A Cognitive Model for Spatial Perspective taking

When communicating with other people, one of the basic things that people must do is take others’ perspectives. Most of the experimental work on spatial language and perspective taking has focused on four frames of reference: exocentric (world-based, such as “Go north”), egocentric (self-based, “Turn to my left”), addressee-centered (otherbased, “Turn to your left”) and object-centric (objectbased, “The fork is to the left of the plate”) (CarsonRadvansky & Logan, 1997; Carson-Radvansky & Radvansky, 1996; Levelt, 1984). Any time egocentric or addressee-centered frames of references are used, spatial perspective taking is needed: egocentric utterances require the listener to take the speaker’s perspective, and addressee-centered utterances require the speaker to have already taken the listener’s perspective. As part of a project to make intelligent agents and robots more useful to people, we have been developing cognitive models of spatial cognition and perspective taking (Trafton, Schultz, Cassimatis et al., under review; Trafton, Schultz, Perzanowski et al., under review). Unfortunately, there are relatively few computational cognitive models of spatial cognition available in order to implement spatial perspective-taking models. One recent entrée to spatial cognition research has been ACT-R/S (Harrison & Schunn, 2003). ACT-R/S extends ACT-R (Anderson & Lebiere, 1998) to implement a theory about spatial reasoning. It posits that spatial representations of objects are egocentric and dynamically updated (Wang & Spelke, 2002). ACT-R/S represents objects using vectors to the visible sides of the object. It has the ability to track these objects through a configural buffer, a data structure analogous to the other buffers of ACT-R that stores each object once it has been identified. The coordinate vectors of the objects in the buffer are then dynamically updated as the agent moves throughout the spatial domain. The configural buffer, unlike the visual and retrieval buffers of ACT-R, can hold more than one object to account for the fact that animals have been shown to track more than one landmark at once while moving through the world. The other spatial buffers within ACT-R/S, the visual and manipulative buffers, are not a central part of the work described here and are described more fully elsewhere (Harrison & Schunn, 2003).