단순 물리 모델을 사용한 역 최적화 기반 자연스러운 캐릭터 동작 생성

Abstract

This thesis introduces an inverse optimization technique to construct a motion controller that controls a human character according to a continuous user command. In addition, the controller generates human-like natural motion while satisfying constraints subject to the type of the motion. The proposed method consists of three components: simplified physical models, motion capture data, and optimization. We assume that the motion capture data contains natural character motion patterns and the characteristics of a human. The physical model is constructed to satisfy its physical constraints such as the relative distance and angle in two character interaction, interpenetration in object manipulation, footstep changes in locomotion, or character balance. Optimization procedure matches the motion pattern between the simplified physical models and the motion pattern of the motion capture data. The cost function of the optimization is to minimize the difference of the motion patterns between the model and the motion capture. Based on the optimization, the resultant motion controller can be constructed including the controllability originated from the simplified physical model and the naturalness inherited from the motion capture data, which we call inverse optimization. To reconstruct the relatively high dimensional full-body character motion from the low dimensional simplified physical models, relative information between the physical model and the character is extracted and stored. The stored relative information is selected at every frame during motion synthesis. To validate the inverse optimization technique, we apply the proposed method to the various motion patterns such as two character interaction, object manipulation, locomotion, and balancing motion. In addition, We emphasize that the different constraints should be considered in motion generation procedure, so as to not lose the naturalness of the generated motion according to the types of the motions. For two character interaction, we emphasize continuous interaction in order to enhance the quality of the generated motion. To maintain the continuous interaction, two consistencies, spatio-temporal and the physical consistencies, are considered. The spatio-temporal consistency is related to keep the relative distance and angle between two characters, and the physical consistency indicates to keep the balance of each character. We develop a new type of simplified physical model, coupled inverted pendulum on carts (CIPC), to consider the two consistencies in two character interaction. Optimization on the CIPC model extracts the relation between two interacting characters, which allows maintaining interaction qualities similar to the motion capture data. For object manipulation, we introduce the spatio-temporal constraints in performance-based animation to generate natural manipulation motion. The spatial constraint indicates the interpenetration between the character’s hand and the object. The temporal constraint is related to the timing of grasping the virtual object. To consider these constraints, we define multiple particle models that represent the hand motion. By training the particle models with motion capture data of object manipulation, the motion controller generalizes the motion capture data to generate object grasping motion without the interpenetration between the character hand and the object even when the location of the object is different from the motion capture data. For locomotion, inversely learned inverted pendulum model (ILIPM) is developed as a simplified physical model to handle the footstep strategy while generating locomotion. ILIPM learns the motion patterns of the motion capture data and the footstep strategies that consist of footstep speed and location. The learned motion and footstep vary according to the changes of the situation such as a user control, an external force, or obstacles. Moreover, regressors are devised to estimate the adequate model control signal that controls the motion of the ILIPM. For balancing motion, double inverted pendulum model (DIPM) is devised by composing an additional pendulum to represent the pelvis balancing strategy.We perform the optimization to learn the human balancing motion while applying the external forces to the physical model. The contributions of this thesis are mainly the development of simplified physical models considering the constraints according to the types of the motions. In addition, the various optimization procedures are constructed considering the physical models.