Imagine a robot smoothly navigating your home, effortlessly handling everyday chores. This is the dream driving research into using Multimodal Large Language Models (MLLMs)—AI that understands both text and images—as the “brains” of in-home robots. But how close are we to this reality? A new study, creating a first-of-its-kind benchmark called MMRo, puts MLLMs to the test, evaluating their ability to perceive, plan, reason, and operate safely in a home environment. Researchers assessed essential skills like recognizing objects (color, shape, material), planning multi-step tasks (like heating a meal), understanding spatial relationships, and handling delicate or sharp objects. The results? While promising in some areas, the study revealed current MLLMs aren't quite ready to take the helm. Even top commercial models struggled with basic perception tasks like accurately identifying colors and counting objects. Planning presented another hurdle—while capable of decomposing complex tasks into sub-steps, many MLLMs couldn't reliably sequence them. For example, getting the steps involved in using a microwave or grasping a delicate item was difficult. Safety is paramount in robotics, and here, the study showed a concerning gap. While some models could identify sharp or hot objects, their performance wasn’t consistently reliable, highlighting the need for significant improvement before these AI systems can safely interact with the physical world. The MMRo benchmark provides a crucial framework for evaluating future MLLMs, guiding the development of more robust and capable AI for in-home robotics. Though a fully autonomous robotic butler might still be a few years off, this research marks an important step towards making that dream a reality.