What makes virtual agents believable?

ABSTRACT: In this paper we investigate the concept of believability and make an attempt to isolate individual characteristics (features) that contribute to making virtual characters believable. As the result of this investigation we have produced a formalisation of believability and based on this formalisation built a computational framework focused on simulation of believable virtual agents that possess the identified features. In order to test whether the identified features are, in fact, responsible for agents being perceived as more believable, we have conducted a user study. In this study we tested user reactions towards the virtual characters that were created for a simulation of aboriginal inhabitants of a particular area of Sydney, Australia in 1770 A.D. The participants of our user study were exposed to short simulated scenes, in which virtual agents performed some behaviour in two different ways (while possessing a certain aspect of believability vs. not possessing it). The results of the study indicate that virtual agents that appear resource bounded, are aware of their environment, own interaction capabilities and their state in the world, agents that can adapt to changes in the environment and exist in correct social context are those that are being perceived as more believable. Further in the paper we discuss these and other believability features and provide a quantitative analysis of the level of contribution for each such feature to the overall perceived believability of a virtual agent.