DRLeague is an environment developed for the training of Deep Reinforcement Learning Agents, inspired by the game Rocket League.
The environment was developed with the use of Unity ML-Agents
To run the environment locally, you have to download Unity 2020.3.0 or Above.
You can clone this repository or download it in zip format.
After import it on Unity, you can go to the directory Assets/Scenes and see the following:
-
Aerial Scene: Scene of the Aerial Minigame with player controlled behavior, for gameplay tests.
-
Barrier Scene: Scene of the Barrier Minigame with player controlled behavior, for gameplay tests.
-
Penalty Scene: Scene of the Penalty Minigame with player controlled behavior, for gameplay tests.
-
MainScene: Scene containing the core gameplay mechanics, inspired by Rocket League, containing the flow of a regular match, with a player control.
-
MLScenes: Folder of Scenes containing the minigames adapted for autonomous controlled behavior, using the ML Agents components. Each scene contains a Game Environment with a pre-trained brain dedicated to the minigame.
Going to the MLScenes folder, you can see the Scenes integrated with the MLAgents framework, related to the implemented minigames. Opening the PenaltyScene for example:
The Object GameManagerToExperiment has a environment with a pre-trained agent, showing a example of the integration. You can press Play in the scene and see the autonomous behavior running.
To train a new brain, we use the ML-Agents Trainer. You can see a tutorial of how to install it here.
Select one of three scenes of MLScenes folder. Disable the GameManagerToExperiment object and enable the TrainingContainer. After this, you can see several environments in the scene view:
Now, open a prompt, according to your installation mode, and run the following command:
mlagents-learn <path/to/your/file.yaml> --run-id=<YourName>
You should see the instances playing the game.
To be added.
The car, arena models and the RoboUtils script are made by Roboleague, under the license CC BY-NC 4.0.
The project is licensed under the MIT license.