ultimate-volleyball

Ultimate Volleyball

About

Ultimate Volleyball is a multi-agent reinforcement learning environment built on Unity ML-Agents.

See 'Ultimate Volleyball Micro-Machine Learning Course' for an updated step-by-step micro-course.

Version: Up-to-date with ML-Agents Release 19

Contents

1. Getting started
2. Training
3. Self-play
4. Environment description
5. Baselines

Getting Started

1. Install the Unity ML-Agents toolkit (Release 19+) by following the installation instructions.
2. Download or clone this repo containing the ultimate-volleyball Unity project.
3. Open the ultimate-volleyball project in Unity (Unity Hub → Projects → Add → Select root folder for this repo).
4. Load the VolleyballMain scene (Project panel → Assets → Scenes → VolleyballMain.unity).
5. Click the ▶ button at the top of the window. This will run the agent in inference mode using the provided baseline model.

Training

1. If you previously changed Behavior Type to Heuristic Only, ensure that the Behavior Type is set back to Default (see Heuristic Mode).
2. Activate the virtual environment containing your installation of ml-agents.
3. Make a copy of the provided training config file in a convenient working directory.
4. Run from the command line mlagents-learn --run-id= --time-scale=1
   • Replace `` with the actual path to the file in Step 3
5. When you see the message "Start training by pressing the Play button in the Unity Editor", click ▶ within the Unity GUI.
6. From another terminal window, navigate to the same directory you ran Step 4 from, and run tensorboard --logdir results to observe the training process.

For more detailed instructions, check the ML-Agents getting started guide.

Self-Play

To enable self-play:

1. Set either Purple or Blue Agent Team ID to 1.

2. Include the self-play hyperparameter hierarchy in your trainer config file, or use the provided file in config/Volleyball_SelfPlay.yaml (ML-Agents Documentation)

3. Set your reward function in ResolveEvent() in VolleyballEnvController.cs.

Environment Description

Goal: Get the ball to bounce in the opponent's side of the court while preventing the ball bouncing into your own court.

Action space:

4 discrete action branches:

• Forward motion (3 possible actions: forward, backward, no action)
• Rotation (3 possible actions: rotate left, rotate right, no action)
• Side motion (3 possible actions: left, right, no action)
• Jump (2 possible actions: jump, no action)

Observation space:

Total size: 11

• Agent Y-rotation (1)
• Normalised directional vector from agent to ball (3)
• Distance from agent to ball (1)
• Agent X, Y, Z velocity (3)
• Ball X, Y, Z relative velocity (3)

Reward function:

The project contains some examples of how the reward function can be defined. The base example gives a +1 reward each time the agent hits the ball over the net.

Baselines

The following baselines are included:

• Volleyball_Random.onnx - Random agent
• Volleyball_SelfPlay.onnx - Trained using PPO with Self-Play in 60M steps
• Volleyball.onnx - Trained using PPO in 60M steps (without Self-Play)