Expressive Cognitive Architecture for a Curious Social Robot

Introduction

Cognitive architectures attempt to capture important aspects of human cognition in generalized conceptual and computational models [39, 58]. While many architectures focus on perception, attention, memory, and learning, few have focused on social interaction with other agents [59].
Additionally, the field of artificial curiosity has gained much attention in recent years [23, 29, 37, 40, 50]. Artificial curiosity, also called intrinsic motivation models, is based on the premise that the agent can quantify how much it can learn from specific actions, and then selects those actions that maximize learnability [49, 66]. The field has become mainstream in the wake of deep learning, where numerous models incorporate intrinsic motivation modules to add intrinsic rewards in an otherwise sparse reward space [1, 29, 37]. Nevertheless, most artificial curiosity studies focus on the game or robot itself, while only few have incorporated social interaction into their scenarios [24, 55].

In this article, we present a novel cognitive architecture that incorporates an Artificial Curiosity component and a Social Expressivity component (see Figure 1). For the robot to externalize its curiosity states, i.e., help users interacting with the robot perceive and understand the rationale behind its curiosity-driven choices, the variables of the artificial curiosity module need to be socially expressed in human terms. The Embodied Curiosity sub-component within the Social Expressivity component translates artificial curiosity variables to verbal and non-verbal social expressions. This way, the robot is both curious, i.e., selects tasks that maximize learnability, and social, i.e., expresses its intrinsic motivation to other social agents (users).

We implement the expressive cognitive architecture in a social robot for children, that plays a tangram puzzle game on a tablet with a child. The robot and child take turns in selecting which puzzle to try and solve next, and then attempt to solve it. The robot implements an artificial curiosity component that estimates the learnability of each puzzle in the selection and expresses its estimation in verbal and non-verbal communication to the child, e.g., “I chose this puzzle because I can learn the most from it,” with an enthusiastic smile. We validate the curiosity module and show that the robot learns, estimates the learnability, and improves when its selection is based on its learnability estimation. We then report on a user study with 92 children who played with either a curious or non-curious robot. We show that children who played with the curious robot chose significantly more unknown and challenging puzzles to solve and had a significantly smaller decrease in estimated curiosity.

The contributions of this article are as follows:

(i) a novel expressive cognitive architecture that links intrinsic motivation to social expressivity

(ii) a validation scheme of an artificial curiosity module;

(iii) a user study with a fully autonomous curious social robot that influences children’s
behaviors; 

(iv) estimation of children’s curiosity based on the same cognitive architecture.

The structure of the article is as follows: we first present related work, then introduce the general architecture, followed by the specific implementation and description of the user study. We then present the validation and user study analysis, followed by discussion and future work.

Read More