Towards Efficient Llm Grounding For Embodied Ai