Embodied Agent Interface: Benchmarking LLMs for Embodied Decision Making