Researchers at MIT and New York University have developed FLEX, a system that lets people correct a robot in conversational language and has the robot infer the broader intent behind the correction, MIT News reported on November 9, 2026. The paper, "Flexible Language-Guided Representation Editing for Adaptive Robot Behavior," is on arXiv and is due to be presented at the Conference on Robot Learning this month.
The study was led by Andreea Bobu, an MIT assistant professor of aeronautics and astronautics who heads the CLEAR Lab, with co-authors Audrey Lee of MIT and Ilia Sucholutsky of NYU.
FLEX edits the set of features a robot's motion planner pays attention to, in three steps that take about five seconds. First, it adds the object named in the user's command, such as a mug, to the list of features to avoid. Second, a large language model, GPT-4o, looks for patterns among the listed features and generalizes them into a rule, for example avoiding fragile objects in general. Third, the robot's vision system masks out features for objects that are no longer in the scene.
The team tested the system on a simulated robot arm across three tasks: serving coffee, moving a candle and carrying a bowling ball, with volunteers typing commands. According to MIT News, the robot followed the literal instructions and also inferred higher-level rules such as avoiding electronics, flammable objects and fragile items. The article does not report numerical success rates; those are in the paper.
The approach matters for anyone building robots that take natural-language feedback. Instead of retraining a policy or writing a rule for every object, a single correction can be generalized by a language model into a constraint the planner applies to new objects it has not been told about.