Motor Skill Learning by Virtual Co-embodiment with an AI Teacher Trained in Human Teaching Behavior

Virtual reality is gaining attention as a tool to facilitate motor skill learning. Numerous studies have been conducted on virtual co-embodiment, in which the movements of the teacher and student are weighted and averaged into a single avatar for motor skill learning. Previous studies have shown that virtual co-embodiment with a human teacher enhances motor skill learning efficiency, and the behavior of the human teacher is important for effective learning. However, this system has some challenges, such as the human playing the teacher’s role must be skilled in teaching and using virtual co-embodiment, and the teacher can be adversely affected by the learner. To solve these problems, we created an AI teacher using long short-term memory, which outputs the behavior of the teacher based on the input of the learner’s behavior data and the state of the experimental environment and trained the AI teacher by supervised learning using behavior training data of a human virtual co-embodiment. We confirmed that this AI teacher can generate behaviors similar to those of the human teacher and investigated the efficiency of motor skill learning using virtual co-embodiment with the AI teacher. Co-embodiment with the AI teacher reduced performance during the learning phase but improved performance during subsequent independent task execution. We further analyzed the assist proportion of the teacher and observed that when the co-embodied with an AI teacher, the assist proportion is lower than when the co-embodied with a human, suggesting that learning efficiency may be enhanced when the co-embodied partner is a mixture of supportive and obstructive.