Resumen de Learning from a robot: creating synthetic psychologically plausible agents
Due to technological advancements, robots will soon become part of our daily lives and interact with us on a frequent basis. Thus, given the robots’ primal goal to socially interact with humans, a fundamental question arises: “How can we create robots that are successful in interaction and are accepted by humans?” and more importantly: “how do we measure success?”. An answer to this question may lie in the psychological plausibility of such robots. User perception is in principle hypothesis testing; hence, if the agent’s behaviour or traits match user’s expectations, it can be considered plausible and therefore accepted. Thus, to construct psychologically plausible robots, one can define a basic psychological engine or a set of features that if fulfilled, can account for believability.
To answer the central question, we divide plausibility into two approaches: psychological and implementation. To be more precise, we are interested in understanding “What are the behavioural traits that allow humans to perceive a robot as a believable agent?” which in turn leads us to examine “what kind of control system does a robot require to be psychologically plausible?”. We offer four possible psychological benchmarks for consideration: autonomy, morphology, social competence and task competence. The suggested benchmarks aim at decomposing psychological plausibility to discrete parts that can be tested empirically and use their interactions in practice for the meaningful design and development of social robots. Although there have been some attempts to examine humans’ responses towards robots and establish standard metrics, few attempts have been made to establish psychological benchmarks.
Thus, the main goal of this thesis is to create a robot that is accepted by its human partners. To do so, we focus on the psychological plausibility of the robot. More specifically, we describe the design, development and study of social robots intended for dyadic interactions and propose four benchmarks aimed at evaluating the robot’s behaviour and plausibility at the following domains: autonomy, morphology, social competence and task competence. The first chapters of this thesis aim at providing a general overview of the morphology and behavioural components of current robotic systems that socially interact with humans. In the following chapters, we go in more detail on our current system implementation and the evaluation methods of our research questions.
More specifically, chapter 2 provides examples of archetypes that are characterised by efforts made to understand and imitate biological organisms regarding functionality, physical appearance, processes and complex life-like behaviours. Given our focus on robots with social character, we are compelled to define “What is a robot?” as we do not find sufficient the existing definitions. In chapter 3, we present the existing behavioural and social strategies employed to create robots that socially interact with humans. From these, we identify three key concepts that are relevant to our proposed taxonomy: the expression of internal states, the usage of gaze and the elicitation of proactive behavior. Chapter 4 revolves around the presentation of the proposed taxonomy where we explain in more detail the motivation behind selecting the criteria above and our evaluation methodologies. In chapter 5, we present our implementation of the H5W_Alpha sociable robot whose control architecture is based on the “Distributed Adaptive Control” theory of brain, body and mind.
Chapter 6 answers the following question: “How do the various robotic features affect the plausibility of the robot?”. Here, we show the first attempt to implement the DAC architecture on a social humanoid robot, endowed with a set of drives that aim at initiating and maintaining an interaction with a user through a game-like scenario. Results indicate that the robot is able to trigger behaviours that aim at satisfying the robot’s needs and illustrate an interplay between drives, emotions, perceived stimuli and actions. Indeed, the robot can behave autonomously, even if not all preconditions are matched. To evaluate the transparency of the robot’s communication channels, we varied the facial features of the robot (eyebrows, eye opening and mouth) and asked participants to rate them in terms of valence and arousal. Results suggest that there is a correlation with valence and mouth and eyes and arousal but not a combination of both. To assess the robot’s social competence, we decomposed social behaviour in a number of discrete cues such as gestures, touch, speech, gaze, facial expressions and proactive behaviour. To assess how these behavioural components affect the robot’s believability we devised five interaction scenarios of increased complexity and asked participants to evaluate the robot. Results suggest that the more the robot appears socially competent, the higher it scores in believability. Additionally, we hypothesised that social competence could trigger empathic responses. We therefore manipulated the robot’s gaze model and facial expressions and looked at the empathic relation between the participant and the robot. Indeed, results indicate that participants showed empathic responses toward the robot, however it was not the result of the robot’s behavior but the predisposition of individuals to show empathy.
Finally, we evaluated the robot’s task and social competence in dyadic scenarios. First, we examined the role of facial expressions and gaze model in an educational task. We conducted this experiment with adults and children. Although results were not conclusive regarding the effects of the robot’s social components or task competence, we identified an impact of the role of gaze in engagement. Having evaluated the robot’s psychological plausibility with adults, we now focused on the psychological plausibility of the robot with children. More specifically we asked “Can we extract valuable information or design guidelines from children’s drawings?”. To answer this question, we exposed children to three different robotic platforms and asked them to evaluate them. Additionally, we asked them to draw the robot they would like to have and interact with and assessed their drawings regarding functionality and morphology. Results suggest that children tend to design multi-purpose robots that are anthropomorphic but are more machine than human-like. Finally, the central question we ask to evaluate the robot’s task competence is: “Does the robot’s help mechanism allow students gain a better understanding of the task and therefore be accepted by students as a peer?”. As a tutoring task, we used the balance beam problem and varied the nature of the robot’s help by providing hits (open/closed) and distractions (jokes/trivia). In this scenario, the robot used the virtual balance (evaluated in the previous study) as a tool to convey content. The aim of this study was to see what are the minimum set of tools and behavioural components needed to efficiently and effectively teach children physics. Results indicate that children enjoyed the interaction and found the feedback of the robot helpful.
