New AI system teaches robots to infer intent, not just follow instructions

Researchers at MIT and New York University unveiled FLEX, a learning pipeline that lets robots infer higher-level intent from everyday language, aiming to reduce the micromanagement typically required to guide machines. Presented at the Conference on Robot Learning, the three-step system—Add, Generalize, Mask—pairs standard perception with a large language model (GPT-4o) to extend user commands into broader safety and planning rules. When told, for example, “Don’t spill on the laptop,” the robot generalizes to avoid all electronics; “Don’t drop it on my mug” prompts avoidance of fragile objects. In simulations involving coffee serving, candle handling and moving a bowling ball, FLEX updated its motion plans within about five seconds and pruned irrelevant features as environments changed. The project, led by MIT’s Andreea Bobu with collaborators including Nathan Dennler, Audrey Lee and Ilia Sucholutsky, targets more intuitive household and assistive robots. Funding support came from the MIT Generative AI Impact Consortium and the MIT Schwarzman College of Computing’s Social and Ethical Responsibilities of Computing program.

Read more