Using Vision Language Models as Closed-Loop Symbolic Planners for Robotic Applications: A Control-Theoretic Perspective