RLSM64 is a reinforcement learning project that uses computer vision to train an AI agent to play Super Mario 64. The agent interacts with the game through a custom OpenAI Gym environment, leveraging frame analysis and reward shaping to encourage exploration, object collection, and path following.
- Computer Vision: Uses OpenCV to process game frames, detect movement, coins, and Mario's pose.
- Reinforcement Learning: Employs Stable Baselines3 (PPO/DQN) for training the agent.
- Custom Environment: Modular
MarioEnvclass for interfacing with the game and managing state. - Reward System: Rewards for movement, collecting coins, exploring new areas, and following paths; penalties for death or leaving walkable areas.
- Exploration Tracking: Tracks visited positions on a fine grid to encourage thorough exploration.
- Path Following (Goal): Ongoing improvements to help the agent recognize and follow level paths for better progress.
RLSM64/
├── env/ # Custom environment and wrappers
├── vision/ # Computer vision utilities
├── rewards/ # Reward calculation logic
├── scripts/ # Training and evaluation scripts
├── config/ # Config files (YAML/JSON)
├── tests/ # Unit and integration tests
- Improve the agent's ability to make progress and follow intended level paths.
- Refine reward shaping to better distinguish between meaningful exploration and random movement.
- Modularize code for maintainability and extensibility.
- Install dependencies from
requirements.txt. - Run training with
python scripts/train_rl_mario.py. - Adjust environment, vision, and reward logic as needed for new levels or behaviors.
- Enhance path detection and following using higher resolution and finer grid tracking.
- Add configuration files for easy parameter tuning.
- Expand test coverage and add evaluation scripts.
For questions or contributions, open an issue or pull request.