PWIL: Primal Wasserstein Imitation Learning

Robert Dadashi, Leonard Hussenot, Matthieu Geist, Olivier Pietquin

This directory contains the source code accompanying the paper: Primal Wasserstein Imitation Learning https://arxiv.org/abs/2006.04678.

Dependencies

PWIL is compatible with Python 3.7.7. You can install the dependencies using:

pip install -r requirements.txt

You will also need to install Mujoco and use a valid license. Follow the install instructions here.

Expert demonstrations

PWIL demonstrations are available in a GCS bucket.

DEMO_DIR=/tmp/demonstrations
mkdir $DEMO_DIR
gsutil cp -r gs://gresearch/pwil/* $DEMO_DIR

Run PWIL

python -m pwil.trainer --workdir='/tmp/pwil' --env_name='Hopper-v2' --demo_dir=$DEMO_DIR