This is an official implementation for IBC from our paper: Demonstration-free Autonomous Reinforcement Learning via Implicit and Bidirectional Curriculum by Jigang Kim*, Daesol Cho* (*Equally contributed), and H. Jin Kim
The instructions below were tested on Ubuntu 20.04, but should work on other Linux distros as well.
Conda package manager is required for installing python dependencies. Follow the link below to install conda:
https://docs.conda.io/projects/conda/en/latest/user-guide/install/
conda env create -f conda_env.yml
conda activate ibc
# Install a version of pytorch appropriate for your machine. For example,
conda install pytorch torchvision torchaudio pytorch-cuda=11.6 -c pytorch -c nvidia
# Install metaworld for sawyer env.
pip install git+https://github.com/rlworkgroup/metaworld.git@master#egg=metaworld
Set the path parameters default_save_path_prefix and workspace_path for your machine by following the instructions in config/paths/template.yaml.
Below are the commands for running IBC for the six environments:
python train.py env=tabletop_manipulation
python train.py env=sawyer_door
python train.py env=fetch_pickandplace_ergodic
python train.py env=fetch_push_ergodic
python train.py env=fetch_reach_ergodic
python train.py env=point_umaze
This repository contains modified open-source code from the official implementation of HGG. It also contains open-source implementations of various RL environments such as earl_benchmark, mujoco-maze, and metaworld.
@article{kim2023demonstration,
title={Demonstration-free Autonomous Reinforcement Learning via Implicit and Bidirectional Curriculum},
author={Kim, Jigang and Cho, Daesol and Kim, H Jin},
journal={arXiv preprint arXiv:2305.09943},
year={2023}
}