Learning Gaze Behaviors for Balancing Participation in Group Human-Robot Interactions

Abstract

Robots can affect group dynamics. In particular, prior work has shown that robots that use hand-crafted gaze heuristics can influence human participation in group interactions. However, hand-crafting robot behaviors can be difficult and might have unexpected results in groups. Thus, this work explores learning robot gaze behaviors that balance human participation in conversational interactions. More specifically, we examine two techniques for learning a gaze policy from data: imitation learning (IL) and batch reinforcement learning (RL). First, we formulate the problem of learning a gaze policy as a sequential decision-making task focused on human turn-taking. Second, we experimentally show that IL can be used to combine strategies from hand-crafted gaze behaviors, and we formulate a novel reward function to achieve a similar result using batch RL. Finally, we conduct an offline evaluation of IL and RL policies and compare them via a user study (N=50). The results from the study show that the learned behavior policies did not compromise the interaction. Interestingly, the proposed reward for the RL formulation enabled the robot to encourage participants to take more turns during group human-robot interactions than one of the gaze heuristic behaviors from prior work. Also, the imitation learning policy led to more active participation from human participants than another prior heuristic behavior.

Not Only WEIRD but “Uncanny”? A Systematic Review of Diversity in Human–Robot Interaction Research

“…Others [205,206] used classifications, such as experts and novice users. Still others [177,207] simply reported the presence or absence or previous interactions between participants and robots. With this in mind, we identified 69 studies reporting that at least some participants were already familiar with robots and 97 where at least some were not.…”

Robotic Mental Well-being Coaches for the Workplace

“…Within the HRI literature, very recently a number of works, e.g., [4,11,17,31], have started to investigate and evaluate the use of adaptive and personalised robots by showing that adaptive configurations are very promising for robotic applications [39]. For example, Axelsson and Skantze [4] presented a fully automated system for building adaptive presentations for embodied agents.…”

“…Hence, we formulated the problem of conducting the well-being practice in HRI as a sequential decision-making problem for selecting the next action in the coaching dialogue flow. At every turn 𝑡 (i.e., the robotic coach asks questions, and the participant answers) [17], the robotic coach environment is captured as an observation state variable 𝑠 𝑡 . The robotic coach, similar to the human well-being coach, may choose an action 𝑎 𝑡 at every turn 𝑡 that allows it to move forward with the conversation for delivering coaching practice by asking for a follow-up question, summarising what the coachee just said, or asking for a new episode.…”

“…This is a well-known problem in HRI. Past works have shown how adaptive models were better performance-wise, but users -interacting with the systemusually cannot perceive the differences across adaptive system's configurations [4,17].…”