Dipartimento di Ingegneria informatica, automatica e gestionale

Goal reasoning is a main objective for robot task execution. Here we propose a deep model for learning to infer a next goal, while performing an activity. Because predicting the next goal state requires a robot language, not comparable to sentences, we introduce a specific metric for optimization, which is related to the representation the robot has of the scene. Experiments of the proposed idea and method have been done at a warehouse with a humanoid robot performing tasks assisting a maintenance technician working at a production line.