Research

Fox sisters: Tale of the Nine Tails

While anthropomorphic animals from folklore are among the most popular creatures in fantasy movies, computer games, and extended reality, visualizations of their behavior rely mostly on animators’ imagination and the direct observation of real animals.  The artist-driven creation describes emotional characteristics, but this method is limited in terms of discovering diverse unseen behaviors that lie beyond creators’ premade conception.  To promote the verisimilitude of imaginary animals, the present study adopts a reinforcement learning (RL) agent simulation. This study investigates its effects on the autonomous behavioral generation and serendipitous responses of the nine-tailed Fox Sister, an anthropomorphic animal in Korean folklore. The developed agent is based on a physics-based controller with 43 skeletal joints. 

To compute autonomous behavior transitions, we trained two different bipedal- and quadrupedal-walking actions to the fox-sister agent in one training sequence. Under the reward function, the given reward and penalty nudged the agent to decide the action depending on the contact of the body part with the physical environment. In several training steps, the Fox Sister agent also generates multiple unexpected behaviors, for instance, a forward roll during a quadrupedal walk responding to the given slopes. Additionally, the RL agent simulation enables the discovery of unknown functions of body components, for instance, using the nine tails to control body balancing in bipedal behaviors. This study addresses RL’s visualization of varying autonomous behaviors of imaginary folktale animals by simulating hypothetical relationships between body structures, motion generation, and environment. This RL simulation-driven character design can inspire creators' imagination and elevate emotional arousal among audiences.

Recreated digital model  of the Fox Sister

Training Process

Training results 1 : behaviour transition