regressor

Run python -m data_gen.res_rgr_data. It will save the generated data in data/res_rgr_data_custom.

Different data generation modes

You could specify the data generation mode in config/data_gen/res_rgr.yaml by changing mpc_data/mode. If you generate resolution regressor training data in random mode, it will synthesize random initial configurations and goals. If you want to re-produce Fig. 4 (a) in the paper, you could change it to same_init mode. For Fig. 4 (b) in the paper, you could change it to same_goal mode. Due to stochasity, it may not produce exactly the same result as in the paper.