The effect of noise injection method on DRL-based controller robustness of human gait model

Human gait models in the dynamic environment have been studied to understand the fundamental neuronal control system of the human. Previously, imitation learning methods with reinforcement learning could successfully reproduce human gait motions using skeletal models. However, imitation learning-based controllers could have lacks of the ability to flexibly adapt to a wide range of state spatial scenarios such as instabilities and falling. Recently, noise injection methods have been introduced to increase the flexibility and robustness of the controller for robot systems. Therefore, the objectives of this study were to train human gait controllers with the noise injection method and to analyze the effect of noise injections on the balance recovery against external forces. A three-dimensional skeletal human gait model and two gait controllers with and without noise injections were developed. The robustness of the gait controllers against external forces was tested via forward dynamics-based gait simulation. The number of simulations without falling against external forces increased when the gait controller was trained with the noise injection method. In this study, the noise injection method during imitation learning could enhance the robustness and stability of the human gait controller.