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).
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}
}
- 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.
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