Employing Models of Human Social Motor Behavior for Artificial Agent Trainers

Many everyday tasks require individuals to work together as a team to achieve a task goal. For many complex or high-stakes multi-agent activities, team members are required to participate in simulated training exercises to develop the taskand team-work (coordination) skills needed tomaximize task performance. Such training, however, can be both timeand labor-intensive, requiring the participation of full teams and expert instructors. One way to minimize these costs is to augment team training scenarios with interactive artificial agents (AAs) capable of robust, ‘human-like’ behavioral interaction. With regard to perceptual-motor tasks specifically, recent research suggests that this can be achieved using task dynamical models derived from the dynamical primitives of human motor behavior. To investigate the degree to which such models can be employed for human team training, we examined whether a task dynamical model of human herding behavior could be embedded in the control architecture of an AA to train novice human-actors to learn various simulated multi-agent herding tasks. Three experiments were conducted that (i) first modeled human team performance during a set of novel herding tasks adapted from [19, 21], (ii) tested an AA utilizing this model to complete the tasks with human novices, and (iii) demonstrated how this AA could effectively train novices in a manner comparable to a human-expert trainer.