Generation of subject-specific motions at various walking speeds using meta reinforcement learning and generative adversarial networks

Abstract

Research on creating a controller that controls movement by modeling human behavior has continued. In particular, recently, research has been conducted in a wide range of fields, such as biomechanics, animation, and robots, such as research for pure dynamics simulation of musculoskeletal models, research to make character movements natural by replicating human behavior, and research to create controllers for bipedal robots. . However, it is still difficult to obtain a controller that can control not only one motion of a specific person, but also multiple motions. In particular, gait is one of the most common exercises by humans, but even for one person, it is difficult to obtain a controller that can reproduce gait according to various speeds of a specific person because the gait motion changes depending on the walking speed. If a personalized controller that can reproduce all gait according to a specific person's situation is obtained, biomechanical applications such as injury simulation according to walking speed or the ground environment, and animation such as virtual reality implementation using gait reproduction that captures the characteristics of a specific person field applications can be made. The purpose of this study is to develop a gait controller that can reflect the subject's characteristics and generate gait motions not only at measured gait speeds but also gait motions at unmeasured gait speeds when there is gait data of a specific person at multiple speeds. It is developed through artificial intelligence techniques. A generative adversarial neural network and reinforcement learning were used to reflect individual characteristics in the gait controller, and a meta-learning technique was used to quickly generate a gait controller for a new subject. The target walking speed information was set as one of the inputs of the gait controller to allow walking at various speeds. To confirm the performance of the learned gait controller, the kinematic data of the created motion and the kinematic data of the measured motion were compared. In addition, a classifier was learned through supervised learning in order to check whether the personalized gait controller was learned, and the result was checked whether the motions generated according to each subject were classified using the learned neural network.