Learning to Control Self-Assembling Morphologies

NeurIPS 2019 (Spotlight)
Winner of Virtual Creatures Competition at GECCO 2019, Prague

[Project Website] [Demo Video]

Deepak Pathak*, Chris Lu*, Trevor Darrell, Phillip Isola, Alexei A. Efros
University of California, Berkeley
MIT
(* equal contribution)

This is a pytorch based implementation for our paper on learning to control self-assembling agents using deep reinforcement learning. We investigate a modular co-evolution strategy: a collection of primitive agents learns to dynamically self-assemble into composite bodies while also learning to coordinate their behavior to control these bodies. We learn compositional policies to demonstrate better zero-shot generalization. If you find this work useful in your research, please cite:

@inproceedings{pathak19assemblies,
    Author = {Pathak, Deepak and Lu, Chris and Darrell, Trevor and
              Isola, Phillip and Efros, Alexei A.},
    Title = {Learning to Control Self-Assembling Morphologies:
              A Study of Generalization via Modularity},
    Booktitle = {arXiv preprint arXiv:1902.05546},
    Year = {2019}
}

Installation and Usage

1. Setting up repository

git clone https://github.com/pathak22/modular-assemblies.gitcd modular-assemblies/
git clone https://github.com/Unity-Technologies/ml-agents.gitcd ml-agents/
git reset --hard 6c5255ecd ..
bash envs/setup_env.sh

python3 -m venv assemblyEnvsource $PWD/assemblyEnv/bin/activate
pip install --upgrade pip

2. Installation

   # Install Pytorch from http://pytorch.org/pip install http://download.pytorch.org/whl/cu80/torch-0.3.0.post4-cp35-cp35m-linux_x86_64.whl
   pip install torchvision
   pip install --upgrade visdom# Install baselines for Atari preprocessingpip install gym==0.9.4 # baselines install latest gym first automatically, but latest gym has moved to mujoco5 so first install old gym and then install baselinesgit clone https://github.com/openai/baselines.gitcd baselines
   git reset --hard b5be53d
   pip install -e .# Additional packagespip install numpy
   pip install matplotlib
   pip install pillow
   pip install opencv-python# foldcd modular-assemblies/src/
   git clone https://github.com/nearai/pytorch-tools.gitcd pytorch-tools/
   git reset --hard 09dccb2
   python setup.py install

   pip install -r requirements.txt

• Quick setup for exact replication:

• CUDNN-5.1, CUDA-8.0, Python-3.5

• Requirements:

• Detailed setup, skip to quick setup for exact replication:

3. Run code

cd modular-assemblies/src/
python test_env.py

Acknowledgement

Builds upon Ilya Kostrikov's Pytorch PPO implementation.