Towards Automated Semantic Interpretability in Reinforcement Learning via Vision-Language Models