Revealing internal mechanisms of motor learning of the human brain is one of a remarkable issue to make intelligent robot systems. Hence, research is being actively conducted in various fields such as neuroscience, robotics, and developmental psychology. According to the Piaget's cognitive development theory, human's abilities including motor skills such as motor babbling and imitating others action develop step by step, and studies of cognitive developmental psychology have been reporting empirical evidence supporting this theory. Since it is known that the human brain has a hierarchical structure in its functional connectivity, brain-like computational models such as hierarchical temporal memory (HTM) have been proposed. In this dissertation, we suggested various brain-like computational models representing human's developmental motor abilities as following Piaget's theory and conducted neurorobotic experiments in terms of cognitive developmental robotics. First, multiple motor behaviors were learned and reproduced on the NAO humanoid robot as suggesting a computational model with a hierarchical self-organizing map (SOM). It successfully implemented on the real-robot and resolved intersection issues with the hierarchical structure. Second, developmental dynamics of a recurrent neural network to learn multiple goal-directed motor behaviors were proposed and examined on the NAO humanoid robot, and a relationship with studies of infant developments was discussed. Third, a deep temporal convolutional neural network (TCNN) for recognizing human action based on body movement with detected pose data was proposed. As examining internal representations of trained model, a relationship with the motor system of the human brain was discussed.
본 연구에서는 발달로보틱스 및 신경로보틱스의 관점에서 사람의 운동 발달 과정을 인공신경망 모델을 이용하여 모델링하고 이를 휴머노이드 로봇에 적용하는 일련의 실험을 수행하였다. 첫 번 째 실험에서는 운동피질의 체성 감각지도의 구조를 모방하여 계층적으로 구성된 자기조직화 맵(SOFM) 기반의 인공신경망 모델을 제안하고, 이를 통해 휴머노이드 로봇이 운동 학습 능력을 갖도록 하였다. 그 다음으로, 재발 성 신경망 모델을 기반으로 한 인지발달로봇 실험을 수행하고 유아의 행동 및 운동 발달 과정과의 연관성을 논의하였다. 마지막으로, 컨벌루션 신경망을 기반으로 사람의 행동을 분류하는 모델을 학습하고, 학습된 특징들을 시각화하여 분석하였다. 이러한 일련의 연구를 통해 사람의 행동 발달 매커니즘에 대한 이해를 높이고, 이를 공학적으로 활용하기 위한 방안에 대해 전반적으로 논의를 하였다.
