Learning Finite-State Machine Controllers From Motion Capture Data

With characters in computer games and interactive media increasingly being based on real actors, the individuality of an actor´s performance should not only be reflected in the appearance and animation of the character but also in the AI that governs the character´s behavior and interactions with the environment. Machine learning methods applied to motion capture data provide a way of doing this. This paper presents a method for learning the parameters of a finite-state machine (FSM) controller. The method learns both the transition probabilities of the FSM and also how to select animations based on the current state.