Assessing the effectiveness of older adults' spatial descriptions in a fetch task

Assessing the effectiveness of older adults’ spatial descriptions in a fetch task Laura A. Carlson 1 (lcarlson@nd.edu) Marjorie Skubic 2 (skubicm@missouri.edu) Jared Miller 1 (jmille39@nd.edu) Zhiyu Huo 2 (zhiyuhuo@mail.missouri.edu) Tatiana Alexenko 3 (ta7cf@mail.missouri.edu) Department of Psychology, 119-D Haggar Hall, University of Notre Dame, Notre Dame, IN 46556 USA Electrical and Computer Engineering Department, University of Missouri, Columbia, MO Computer Science Department, University of Missouri, Columbia, MO Abstract older adults (Alexenko et al., 2013), parsing the natural language descriptions and coding them into chunks that can be converted into robot commands, recognizing key furniture items within a cluttered environment that are included in the descriptions, and identifying spatial relations within the horizontal plane (e.g., behind the couch) and the vertical plane (e.g., on top of the table) (Skubic et al, 2012). Given that the robot algorithms are driven by the properties of the spatial descriptions, in the current paper we examine how the communicative task of the speaker impacts the features of the descriptions, and present data that reflect the effectiveness of the descriptions. The current paper examines spatial descriptions provided by older adults in the context of a fetch task in a virtual house environment that mimics an eldercare setting. Sixty-four older adults provided directions for how to find a target or where to find a target to a robot or human (named Brian) avatar. There were systematic differences in the form and structure of the descriptions based on the communicative task. Specifically, how descriptions were longer, contained more detail, and were dynamically structured as compared to where descriptions. However, where descriptions were found to be more effective in conveying the target location, as assessed with a subsequent target selection task. Implications for the development of robot algorithms for the comprehension of naturalistic spatial language across these two communicative tasks are discussed. Spatial Directions and Spatial Descriptions A fetch task is one in which a speaker specifies the location of a desired target for an addressee whose goal is to retrieve the target. There are two ways in which the location can be indicated by the speaker. The speaker could provide directions that tell how to get to the target location or the speaker could provide descriptions that specify information about where a given target location is. Research has shown systematic differences in the type and structure of the language that is used for each of these communicative tasks. For example, Plumert et al. (1995) found that written directions on how to find a target in a hierarchically organized doll-house environment were more likely to provide more detailed messages and contain more spatial units that tended to be organized in a descending sequence (floor a room a reference object. e.g., The keys are on the first floor in the living room on the table.) as compared to written descriptions of where to find a target that were less detailed and organized in an ascending sequence (reference object a room a floor, e.g., The keys are on the table in the living room on the first floor). This distinction between how and where has also been characterized as dynamic and static (Wahlster. 1995, Fasola and Mataric 2012) spatial language, respectively, with dynamic stepping the addressee through the environment in a point by point fashion and static offering spatial information that does not embed the addressee in the environment. Dynamic spatial directions are also inherently sequential, while static descriptions are not. Nevertheless, static descriptions are often overlooked or treated the same Keywords: Human-robot interaction (HRI); spatial language; dynamic and static; how and where; effectiveness; fetch task; assistive robotics; eldercare. Introduction An emerging line of research in human-robot interaction involves the development of assistive devices for use in eldercare settings, either as social companions (e.g., Heerink et al., 2008; Kidd, Taggart, & Turkle, 2006; Libin & Cohen- Mansfield, 2004; Shibata, Kawaguchi, & Wada, 2011; Wada et al., 2003) or as task-oriented robots assisting with navigation (Montemerlo et al., 2002), managing medication (Tiwari et al., 2011)], and providing reminders (Pollack et al., 2002). Older adults also report wanting help with tasks such as cleaning, heavy lifting, and fetching objects (Beers et al., 2012). They also prefer to speak naturally to these assistive devices, rather than use a more constrained interface (Scopelliti, Guiliani, & Fornara, 2005). To accommodate these preferences, recently we gathered a corpus of spatial descriptions from older adults who interacted with an avatar within a virtual house setting in the context of a fetch task. Our primary goal in this project is to develop robot algorithms for the online comprehension of these natural language spatial descriptions and to test these algorithms in an analogous physical environment with a physical robot. In working toward this goal, on the basis of the corpus, we have identified key components that need to be developed for the robot including speech recognition for