MIT researchers developed a framework that lets a user correct a robot’s behavior during deployment using simple interactions, such as by pointing to an item, tracing a trajectory, or nudging the robot’s arm.