Adversarial Intrinsic Motivation for Reinforcement Learning

This repository contains code to accompany the paper Adversarial Intrinsic Motivation for Reinforcement Learning, published at the Thirty-fifth Conference on Neural Information Processing Systems (NeurIPS 2021).

This repo contains the code to run the simple grid world examples, as well as the experiments on the Fetch domain.

Requirements

A lot of the code, such as the TD3 implementation and the execution setup is taken from the stable-baselines repository and the accompanying RL Baselines zoo repository.

You will not need to install these repositories, but the dependencies are similar to those libraries. These requirements can be satisfied as follows:

pip install -r requirements.txt

Examples

Some examples of how to run the experiments.

Grid World Experiments

The code for all the grid world experiments should be the easiest to follow and run. This code is all in the directory grid_world_experiments. An example of a run here is:

python main.py --reward aim --dir aim_results

Fetch Robot Experiments

A single run on the FetchReach-v1 domain can be executed by running the following command:

python -u train.py --algo her --env FetchReach-v1 --tensorboard-log $PathToTensorboardLogs --eval-episodes 100 --eval-freq 2000 -f $PathToResultsFile --seed $SEED