This repo is a collection of a few reinforcement learning projects I've been working on to learn more about AI training. Everything is built using Python, Stable Baselines3 (PPO), and Gymnasium.
| Simulation Demo | Controller Preview |
|---|---|
 |
 |
I built an AI that learns to navigate a grid filled with random obstacles to reach a goal. I went through a few iterations to get the behavior right:
- Model 1 (8-Sensor Radar): The baseline model. It only knew what was directly next to it, which made it act pretty blind.
- Model 2 (Ray-Casting): Upgraded the AI with longer-range "LIDAR" vision so it could see walls before hitting them.
- Model 2.1 (Stable Train): Fixed a major issue where the AI would get stuck in infinite loops. It was too afraid to backtrack because of the point penalties, so I had to rebalance the reward system.
- Model 3 (Spatial Awareness): The final and most advanced version, capable of actually planning a path out of U-shaped traps and complex obstacle clusters.
| Simulation Demo | Controller Preview |
|---|---|
 |
 |
For the next step, I wanted to try Multi-Agent Reinforcement Learning (MARL). This project features two separate neural networks training in the same arena against each other.
- The Predator: Learns to chase down the prey as fast as possible.
- The Prey: Learns evasive maneuvers to stay alive.
- Because Stable Baselines3 is built for single agents, I set up a custom training loop where they take turns learning and adapting to each other's strategies.