🤖MIT's FLEX Lets You Tell Robots 'Don't Spill' in Plain English

Your robot can finally understand 'don't spill on my laptop'

MIT engineers just showed a robot can take a casual sentence like "Don't spill on my laptop" and figure out the rest on its own. The platform is called FLEX, and it's a flexible feature-editing pipeline for robot motion planning. Why should you care? Because the usual way to make a robot safe is to hand-write every constraint. Someone has to think of fragile objects, flammable stuff, and electronics in advance. FLEX pushes that work onto a large language model that reads the instruction and generalizes from it. If you're building task planners for manipulation robots, this is a different architecture to think about, not just a prompt trick. Here's how it works. The "Add" step puts relevant features into the robot's list of things that matter. The "Generalize" step uses an LLM to expand the avoid-list into broader rules. The "Mask" step removes objects that are no longer in the environment. Volunteers tested it in simulation on tasks like serving coffee, moving a candle, and carrying a bowling ball. The whole loop runs in about five seconds. The catch? These results come from simulation, not a warehouse floor. Five seconds per replanning cycle is fine for a tabletop arm and probably too slow for anything moving fast. Watch for the CLEAR Lab to move this onto physical hardware.