Xiang-Zhang-98/Hybrid-action-for-insertion

Python

This repo suppports our ICRA 22 submission "Learning Insertion Primitives with Discrete-Continuous Hybrid Action Space for Robotic Assembly tasks" This repo is developed based on the code from MP-DQN. paper code

Installation

Setup MuJoCo

Download the MuJoCo version 2.0 binaries
- Download Mujoco200 and activation key from https://www.roboti.us/download.html https://www.roboti.us/license.html
Extract the downloaded mujoco200 directory into ~/.mujoco/mujoco200
Append the following lines to ~/.bashrc and source ~/.bashrc

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/robodemo/.mujoco/mujoco200/bin
export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libGLEW.so

Create conda environment

conda create --name insertion python=3.7 pip
conda activate insertion
pip install -r requirements.txt
cd pegibhole_env
pip install -e .
cd ..

Training

python TS-MP-DQN/run_peginhole_3prms.py # train peg-in-hole policy
python TS-MP-DQN/run_policy_3prms.py --policy_file /proposed_square_peg1/19999 # run policy
python TS-MP-DQN/run_transferlearning_3prms.py --episodes 5000 --loadingpath "/proposed_square_peg1/19999" --pegshape "pentagon" --fintune_start_episode 1000 # transfer learning

Citing

If this repository has helped your research, please cite the following:

@inproceedings{zhang2022learning,
  title={Learning insertion primitives with discrete-continuous hybrid action space for robotic assembly tasks},
  author={Zhang, Xiang and Jin, Shiyu and Wang, Changhao and Zhu, Xinghao and Tomizuka, Masayoshi},
  booktitle={2022 International Conference on Robotics and Automation (ICRA)},
  pages={9881--9887},
  year={2022},
  organization={IEEE}
}