rddy / mimi

Code for the paper, "First Contact: Unsupervised Human-Machine Co-Adaptation via Mutual Information Maximization"

MIMI: Mutual Information-Maximizing Interface

MIMI is an algorithm for training an interface to map user command signals to system actions through unsupervised human-in-the-loop reinforcement learning.

Usage

1. Clone mimi into your home directory ~
2. Download data.zip and decompress it into mimi/
3. Setup an Anaconda virtual environment with conda create -n mimienv python=3.6
4. Install dependencies with pip install -r requirements.txt and pip install pyglet==1.5.11
5. Replace your_install_dir/gym/envs/box2d/lunar_lander.py with deps/box2d/lunar_lander.py
6. Install the mimi package with python setup.py install
7. Jupyter notebooks in mimi/notebooks provide an entry-point to the code base, where you can play around with the environments and reproduce the figures from the paper.

Citation

If you find this software useful in your work, we kindly request that you cite the following paper:

@article{mimi2022,
  title={First Contact: Unsupervised Human-Machine Co-Adaptation via Mutual Information Maximization},
  author={Reddy, Siddharth and Levine, Sergey and Dragan, Anca D.},
  journal={arXiv preprint arXiv:2205.12381},
  year={2022}
}

Latent Space Exploration (Experimental)

Explore the latent space of a generative model of MNIST images using hand gestures

1. Clone this repo
2. Set dvae_dir in mimi/utils.py
3. Download the MNIST dataset to mimi/data/mnist/mnist.npz
4. Run notebooks/mnist.ipynb