A robot told to keep coffee away from a laptop could infer that a nearby phone deserves similar protection. That is the kind of generalization demonstrated by FLEX, a language-guided system developed by researchers at MIT and New York University to update robot motion planning from human corrections.
According to MIT News on November 9, 2026, the team tested FLEX in simulations involving a robot arm serving coffee, moving a candle and carrying a bowling ball. Volunteers typed natural-language corrections during the tasks. The study, led by MIT’s Andreea Bobu, was scheduled for presentation at the Conference on Robot Learning that month; the supplied material does not explicitly establish its peer-review status.
FLEX edits the features a robot considers when planning its movements through three steps. “Add” introduces a task-relevant feature identified by a correction, such as a mug to avoid while carrying a bowling ball. “Generalize” uses GPT-4o to infer a broader preference, such as protecting fragile objects. “Mask” uses the robot’s computer vision system to remove features associated with objects no longer present. MIT News reports that the full process takes about five seconds, without specifying the hardware, timing basis or variability.
In the described demonstrations, corrections about a laptop, a stack of paper and a mug were extended into avoidance preferences concerning electronics, flammable objects and fragile objects, respectively. The intended benefit is to reduce the need for people to list every object a robot should avoid, while allowing its planning to adapt as its surroundings change.
These examples support feasibility within the reported simulations. They do not establish how reliably FLEX infers a person’s wishes or whether it outperforms existing approaches. The report discusses methods that add individual features, but supplies no measured comparison, success rates, effect sizes or uncertainty estimates. Participant and trial counts are also unspecified.
The named paper, “Flexible Language-Guided Representation Editing for Adaptive Robot Behavior,” was not supplied, so its detailed methods and results cannot be assessed here. Failure cases, testing on unseen situations and possible overlap with model-training data remain unclear. An inferred rule could also differ from what a user intended. Transfer to physical robots, unfamiliar homes and diverse user preferences remains unestablished, and no independent evaluation or replication is supplied.
Cover illustration generated with AI; conceptual, not research evidence.