Leveling Up Efficiency

• 

  Testing the LM model

• 

  Training the LM model

• 

  Playable Mario Game addition

🚀 Inspiration -We were inspired by the idea of teaching AI to master platforming games—just like human players learn through trial and error. By integrating reinforcement learning into a classic game, we wanted to explore how AI could optimize movement, avoid obstacles, and reach a goal efficiently.

🎮 What It Does -Our AI-controlled Mario learns the best path to the flag through Proximal Policy Optimization (PPO). The AI makes real-time decisions, navigates platforms, avoids Goombas, and adapts to challenges, all while improving its strategy through reinforcement learning.

🛠️ How We Built It -Frontend: React + TypeScript + Vite for a real-time game interface. -Backend: FastAPI + WebSockets for AI processing and streaming. -AI Model: Stable-Baselines3 (PPO) trained in a custom Gym environment using OpenCV for rendering.

⚡ Challenges We Ran Into -AI initially kept running left instead of toward the goal. -Collision detection issues with the Goomba required fine-tuning hitboxes. -Balancing rewards to ensure AI learns optimal movement without unintended behaviors.

🏆 Accomplishments We're Proud Of -Successfully built an AI that learns to complete levels dynamically. -Replaced Pygame with OpenCV for better AI processing. -Implemented real-time gameplay streaming with WebSockets.

📚 What We Learned -Fine-tuning rewards is crucial for reinforcement learning. -AI can develop unexpected strategies if not trained carefully. -WebSockets are highly effective for streaming AI gameplay.

🚀 What's Next for Untitled -More complex levels with moving obstacles and power-ups. -Train AI with human-like behavior (e.g., speedruns, adaptive difficulty). -Deploy a web-based AI playground for users to experiment with AI learning.

Built With

• baselines3
• fastapi
• gym
• opencv
• phaser.js
• react
• typescript
• uvicorn
• vite
• websockets