Unity Reinforcement Learning

Development of a simple computer game that uses reinforcement learning to teach an agent progressing through the environment.

Getting Started

Dependencies

• Python 3.8
• PyTorch 1.7.1
• ML-agents 0.28.0
• Unity Editor 2020.3.30r1 with ML Agents 2.0.1 package

Installing

• Clone the repository
• Create en empty Unity 3D project
• Move the repository files to the project directory
• Add one of the existing scenes from Assets/Scenes to the project Hierarchy
• Install Python 3.8
• In the project directory open terminal and create a virtual environment:

python -m venv venv

• Activate the environment:

venv\Scripts\activate

• Install PyTorch 1.7.1

pip3 install torch~=1.7.1 -f https://download.pytorch.org/whl/torch_stable.html

• Install ML-agents 0.28.0

pip install mlagents==0.28.0

• Launch the Unity project > Window > Package Manager > Unity Registry > Search for ML Agents package > Install ML Agents 2.0.1

Executing program

To use an existing ML model:

• In Unity select the RollerAgent object
• In the Behavour Parameters module select a model from the Assets/ML-Models directory
• Press Play in Unity Editor

To train a new model:

• In Unity select the Agent object
• In the Behavour Parameters module set the Behavour Type to Default
• Open terminal in the project directory and activate the virtual environment:

venv\Scripts\activate

• Train a new model:

mlagents-learn <(optional) path to config file, e.g. config/new_config.yaml> --run-id=<unique name, e.g. run1>

• Press Play in Unity Editor
• The model file will be saved in results/run-id directory with .onnx extension

Help

ML-Agents docummentation