IBC

Accepted to the Fortieth International Conference on Machine Learning (ICML 2023)

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.

Installation

1. Install Conda package manager

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/

2. Create a Conda environment

conda env create -f conda_env.yml
conda activate ibc

3. Manually install other dependencies

# 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

Running experiments

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

Acknowledgements

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.

BibTeX

@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}
}