Engineers at MIT and New York University have built an AI platform that lets robots read the larger meaning behind a person's words, rather than following commands literally. Called FLEX, it was announced by MIT News on 9 November 2026 and will be presented at the Conference on Robot Learning.

People rarely say exactly what they mean. Told “Don't spill on the laptop!” while carrying coffee, a robot should understand that liquids and electronics do not mix, and steer clear of every device nearby.

The coffee test

In simulated trials, FLEX did exactly that. Using a large language model, the robot reasoned from the laptop warning to a wider rule and avoided both the laptop and a nearby phone, without being told to. Study author Nathan Dennler, an MIT postdoc, said language models absorb this kind of common sense from the vast text they were trained on.

Three steps, five seconds

FLEX works in three steps taking about five seconds. “Add” lists the object named in a command; “Generalize” uses the language model, here GPT-4o, to find the pattern behind it, such as fragility; “Mask” uses computer vision to drop objects no longer present.

Tested on three household tasks

The team tested FLEX in simulations of a robot arm serving coffee, moving a candle and carrying a bowling ball. Volunteers typed commands such as “Stay away from the stack of paper” and “Don't drop it on my mug.” Each time, the robot reasoned further, avoiding electronics, flammable objects and fragile things.

The work was led by MIT's Andreea Bobu. “What excites me about FLEX is that it can pick that up from a sentence or two instead of asking you to spell out every rule,” she said. The tests ran in simulation; a physical robot is the next step.