Team: Emma Drobina, Dr. Jenay Beer

Electronic tutoring agents, such as robots, have much potential to provide customized and engaging education to students.  In math education specifically, robots have been successfully applied to tutor grade school students through customized, adaptive, and emotionally engaging interaction.  However, much work is needed to fully understand how a robotic tutor can re-engage and maintain student participation during a tutoring session. Human tutors utilize both verbal and non-verbal strategies for engagement, to increase attention, motivation, and involvement. These strategies include personalized instruction, verbal encouragement, using gestures to maintain and direct attention, making eye contact, and speaking in a positive tone.  Therefore, the goal of this study was to investigate effective and social strategies for a robotic tutor to provide students feedback, based on the students’ math performance and emotional state. We simulated a robotic tutor utilizing 20 different tutoring strategies. Via Mechanical Turk, we showed 110 participants video clips of our simulated robot. We asked participants to identify via multiple choice (1) the situation where the robots’ behavior and strategy is most appropriate (e.g., a scenario where a student expresses frustration); and (2) the role the robot takes in relation to the student (e.g., the robot asks as a peer or instructor).

Afterwards, we reviewed the results and tabulated the top three answers for each emotion identification question.

chart of intended & identified emotions. intended emotion: relief. identified emotion: relief; satisfaction; happiness. intended emotion: satisfaction. identified emotion: satisfaction; happiness; trust. intended emotion: happiness. identified emotion: happiness; excitment; satisfaction. intended emotion: stress. identified emotion: stress; frustration. intended emotion: sadness. identified emotion: sadness; frustration. intended emotion: resentment. identified emotion: frustration; isolation; sadness. intended emotion: trust. identified emotion: trust, satisfaction. intended emotion: excitement. identified emotion: excitement; happiness; satisfaction. intended emotion: frustration. identified emotion: frustration; sadness. intended emotion: isolation. identified emotion: frustration, stress.

The emotion we intended the robot to respond to, vs. the emotion participants identified the robot as responding to.

Our results indicate that the robot’s responses were generally appropriate to the emotion the student displayed, as respondents tended to select the correct emotion or a closely related one (e.g. frustration instead of stress). We also examined whether participants correctly identified the role the robot took in the tutoring session.

A pie chart. 35% of roles were identified incorrectly. 65% of roles were identified correctly.

We counted each scenario where the majority of participants identified the robot’s role correctly as an overall correctly identified role; likewise, if the majority of participants identified the robot’s role incorrectly, it was counted as an incorrectly identified role. While the majority of roles were classified correctly, our most interesting finding was that all misidentified scenarios were cases where the peer condition was incorrectly identified as the instructor condition. This indicates that more careful consideration is needed to correctly establish a peer relationship between human student and robot tutor, which has implications that should be taken into account in terms of the overall trend towards developing tutors in HRI.

I was the lead researcher on this project, which was made possible by a Magellan Grant from the University of South Carolina’s Office of Undergraduate Research. With assistance from Dr. Jenay Beer, I analyzed both the HRI/HCI and education literature, to identify the behaviors of tutors and the emotions of students in the classroom. I selected five positive emotions (happiness, satisfaction, enthusiasm, trust, and relief) and five negative emotions (stress, despair, frustration, resentment, and isolation) that students might feel, as well as two roles that a tutor might play (one where the robot is a peer tutor, and one where it is a traditional instructor). I then designed 20 scenarios, one for each combination of role and emotion, and created behavior for the robot based on each scenario. Once the survey was complete, I also analyzed the data and created graphical breakdowns of the participant responses. Finally, I presented my research as a poster during the April 2018 Discover USC research showcase at University of South Carolina.

This project taught me how to design research, from start to finish. I reviewed the literature to identify a need and proposed an idea to Dr. Beer. I wrote the proposal for the Magellan Grant, and I designed the survey and wrote the questions. Once the data was collected, I analyzed it, created the graphs, and presented it to an audience. Overall, I became much more familiar with the field of educational technology and with all the phases of project design.