Researchers from MIT and New York University presented FLEX at this month’s Conference on Robot Learning – a flexible feature‑editing pipeline that lets robots interpret natural‑language commands beyond their literal meaning. Users type conversational instructions; the robot follows them literally and also infers higher‑level intentions.\n\nIn a simulated coffee‑serving scenario, a user types “Don’t spill on the laptop.” FLEX invokes a large language model (LLM) such as GPT‑4o, which recognizes that liquids and electronics often cause damage, and therefore adds both the laptop and a nearby phone to its avoidance list.\n\nFLEX operates in three steps:\n1. Add – Based on the user’s utterance, the robot adds relevant objects to a feature list (e.g., “Don’t drop it on my mug” adds the mug).\n2. Generalize – The LLM searches its massive text corpus for patterns and automatically abstracts a broader rule, such as avoiding all fragile items.\n3. Mask – The robot’s vision system checks the current scene, removes objects that are no longer present, and keeps the feature list concise. The whole pipeline runs in about five seconds.\n\nThe team tested FLEX on three tasks – serving coffee, moving a candle, and transporting a bowling ball. With only a few natural‑language prompts, volunteers enabled the robot to reason about avoiding electronics, flammable objects, and fragile items. The results show that robots can acquire user preferences from minimal language input instead of exhaustive rule specification, bringing them closer to practical home assistants.\n\nReview: FLEX illustrates a promising integration of LLM‑derived common sense with real‑time robot perception, paving the way for truly conversational household robots.