Text this: Evaluating a Hybrid LLM Q-Learning/DQN Framework for Adaptive Obstacle Avoidance in Embedded Robotics