Proactive learning of cognitive exercises with a social robot

Abstract

We introduce INtuitive PROgramming 2 (INPRO2), an improvement over our previous INPRO framework for learning board exercises via demonstrations. INPRO2 makes use of our Online Action Recognition through Unification (OARU) algorithm, which maintains and extends as needed a library of STRIPS action schemata that represent the dynamics, rules and goal of the exercise. OARU operates on a sequence of states shown by the user. Each state transition is either used to learn a new action, or is recognized as an instance of one action currently present in the library, possibly refining it. We have extended OARU to support negative examples (i.e. invalid moves that show forbidden state transitions) in order to increase the complexity of the exercises that can be learned. This new OARU's feature is exploited through another crucial element of INPRO2: its ability to proactively ask for the legality of certain moves to the user in critical situations, and fix overly permissive actions. We show an example of a typical INPRO2 learning session. We also outline a plan for a user study that will serve to assess the proactive behavior of the robot.