Emerging Technologies · By Brian Fulcher
Prototype or demonstration. The team tested interactive corrections in simulation and used physical robot demonstrations in a separate offline analysis.
Last reviewed: November 9, 2026.
You tell a robot carrying coffee to give your laptop a wide berth. Should it also avoid the phone beside it? A person can often infer why you asked. A robot planner needs a way to turn that correction into something it can measure.
What changed
MIT described FLEX on November 9, 2026, ahead of the Conference on Robot Learning. The MIT and New York University team first posted the paper on October 20. FLEX changes which details a robot considers when planning a movement after someone types a correction.
How the correction changes a route
The research team’s project page describes three operations. FLEX can add something the planner overlooked, broaden a specific concern to related objects, or set aside an irrelevant detail. A language model helps decide which edits to make. A route planner then weighs the updated features when evaluating paths.
In the coffee example, a correction about the laptop can add that object and then widen the concern to electronics. The planner can then account for the phone too. The demonstrations help the system learn how much weight to give the different features. That separation matters. Knowing which objects to protect is a different problem from deciding how far away to travel.
What the researchers actually tested
The paper’s methods and results describe simulated navigation and tabletop tasks. In a study with 12 university recruits, participants watched simulated arm movements and typed corrections toward preferences the researchers had defined in advance. The authors report that participants rated FLEX’s understanding of intent higher than the two comparison methods they tried. Physical robot demonstrations were used separately to test offline learning from those corrections.
The study leaves open how the system would behave during a live sequence of corrections on physical hardware. The authors also identify unusual personal preferences and tasks involving flexible objects or rapidly changing scenes as limitations.
Who might use it
Researchers can inspect the public paper and project repository. The repository showed a README and license when checked. I found no verified turnkey release, consumer purchase option, or deployment program for FLEX. A future system could help someone adjust a supervised assistive task or a workbench robot without rewriting a motion plan. Both uses would need testing in those settings.
Root Access Take
I’d make every inferred preference visible and reversible. If I tell a robot to avoid my phone, I want to see whether it inferred electronics, fragile objects or everything on that side of the desk. I also want a separate stop control and physical limits that an inferred preference cannot override. A useful correction should reduce the work of supervising the robot.
For another example of how task evidence differs from a staged result, see the whole-body robot control explainer.
What to watch next
Look for a supervised hardware study where people can give their own corrections, notice an overbroad inference and undo it. The important measurements would include time to recover from a bad guess, unexpected movements and whether people can tell what changed.
Primary sources
MIT News, November 9, 2026 · Lee and colleagues, arXiv v1, October 20, 2026 · MIT CLEAR Lab project page · FLEX repository, checked November 9, 2026.