Skip to content
Go to file

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

How to run HAAR

Welcome to our code release of Hierachical Reinforcement Learning with Advantage-Based Auxiliary Rewards, accepted to NeurIPS 2019.

Check out the videos and our paper.

We adapted the code from SNN4HRL heavily, and also modified some files in rllab. For historical reasons, some folder names remain the same as in SNN4HRL.

You will have to configure rllab(with mujoco) properly. Our configuration is the same as SNN4HRL, so we encourage you to check out their readme first and have all the dependencies properly installed.

The first step of running HAAR involves pre-training a set of low-level skills, which can be achieved by using SNN4HRL

cd rllab/sandbox/snn4hrl
python runs/

Now you should obtain a .pkl file storing the low-level policy.

Change the path pkl_path in rllab/sandbox/snn4hrl/runs/ to where your pre-trained low-level skills lie (as a pickle file) before running it.

python runs/

This will give you the result of non-annealed HAAR on Ant Maze environment. The result can be found in rllab/data/local/tmp/exp_name, as a .csv file. The column named wrapped_succesRate indicates the success rate of the ant.

If you desire to run annealed HAAR, change the option in the configuration files inside rllab/sandbox/snn4hrl/runs/configs. Set time_step_agg_anneal to True.

You can also transfer both high and low-level skills learned from a previous task to a new task with

All experimental results are reproducible with this code release. To reproduce/design different experiments, we encourage you to look at files in runs/ and tweak the parameters in configs/.

If you have any questions, please open an Issue on this GitHub page.


Code accompanying HAAR paper, NeurIPS 2019 - Hierarchical Reinforcement Learning with Advantage-Based Auxiliary Rewards




No releases published


No packages published
You can’t perform that action at this time.