Imagine a robot that can organize your fridge exactly how you like it, without you having to lift a finger. That's the promise of APRICOT, a new AI system from Cornell and Stanford researchers. APRICOT combines the power of large language models (LLMs) with a clever approach to learning your preferences. Instead of painstakingly telling the robot exactly where everything goes, you simply show it a few examples. APRICOT then uses these demonstrations to generate a set of possible preferences. To refine its understanding, it strategically asks clarifying questions, like "Do you prefer to keep all dairy products together?" This active learning process allows APRICOT to quickly zero in on your unique organizational style. But what about those pesky real-world constraints? Fridges are notoriously packed, and a robot needs to avoid collisions. APRICOT's constraint-aware task planner tackles this challenge by generating plans that respect physical limitations while maximizing preference satisfaction. It even uses a "reflection" process where it iteratively refines its plan based on feedback from a world model (a 3D representation of the fridge). The researchers tested APRICOT on a real robot in various cluttered fridge scenarios. Impressively, it successfully organized items while respecting both preferences and spatial constraints, even adapting to changes in the environment, like when a user rearranged items mid-task. While still in the research phase, APRICOT represents a significant leap towards personalized robotic assistance in the home. Future research aims to enhance its ability to handle an even wider range of preferences and more complex constraints, paving the way for truly helpful home robots.