Objective: We aim to model the cognitive decision-making behaviour of autonomous virtual characters, in order to generate relevant learning situations in collective and complex activities and to foster learning.

Motivations: Virtual reality now makes it possible to design increasingly credible environments. After extensive work on the interaction of a user with their environment, and then on collaborative and cooperative interactions between several “real” users, research is now turning to the interaction of a user with autonomous virtual characters. The range of applications for this type of interaction is wide, and it would reduce the number of real participants required: crisis management training (handling a large number of casualties), modelling of virtual factories with many operators, modelling of children, etc. The new possibilities offered by virtual reality and agent technology in terms of scenarisation make it possible to design autonomous virtual characters that evolve in virtual environments. However, credibly modelling “humanised” behaviour remains a difficult problem, which artificial intelligence has been grappling with for many years.

Proposed approach: Our approach aims to account for the variability of human behaviour, decision-making processes in normal and degraded situations, erroneous behaviour, situated and contextualised cognition, and the human factors that influence these behaviours. We seek to model the impact of external characteristics (context and environment) on the characters' internal characteristics (emotions, physiology). An agent's perceptions can be modulated according to physiological criteria (hunger, fatigue, stress, etc.), its personality (curiosity, empathy, etc.) and its emotions (anger, fear, etc.). We also seek to model how the cognition of these autonomous virtual characters can be affected by external characteristics and by their internal characteristics (personality, emotions, physiology). We propose models and architectures that account for variable and erroneous processes. Depending on certain static or dynamic characteristics, autonomous virtual characters perceive and decide differently, and their goals change according to:

• physiological states (e.g. fatigue, stress, injuries, contamination);
• emotional states (e.g. fear, anger, pride, trust);
• personality traits (e.g. leadership, caution, courage, empathy, altruism, expertise).

We are also interested in the emergence of collective behaviour from highly cognitive individual behaviours.

Novelty: Our contribution in this field is to couple models developed in cognitive psychology with knowledge models. In particular, we have drawn on models from the field of human reliability, and more specifically from industrial risk research, such as:

• the Contextual Control Model (COCOM) proposed by [Hollnagel, 1993];
• the Cognitive Reliability and Error Analysis Method (CREAM) [Hollnagel, 1998];
• the Borderline Tolerated Conditions of Use (BTCU) model proposed by [Amalberti, 2001], [De la Garza, 2005];
• the ACTIVITY-DL formalism that we proposed for activity modelling.

Our work also addresses the coupling of these models with models already well established in the decision-making modelling of autonomous virtual characters, such as:

• the SOAR cognitive architecture [Laird et al., 1987];
• the OCEAN personality model [Costa et al., 1992];
• the OCC model of emotions [Ortony et al., 1988].