Skip to content

Files

This branch is 780 commits ahead of, 4633 commits behind google-research/google-research:master.

dql_grasping

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
Mar 5, 2019
Jan 11, 2019
Jan 11, 2019
Jan 22, 2020
Jan 22, 2020
Jan 22, 2020
Jan 22, 2020
Jan 22, 2020
Jan 22, 2020
Jan 22, 2020
Jan 11, 2019
Feb 13, 2020
Jan 22, 2020
Jan 22, 2020
Jan 22, 2020
Feb 13, 2020
Dec 3, 2019
Jan 22, 2020
Feb 13, 2020
Jan 22, 2020
Jan 22, 2020
Jan 22, 2020
Jan 22, 2020
Jan 22, 2020
Jan 22, 2020
Jan 22, 2020
Feb 14, 2020
Jan 22, 2020
Jan 22, 2020
Feb 12, 2020
Feb 12, 2020
Jan 22, 2020

Deep Reinforcement Learning for Vision-Based Robotic Grasping: A Simulated Comparison of Off-Policy Methods

This codebase implements learning algorithms and experiments from Deep Reinforcement Learning for Vision-Based Robotic Grasping: A Simulated Comparison of Off-Policy Methods (ICRA 2018).

grasping in pybullet

If you use this codebase for your research, please cite the paper:

@article{quillen2018deep,
  title={Deep Reinforcement Learning for Vision-Based Robotic Grasping: A Simulated Comparative Evaluation of Off-Policy Methods},
  author={Quillen, Deirdre and Jang, Eric and Nachum, Ofir and Finn, Chelsea and Ibarz, Julian and Levine, Sergey},
  journal={IEEE International Conference on Robotics and Automation},
  year={2018}
}

Features

  • Several grasping environments with varying degrees of grasping difficulty.
  • Customizable DQL, MC, Supervised, Corr-MC, DDPG, PCL algorithms.
  • MC returns and elibility traces for biased returns.
  • Bash scripts for gathering data from random policies and running synchronous on-policy or off-policy experiments that alternate between training and evaluation.
  • Scripts to run grid search over hyperparameters.

Getting Started

The recommended way to set up these experiments is via a virtualenv

sudo apt-get install python-pip
python -m pip install --user virtualenv
python -m virtualenv ~/env
source ~/env/bin/activate

Then install the project dependencies in that virtualenv:

pip install -r dql_grasping/requirements.txt

The first step is then to collect off-policy grasping data with a random policy.

sh dql_grasping/run_random_collect_oss.sh

Then you can train with onpolicy re-collection. By default this runs Deep Q-Learning on the env_procedural environment.

sh dql_grasping/run_train_collect_eval_oss.sh