Multi-Modal Grounded Planning and Efficient Replanning For Learning Embodied Agents with A Few Examples