RoboDriver

A Unity ML-Agents environment for robotics transfer learning research. This repo contains the transfer learning and robotics parts.

Agent trained using ML-Agents

Agent's observations

Training process of an agent which takes 2 decisions per second

Training process of an agent which takes 25 decisions per second

Python Installation

1. Clone the project using a version control system of your choice.
2. Download and install the Anaconda. It is important to select Just Me in Install for: step and add Anaconda to the path to allow you to use the included batch files.
3. Open Anaconda Navigator once to finish up the installation. Just close it when it finishes loading.

Automatic installation using batch file:

4. Run Install.bat

Manual installation (If Conda is not in Path and batch files throw errors):

4. Open Anaconda Prompt, navigate to project root using cd path_to_folder
5. Run conda create -n RoboDriver python=3.7
6. Run activate RoboDriver
7. Run pip install mlagents
8. Run pip install matplotlib
9. Run pip install jupyterlab

Communicating with environment via UnityEnvironment class (LowLevelAPI)

Documentation for LowLevelAPI, more LowLevelAPI and GymWrapper.

Automatic running:

1. Run start_jupyter_lab.bat

Manual running (If Conda is not in Path and batch files throw errors):

1. Open Anaconda Prompt, navigate to project root using cd path/to/folder
2. Run activate RoboDriver
3. Run jupyter lab
4. Navigate to notebooks folder and run the notebook

Unity Installation

1. Install Unity Hub and Unity Engine 2020.3.25f1.

Unity project structure

To edit agent's settings open RoboDriverAgent prefab

Agent's scene