Pikachu-Volleyball-RL

In this project, the primary goal is to train an AI agent capable of playing the classic game Pikachu Volleyball. To achieve this, a combination of Reinforcement Learning (RL) techniques, Evolution Strategy (ES), and Selfplay mechanisms are employed. The project aims to create an AI agent that can outperform rule-based policies and exhibit robust gameplay across various scenarios.

Requirements

• Python 3
• Stable Baselines3
• Gymnasium
• pytorch

Installment

1. Clone the project recursively

   git clone --recurse-submodules {project_url}

2. Install gym-pikachu-volleyball

   pip install ./gym-pikachu-volleyball

3. Install requirement packages

   pip install -r requirements

Training

After install necessary packages, you can run python train_[algorithm].py in script folder to train the agent by yourself.

Currently there are 4 algorithms in training script

• train_ppo.py: train agent using ppo model
• train_es.py: train agent using evolution strategy
• train_ppo_selfplay.py: train agent using ppo ppo with selfplay
• train_es_selfplay.py: train agent using evolution strategy with selfplay

Evaluation

To evaluate the model you haved trained, update the file name in corresponding eval file to your model path, and then run one of below script

• eval_ppo.py: evaluate ppo performance
• eval_es.py: evaluate evolution strategy performance

For evaluating two agents against each other, use the --second_model argument. To visualize agent actions, set --eval to false in order to render the environment.

You can also try the model that have been trained in model folder.

Result

See the detail training process and result on TRAINING.md

Acknowledge

This project is inspired by slimevolleygym, where various algorithms were employed to train an agent in playing Slime Volleyball.

License

This project is licensed under the MIT License.