This repository contains research code for the ICLR 2021 paper Domain-Robust Visual Imitation Learning with Mutual Information Constraints.
-
To replicate the experiments in this project, you need to install the Mujoco simulation software with a valid license. You can find instructions here.
-
The rest of the requirements can be installed with conda, by utilizing the provided environment file:
conda env create -f environment.yml
conda activate autonomous_imitation
We explain how to collect expert/prior data and perform observational imitation to
replicate the paper's experiments in the notebook training_notebook.ipynb.
This notebook can be accessed via executing jupyter notebook
after activating the conda environment.
To reproduce the plots, we provide two functions to process and display the results collected:
-
plot.py - to obtain a plot of multiple algorithms for a single observational imitation problem.
-
multi_plot.py - to compare plots for multiple observational imitation problems partially sharing the same axis.
e.g. to plot the results for 1-Linked Inverted Pendulum -> 1-Linked Colored Inverted Pendulum:
python plot.py experiments_data/InvertedPendulum_to_colored --min_score 5 --max_score 50 --logdir . --epochs 20 --cumulative --show_legend --yaxis 'Scaled cumulative reward'
For further details on their usage, please run:
python plot.py -h
python multi_plot.py -h
@inproceedings{cetin2021domainrobust,
title={Domain-Robust Visual Imitation Learning with Mutual Information Constraints},
author={Edoardo Cetin and Oya Celiktutan},
booktitle={International Conference on Learning Representations},
year={2021},
url={https://openreview.net/forum?id=QubpWYfdNry}
}