Curiosity Driven Goal Exploration of Learned Disentangled Goal Spaces

This folder hosts the code to reproduce the results presented in the paper Curiosity Driven Goal Exploration of Learned Disentangled Goal Spaces. In this paper, we study the impact of the structure of the representation when it is used as a goal space in Intrinsically Motivated Goal Exploration Processes. Experiments are performed on a simple task in which multi-joint arm must handle and gather an object in a 2D space in the presence of a distractor which cannot be handled and follows a random walk.

Running the experiments

To run a single experiment, you can run one of the three following python scripts:

• rpe.py to perform a Random Parameterization Exploration
• mge_efr.py to perform a Modular(Random) Goal Exploration using Engineered Features Representation
• mge_representation.py to perform a Modular(Random) Goal Exploration using a learned Representation

Examples of some exploration algorithms together with a demonstration of the environment are provided in the notebook:

• notebooks/ArmBalls.ipynb

In order to reproduce the results, you can also run a full campaign batch by running the following python scripts:

• script_rpe.py to perform a Random Parameterization Exploration campaign
• script_mge_efr.py to perform a Modular(Random) Goal Exploration using Engineered Features Representation campaign
• script_mge_rep.py to perform a Modular(Random) Goal Exploration using a learned Representation campaign and executing the generated scripts.

Alternatively, one can use the data provided in the file results/armballs_dataset.pkl (see notebook).

Finally, to generate the different figures out of the raw results, you can use the notebooks:

• results/Experiment_Visualization.ipynb to compare the performances of the different algorithms or see the individual results.

Installation

The code is intended to be run on python 3.5 or higher.

To install, clone repository, then:

cd Curiosity_Driven_Goal_Exploration
pip install -e .

Dependencies (among others):

• explauto
• pytorch
• visdom