Urban search and rescue (US&R) refer to operations conducted in collapsed man-made structures. It has been recognized as a useful domain for studying human-AI interaction. Human-AI teaming in the domain of US&R is a widely researched area, due in part to the complications that arise out of introducing AI into an unpredictable environment such as a collapsed building. In this study we investigate different AI communication styles in a team-based experimental search and rescue scenario. By inviting human participants via a simulated Minecraft based reconstruction of urban search and rescue mission maps, we collect data gathered by incorporating a "Wizard of Oz" design, with the researchers playing the role of an AI advisor, giving guidance to the team of participants during the experiment. The focus areas for the study are the adherence to guidance under different communication styles, usage of the styles, and participants response to these styles. While the objective of the Minecraft based experiment for the participants was to save as many victims trapped in collapsed buildings via fifteen-minute run experiments, this study does not evaluate team performance and instead looks at how teams adhere to the advisor's guidance, disregard the guidance, and ask for additional information. We present the results of our experiments via two specific guidance conditions (explicit vs information shaping) and modes of communication (voice vs text). Our results indicate that there was a greater adherence with information shaping guidance when compared with explicit guidance and participants seem to respond more to voice over text. The results and discussion presented in this study would help drive the design of human-AI teaming systems especially when the AI's roles is to provide guidance to human teams.
Keywords
AI, HUMAN FACTORS, SIMULATIONS, TEAM TRAINING
Additional Keywords