Cognitive models, for instance, raise the level of abstraction at which synthetic characters can be animated just as physical models permit characters to be animated realistically without pixel-level control [19]. These cognitive models often form a hierarchy with various levels of behavioural interpretation. A set of emotions built on cognitive ideas drives animated characters such as Lyotard, a synthetic cat implemented by Bates [4]. Bates' architecture, Tok, and the emotion unit, Em, are based on the OCC (Ortney, Clore and Collins [44]) cognitive structure of emotions. This yields a more tractable hierarchical approach to emotion synthesis and permits the high level development of behaviourally convincing characters.
Blumberg [6] discusses ethological models of competing behaviours as an effective way of generating autonomous behaviour in synthetic characters. These ethological analogies, permit a good visualization and a hierarchical explanation of what is actually a complex set of rules and functions. Much like the emotions discussed above, goals and other factors can compete in the hierarchy as well.
Additionally, mechanisms that have been used by cognitive scientists from experimental studies could be used as priors in artificial behaviours. The effect of short term memory and the observed decay rates over time have been investigated by [33]. Very similar techniques are used in the short term memory of Blumberg's virtual dog, Silas, where memory decays more rapidly as stimuli increase and accumulate in memory.
Other less graphical examples of cognitive-based behaviour include Weizenbaum's Eliza [69] program. The system uses a series of natural language rules as well as pattern matching to emulate a therapy session. Since the introduction of Eliza, similar dialogue systems are becoming more common and are routinely evaluated with Turing tests to determine how convincing the interactive behaviour is to human users. However, the criticism of these cognitive/computational models is often their dissociation from the physical reality and their lack of grounding in perception and data. Often, these employ a top-down Von-Neumann design (without mentioning bottom-up concepts) and lack grounded or embodied intelligence [11]. This puts into question the applicability of such rule-based mechanisms to domains beyond their particular task.
A more direct motivation for this thesis are the recent cognitive investigations of the primate pre motor cortex that suggest that the development of imitation skills and higher order intellectual activity is related to imitation-specific neurons [50]. Mataric uses such concepts to argue for imitation based learning and discusses the improved acquisition of complex visual gestures in human subjects [14]. In addition, Mataric's robotics research has also been inspired from this cognitive insight as well as ethological models for multi-agent interaction [39].