The Oz group at CMU, led by Joseph Bates (another of the original area chairs for Autonomous Agents 97), has made some major contributions to the development of computer agents as characters. Not the least of these has been the tireless efforts of the group to legitimize the very hard, and sometimes very mainstream AI-ish, problems to be solved in creating interesting, engaging, believable, interactive characters. A postulate of the OZ project's work is that the body of artistic knowledge developed for film animation, acting, and writing fiction, not only can be transferred from the original sources into interactive computing, but ultimately must be. The Oz group sought input from artists on the projection of personality into characters, and suggested early on that progress in these areas would generalize to products for entertainment, commercials, computer games, and corporate training.
Driving this research is the idea that much like musical conception, the artistic spark needed to inspire a truly great character might first be envisioned in an unencumbered purely artistic world, but ultimately must be manifested in the real world -- here in sequences of computer code, and graphical bit mappings. Along the way, great attention to detail is required so that the there originally there is not lost in the shuffle of timing, protocols, AI planning, knowledge representation, and the like.
One difficult problem the group has faced in bringing this work into the mainstream of AI is that the empirical testing of progress is difficult. The goals of the project are wedded to an artistic sensibility, albeit in the AI paradigm. Testing whether or not a character ``works'' as a plausible, interesting, entity, is a far cry from measuring execution time, polygons, or even the quasi-rigorous educational effectiveness of a program: there is no big-oh of plausibility.
One lesson that seems to have fallen out of the corpus of their research is that creative work on interesting interactive believable characters, just like characters for other mediums, requires that one must each time create, from scratch, a substantial amount of what is intuitively novel and intriguing about a new character. One way to consider this is that while tools and frameworks may be extrapolated, original art, once reused, is no longer original. The group has has additionally worked on usable models of emotion (Scott Reilly [Reilly1996]), interactive fiction (Peter Weyhrauch), natural language for believable agents (Bryan Loyall [Loyall & Bates1995]), and most recently on what we might call ``action discourse'' (Phoebe Sengers [Sengers1998]) wherein actions based on disparate motives are constrained to be presented in realistic sequences (and hence related to the work of Lester, et al., and others on diectic believability below). The tools generated by this group have been used by a number of others (e.g. Hayes-Roth et al. in the animated puppets work in the Virtual Theater project at KSL [see below]), and Wurst and McCartney at the University of Connecticut).
Like many agents researchers, they are probably best known by the look of some of their characters, i.e., as the ``Woggles guys and gals from CMU''. The woggles might be described as very cute, squishy, minimalist, highly expressive, playful, social balls of graphical fur. In their classic implementation they are, above all, playful. (See figure 3)
Most recently the group has spawned a commercial enterprise, Zoesis, which finally settled in Boston this past year. Necessarily, with the migration to building a commercial product, the group now has to be more protective of their current ideas. Zoesis still maintains strong ties to those remaining at CMU however, and its members remain active at conferences.