This is the official codebase for Manipulation Primitive-augmented reinforcement Learning (MAPLE), from the following paper:
Augmenting Reinforcement Learning with Behavior Primitives for Diverse Manipulation Tasks
Soroush Nasiriany, Huihan Liu, Yuke Zhu
UT Austin Robot Perception and Learning Lab
IEEE International Conference on Robotics and Automation (ICRA), 2022
[Paper] [Project Website]
This guide contains information about (1) Installation, (2) Running Experiments, (3) Setting Up Your Own Environments, (4) Acknowledgement, and (5) Citation.
- Current codebase:
git clone https://github.com/UT-Austin-RPL/maple - (for environments) the
maplebranch in robosuite:git clone -b maple https://github.com/ARISE-Initiative/robosuite
- Download MuJoCo 2.0 (Linux and Mac OS X) and unzip its contents into
~/.mujoco/mujoco200, and copy your MuJoCo license key~/.mujoco/mjkey.txt. You can obtain a license key from here. - (linux) Setup additional dependencies:
sudo apt install libgl1-mesa-dev libgl1-mesa-glx libglew-dev libosmesa6-dev software-properties-common net-tools xpra xserver-xorg-dev libglfw3-dev patchelf - Add MuJoCo to library paths:
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$HOME/.mujoco/mujoco200/bin
- Create the conda environment:
conda env create --name maple --file=maple.yml - (if above fails) edit
maple.ymlto modify dependencies and then resume setup:conda env update --name maple --file=maple.yml - Activate the conda environment:
conda activate maple - Finish maple setup: (in your maple repo path do)
pip install -e . - Finish robosuite setup: (in your robosuite repo path do)
pip install -e .
Scripts for training policies and re-playing policy checkpoints are located in scripts/train.py and scripts/eval.py, respectively.
These experiment scripts use the following structure:
base_variant = dict(
# default hyperparam settings for all envs
)
env_params = {
'<env1>' : {
# add/override default hyperparam settings for specific env
# each setting is specified as a dictionary address (key),
# followed by list of possible options (value).
# Example in following line:
# 'env_variant.controller_type': ['OSC_POSITION'],
},
'<env2>' : {
...
},
}
See parser in scripts/train.py for a complete list of options. Some notable options:
env: the env to run (eg.stack)label: name for experimentdebug: run with lite options for debugging
During training, the results will be saved to a file called under LOCAL_LOG_DIR/<env>/<exp_prefix>/<foldername>.
Inside this folder, the experiment results are stored in progress.csv. We recommend using viskit to plot the results.
Note that this codebase is designed to work with robosuite environments only. For setting up your own environments, please follow these examples for reference. Notably, you will need to add the skill_config variable to the constructor, and define the keypoints for the affordance score by implementing the _get_skill_info function.
If you would like to know the inner workings of the primitives, refer to skill_controller.py and skills.py. Note that we use the term "skill" to refer to behavior primitives in the code.
Much of this codebase is directly based on RLkit, which itself is based on rllab.
In addition, the environments were developed as a forked branch of robosuite v1.1.0.
@inproceedings{nasiriany2022maple,
title={Augmenting Reinforcement Learning with Behavior Primitives for Diverse Manipulation Tasks},
author={Soroush Nasiriany and Huihan Liu and Yuke Zhu},
booktitle={IEEE International Conference on Robotics and Automation (ICRA)},
year={2022}
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent