People rarely tell robots everything directly. If a person says "don't spill coffee on the laptop," they mean much more: don't touch the electronics, don't knock over the cup, don't cause an accident. A new project by engineers at the Massachusetts Institute of Technology (MIT) and New York University teaches machines to grasp this hidden meaning.
The development is called FLEX - a flexible feature-editing system for robot motion planning. The results will be presented at the annual Conference on Robot Learning (CoRL 2026), which will be held in Austin, Texas, from November 9 to 12. The paper is also posted on the arXiv preprint server.
FLEX is based on a large language model (LLM). This is a type of artificial intelligence trained on huge arrays of text from the internet. Such a model can find general patterns. For example, it "knows" that liquids and electronics do not mix well, because texts often contain stories about spilled coffee and a broken computer.
The system works in three steps. First, a person gives a command, and the robot adds a new object to the list of things to avoid. Then the LLM comes into play: it generalizes the specific instructions and derives a rule - say, "avoid all fragile objects." In the third step, a computer vision system checks which objects have already disappeared from view and removes them from the list. The entire cycle takes about five seconds.
The tests were conducted in simulation. Volunteers typed commands for a virtual robotic arm that performed three tasks: carrying a cup of coffee, moving a candle and transporting a bowling ball. The phrases were simple: "Don't spill on the laptop," "Stay away from the stack of papers," "Don't drop on the mug." The robot carried out what was said literally, but in addition figured out on its own not to touch electronics, not to approach flammable objects and not to drop its load onto fragile things.
"Everyone has different homes and different values, so no robot from the factory knows what matters to you specifically," says the head of the research, Andrea Bobu, a professor of aeronautics and astronautics at MIT. "FLEX can pick that up from one or two sentences."
However, the authors emphasize: for now this is only a laboratory demonstration in a virtual environment. There is no talk of timelines for such robots appearing on sale. In addition, the study shows that the system is capable of making generalizations based on textual experience, but does not prove that the robot will always behave safely in a real home.