Hierarchical Abstraction with Language

This package contains the code for training goal conditioned policy with relabeling used in the Neurips 2019 paper "Language as an Abstraction for Hierarchical Deep Reinforcement Learning"

Disclaimer: This is not an official Google product.

Pre-requisite

Besides those listed in requirements.txt, you must also get the clevr-robot environment in order to run this out of box. However, if you don't wish to run on the clevr-robot environment, you can modify experiment_setup.py to use different environments.

Usage

python -m hal.train