The Pulse

FLEX lets robots generalize human corrections

MIT and New York University researchers introduced FLEX, a system that lets robots use conversational corrections to infer broader preferences. In simulations, it generalized warnings about specific objects to categories such as electronics

AI.info Team · Written with AI · fact-checked · 2026-11-10

MIT engineers have developed FLEX, a learning platform that lets robots infer broader intent from conversational instructions, MIT News reported. The system, developed with researchers at New York University, is designed to make a correction apply beyond the specific object a person mentions. The study is set to be presented at the annual Conference on Robot Learning this month.

From a correction to a preference

FLEX stands for a flexible feature-editing pipeline for robot motion-planning. The researchers focused on helping robots identify a person’s underlying preferences from a limited number of natural-language commands, according to MIT News. Rather than only treating each instruction as a separate obstacle to avoid, the system aims to draw a broader rule from it.

The process begins with “Add,” which records an object mentioned in a correction as something the robot should avoid. In the example described by MIT News, a person asks a robot carrying a bowling ball not to drop it on a mug, and the robot adds the mug to its list of objects to avoid. That step resembles earlier approaches, in which each correction adds a feature and the robot updates its trajectory.

Three stages of learning

The second step, “Generalize,” uses a large language model — GPT-4o in this case — to look for patterns among the objects a person has identified. MIT News says the model can generalize from a mug to avoiding fragile objects, even when the person has not named that broader category. The researchers’ account presents the language model as drawing on patterns in its training data.

The third step, “Mask,” uses the robot’s computer vision system to check which objects are no longer present and remove them from its feature list. According to MIT News, the full three-step process takes about five seconds. The approach is intended to avoid a growing list that includes objects that may no longer matter in a changed environment.

Tests in simulated tasks

The researchers tested FLEX in simulations of three tasks: serving a cup of coffee, moving a candle and carrying a bowling ball. Volunteers interacted with a virtual robot by typing natural-language instructions, including warnings about spilling on a laptop, approaching a stack of paper and dropping cargo on a mug, MIT News reported.

In the reported examples, FLEX followed the instructions literally and also generalized them: a warning about a laptop led the robot to avoid electronics, while the paper and mug examples led it to avoid flammable objects and fragile items. The tests described by MIT News were simulations involving a robotic arm, rather than demonstrations of household robots operating in people’s homes.

Researchers’ aims

The study was led by Andreea Bobu, an assistant professor of aeronautics and astronautics at MIT. The co-authors are Audrey Lee SM ’26 and Ilia Sucholutsky of New York University. Bobu’s group develops autonomous agents that learn to work for, with and around people, and she is focused on assistive and household robots.

Bobu said FLEX could let a robot learn what matters to someone from a sentence or two instead of requiring the person to state every rule. The work received partial support from the MIT Generative AI Impact Consortium and the MIT Schwarzman College of Computing’s Social and Ethical Responsibilities of Computing program.