A Computational Framework for Expressive, Personality-based, Non-verbal Behaviour for Affective 3D Character Agents

Badler defined virtual humanoid characters as computer models of humans that can be used in several applications such as training and entertainment. For the humanoid characters to be credible and human-like, they must exhibit realistic and consistent nonverbal behavior. It is this consistency that ultimately instills in human users a sense that the characters have distinct personalities. Despite this importance, relatively little work has so far been done on the consistency of a 3D character’s behaviour during interaction with human users and their environments. Current 3D virtual character systems lack the ability to maintain the consistency of their behaviour during real-time interaction which can lead to users’ frustration and resentment.This thesis presents the design, implementation, and evaluation of a system named “RealAct” that controls the non-verbal behaviour of virtual characters. To make the virtual characters behave in a believable and consistent manner, the system controls non-verbal behavior such as gaze, facial expression, gesture and posture to give the impression of a specific personality type. The design and development of different modules of the RealAct system, e.g. for controlling the behaviour and generating emotion, is directly modelled from existing behavioural and computational literature. In addition to these core modules, the RealAct system contains a library of modules that are specifically geared toward real-time behavior control needs such as sensory inputs, scheduling of behaviour, and controlling the attention of the character.To evaluate and validate different aspects of the RealAct system, four experimental studies using both passive video-based and presential real-time paradigms were performed. The results of these experiments show that the amount of extraversion and emotional-stability that participants attributed to virtual characters depended on a combination of facial expression, gaze and posture and gestures that they exhibited. In summary, it was shown that the RealAct is effective in conveying the impression of the personality of virtual characters to users. It is hoped that the RealAct system provides a promising framework to guide the modelling of personality in virtual characters and how to create specific characters.