Codebase for RAIL : Risk Averse Imitation Learning, presented at the Deep Reinforcement Learning Symposium at NIPS 2017 and published as Extended Abstract in AAMAS-2018.

• Set up MuJoCo on your machine. Please download mjpro131 for compatibility with the rest of the code.
• Install Anaconda Python 2.7 and set it as the default python:

$ export PATH="/home/username/anaconda2/bin:$PATH"
$ which python
/home/username/anaconda2/bin/python

Make changes in the .bashrc for permanent results.

• Install the required packages.

pip install mujoco-py==0.5.7
pip install theano
pip install gym

• Clone the OpenAI-imitation (GAIL) repository. git clone https://github.com/openai/imitation.git

• Clone the RAIL bitbucket repository. git clone https://abhisheknaik96@bitbucket.org/intelpclfad/rail.git

• Add the path to the RAIL repository to $PYTHONPATH

• Copy the expert_policies directory from GAIL to RAIL.

• Run the initialization script:

username@machine:/path/to/RAIL$ chmod +x initialize.sh
username@machine:/path/to/RAIL$ ./initialize.sh

• Run the training script. Example usage:

username@machine:/path/to/RAIL$ chmod +x run/run_all/*
username@machine:/path/to/RAIL$ ./run/run_all/run_hopper.sh > training_logs/hopper_log.log

In case of any issues, feel free to contact the authors/developers :

• Anirban Santara (nrbnsntr@gmail.com)
• Abhishek Naik (abhisheknaik22296@gmail.com)