This project explores the integration of Reinforcement Learning (RL) and classical pathfinding algorithms to create an adaptive and challenging environment for game AI research. The core objective was to train a Rabbit agent using Tabular Q-Learning to maximize its survival and resource consumption in a 2D grid-based world, while being pursued by a computationally optimal Hunter agent utilizing the A* search algorithm (with Manhattan distance heuristic). The Rabbit agent's state representation was kept small and manageable (approx. 2 * 10**4 states) by using a relative "Radar" vision system, rather than absolute coordinates. Its behavior was shaped by a reward function that heavily penalized capture (−100), slightly penalized movement (−1), and highly rewarded carrot consumption (+50), which acted as a reset-free goal to prioritize long-term efficiency and survival. Training results, analyzed via custom metrics visualization, demonstrated the successful learning progression of the agent. The Success Rate (survival vs. capture) stabilized at 40% - 55%, while the Efficiency metric (steps taken per carrot) sharply decreased from initial values of 50-70 steps to an average of 15-20 steps. This confirms the Q-Learning agent successfully learned to navigate the maze efficiently and flee from the deterministic, near-optimal A* adversary, proving the viability of RL-based agents as dynamic game characters. The project utilized Python 3 and Tkinter for visualization, with key assistance from Generative AI for logic verification and crucial system optimizations, such as decoupling the simulation logic from the rendering process for high-speed training.