Interactive Training for Synthetic Characters

Compelling synthetic characters must behave in ways that reflect their past experience and thus allow for individual personalization. We therefore need a method that allows characters to learn. But simply adding traditional machine learning algorithms without considering the characters’ own motivations and desires will break the illusion of life. Intentional characters require interactive learning. In this paper, we present the results of Sydney K9.0, a project based on the Synthetic Characters creature kernel framework. Inspired by pet training, we have implemented a character that can be trained using the “clicker training” technique. Clicker training utilizes the natural desires of an animal and employs operant conditioning procedures for shaping their behavior. The necessary plasticity of system interconnections shaped by associations and rewards that is required by clicker training was integrated into the creature kernel framework. The implemented system also includes a module named DogEar that is designed for collecting real-world acoustic data, such as human voice commands, integrated into the creature kernel’s perception system. This provides a seamless interface between the simulated and real worlds. Detailed implementation and interaction results are presented.