Paper Reading #1: Designing Effective Gaze Mechanisms for Virtual Agents

    The paper “Designing Effective Gaze Mechanisms for Virtual Agents” found at https://graphics.cs.wisc.edu/Papers/2012/APMG12/APMG12.pdf, was published through the University of Wisconsin-Madison. Sean Andrist and Tomislav Pejsa were two second year graduate students who were led by Proffessors Bilge Mutlu and Michael Gleicher. Dr. Mutlu directs the human computer interaction lab while Dr. Gleicher directs the graphics lab. He previously was a researcher at the Autodesk Vision Technology Center and at the Apple Computer’s Advanced Technology Group.

In their research, they created a model of gaze behavior. This model supplied various variables, which were manipulated to find what situation improved the interaction between the user and the agent. More specifically, they wished to improve learning and feelings of rapport especially in key areas like education, collaboration, and therapy. This paper focused on an educational scenario. However, a challenge that they encountered was developing effective gaze cues for virtual characters. Humans send non-verbal signals out constantly, and it is hard for a computer to mimic these subtle and yet complex movements. Yet, even more so, it’s hard to identify and manipulate a few of these patterns when there are so many. However, they were able to confirm their three hypotheses that, “The presence of an embodied agent will result in better recall performance than only hearing audio with no accompanying agent.”,  “An agent whom employs more affiliative gaze (maintains higher head alignment with the participant) will garner higher subjective evaluations than one which uses more referential gaze (maintains higher head alignment with the information being referred to).”, and  “Viewing an agent using more referential gaze should result in better recall performance. This will especially be true when the information to be recalled relies on building association to objects in the environment.”(5).

Their model comprises of six main components. Target, agent, and environmental parameters; head latency; velocity profiles for head and eye motion; oculomotor range specs (OMR, which makes sure the eyes don’t roll to the back of the head); head alignment preference; and the vestibule-ocular reflex (VOR, which is where the eyes stop moving when it reached its target). Their focus was in varying the head alignment preference. If it was set to zero percent, the head would stop moving once the eyes reached the target. On the other hand, if it was set to one hundred percent, the head keeps moving until it is aligned to look the same direction as the eyes. In order to make sure their model was accurate; they ran an experiment to validate it by checking its communicative accuracy and perceived naturalness. This experiment resulted in confirming that their model is just as accurate as when humans are acting in it.

            So, they used this model to develop an experiment that was implemented through a custom framework built on top of the unity game engine. The subjective and objective results prove their hypotheses and are displayed in the graphs below.

When looking at related work, both psychology and CHI literatures can be referenced. Some assumptions made were based on past psychology papers, but some newer documents are better associated with what this particular experiment is trying to accomplish. What makes this study a novelty is that it identifies unique parameters that can be mapped to specific outcomes. It essentially creates and uses low-level gaze variables for a high-level outcome. These are some of the corresponding papers:

In Psychology:

•          Effects of eye contact, posture and vocal inflection upon credibility and comprehension
•          Communicative effects of gaze Behavior
•          Effect of teacher’s gaze on children’s story recall
•          Facilitative effects of gaze upon learning
•          Does your gaze direction and head orientation shift my visual attention?

In CHI:

•          Experimenting with the gaze of a conversational agent
•          The impact of eye gaze on communication using humanoid avatars
•          A storytelling robot: Modeling and evaluation of human-like gaze behavior.
•          Modeling gaze behavior for conversational Agents
•          Automated eye motion using texture synthesis
•          Animating gaze shifts for virtual characters based on head movement propensity

In this particular experiment, the group took the whole system and divided it up into smaller parts as shown by the many graphs. They evaluated these parts through qualitative measures in both subjective and objective ways.

             Overall, I think this project revealed an important part of information that other researchers will be able to use. However, this is just one small part of gaze creation. Since they only focused on head alignment, further research can be made into target, agent, and environmental parameters; head latency; and velocity profiles for head and eye motion. I believe the evaluation that was performed was effective because they were focusing on specifics instead of a general consensus; thus, the reason for dividing it into many parts. Even though the idea of programming a certain gaze pattern for an agent is not a novel idea, the way they approached singling out one factor was helpful. This was an interesting study to read especially since something similar was looked at when I worked under Dr. Murphy.