Imagine a robot butler collaborating seamlessly with you in your home, understanding your requests and taking action. This is the dream of embodied AI, where artificial intelligence interacts with the physical world. However, a new research benchmark called PARTNR reveals a significant challenge: current AI struggles with teamwork, especially in complex, real-world-like scenarios. PARTNR tests AI agents on over 100,000 household tasks in simulated environments. These tasks, expressed in natural language, range from simple instructions like "Take the glass and bowl to the kitchen and wash them" to more complex, multi-step scenarios requiring coordination, like setting a table for dinner. The benchmark simulates a human partner, adding an extra layer of complexity, as the AI must understand, anticipate, and adapt to human actions. The results? While humans achieve a 93% success rate on PARTNR tasks, state-of-the-art AI models only manage around 30% when deprived of perfect information. Even more striking, AI agents working as a team are often *slower* than a single AI agent, a stark contrast to human teams who outperform individuals. This "coordination burden" arises because the AI struggles to track its partner's actions, leading to duplicated efforts and unnecessary steps. For example, one AI agent might move a dish to the sink, only for its partner to move it back to the table, unaware of the previous action. The research also highlights the fragility of AI when faced with real-world imperfections. When relying on realistic perception systems that can make mistakes, the AI's performance drops even further. Similarly, if the AI's physical skills, like grasping an object, are less than perfect (as they would be in the real world), success rates plummet. However, there is a silver lining. The research explored fine-tuning smaller AI models, which showed promising results. These smaller models, while not as successful as larger ones in ideal scenarios, were significantly faster and proved more effective when collaborating with *real* humans in a separate experiment. This suggests that smaller, faster AI models may be better suited for real-world human-robot collaboration, providing a more seamless and responsive user experience. The PARTNR benchmark provides a crucial testing ground for future development in embodied AI, highlighting the importance of robust coordination, error recovery, and perception for creating truly collaborative robots.