Game environments provide needful settings to support artificial intelligence (AI) research. However, due to static configurations the robust generalization and adaptive learning are limited. Current Reinforcement Learning models, such as Deep Q-Networks (DQN), often face difficulty with sparse reward signals and environment overfitting, which leads to restricting their use in variable, complex procedures, making it necessary to make use of hybrid approaches to improve robustness. On the other hand, heuristic algorithms have less learning flexibility but assures tactical accuracy. This paper introduces an adaptive framework combining generative artificial intelligence, DQN reinforcement learning, and heuristic planning. A lightweight Generative Adversarial Network (GAN) combines semantic tasks and creates multiple levels through guided token placement. By interacting with the environment the DQN agent learns the best strategies, while the challenging situations are helped by an A* planner. A feedback loop adjusts the environment’s complexity to ensure consistent progress by monitoring performance metrics like completion time, success rate and collision frequency. Experiments done in puzzle solving, maze navigation, and adversarial tasks have shown that the hybrid model achieved 12–22% greater success rates, 18% better collision avoidance, and 23% quicker task completion compared to single-module baselines. The proposed framework shows improved generalization across different tasks, shows adaptive gaming and future AI research applications.