Generating Various 3D Motions by Emergent Imitation Learning

This paper proposed emergent imitation learning (EIL) using deep reinforcement learning (deep RL) to generate various 3D motions for a character agent in a virtual 3D physical space. In the conventional method, agents acquire a motion by learning to imitate a predefined reference motion as a teacher motion and then achieve a specific task. In this study, we introduce a "generational change" to the conventional method to elicit a motion derived from a reference motion. When a generational change occurs, new agents begin learning using the current agent's motion as a teacher. Experimental results showed that the motions derived using the proposed method differ from the teacher's motions and are more diverse, and the task achievement rating is higher than those obtained by a conventional method.