Tell a robot “don’t spill on the laptop” and it will dodge the laptop. Ask a person the same thing and they quietly understand that phones, tablets and anything with a battery are also off limits. On November 9, 2026, MIT News described a system that tries to close that gap: FLEX, a pipeline from MIT and New York University researchers that lets robots read meaning beyond the literal words of a correction.
The work is slated for presentation at the Conference on Robot Learning this month. It targets a practical problem for household and assistive robots: nobody wants to list every rule a robot should follow, yet every home is different.
The problem with literal corrections
Motion-planning robots do their trajectory math with numbers, and, as study author Nathan Dennler of MIT’s Computer Science and Artificial Intelligence Laboratory put it, consumers cannot be expected to learn that math. Researchers have long tried to let people steer robots in plain language: clarifying a goal, nudging a path, or flagging an unsafe route.
Those methods mostly use language to say how to physically perform a task. According to the MIT release, the team wanted the robot to learn more from each sentence, inferring an underlying preference from only a few commands. The study was led by Andreea Bobu, an assistant professor of aeronautics and astronautics who heads MIT’s Collaborative Learning and Autonomy Research (CLEAR) Lab, with co-authors Audrey Lee and Ilia Sucholutsky of NYU.
How FLEX works: Add, Generalize, Mask
FLEX is a three-step process, described in the paper “Flexible Language-Guided Representation Editing for Adaptive Robot Behavior.”
- Add. When a person says something like “Don’t drop it on my mug!”, the robot adds the mug to its list of environmental features that matter for the task. This is how earlier approaches already worked.
- Generalize. This is the new part. A large language model, GPT-4o in this study, looks at the mug and every other feature the person has flagged, then finds the pattern. Here it concludes the robot should avoid fragile objects, even though nobody said so.
- Mask. The robot’s ordinary computer vision checks which objects are still in the scene and removes features for things that are gone, so the list does not grow forever.
MIT says the whole loop takes about five seconds. The Mask step addresses a real weakness of the Add-only approach: as Dennler noted, feature lists get long, and some entries stop being relevant when the environment changes.
What the team tested
The researchers tried FLEX in simulation with a virtual robot arm doing three tasks: serving a cup of coffee, moving a candle, and carrying a bowling ball. Volunteers typed natural-language commands such as “Don’t spill on my laptop,” “Stay away from the stack of paper,” and “Don’t drop on the mug.”
In all three scenarios, MIT reports, the robot followed the literal command and also reasoned at a higher level: it avoided spilling on electronics, kept away from flammable objects, and avoided dropping cargo on fragile things. In the coffee example, the updated path steered clear of both the laptop and a nearby phone without the person naming the phone.
The MIT release does not report numerical success rates, and the demonstrations were in simulation, so real-world performance with physical robots remains to be shown.
What it means
The interesting idea is where the common sense lives. FLEX does not hard-code a rule about electronics. It borrows the broad background knowledge a language model absorbed from text, such as the idea that liquids and devices do not mix, and uses it to extend one correction into a category of things to avoid.
For anyone building assistive or household robots, that points toward a different kind of onboarding. As Bobu said in the release, no robot comes out of the box knowing what matters to you, and picking that up from a sentence or two is the difference between a robot people tolerate and one they want at home.
It also suggests design questions for the field. How should a robot confirm a generalization before acting on it? What happens when a language model generalizes wrongly? The release does not address these, and they are likely to matter once such systems leave simulation.
The takeaway
FLEX is a research result, not a product. Its contribution is a compact way to turn conversational corrections into lasting, broader behavior: add what was said, generalize what was meant, and mask what is gone. If it holds up on real hardware, robots that learn your household’s preferences by listening, rather than by programming, come a step closer.