Abstract ▼ AI-Helper

In this paper, we propose deep learning-based force control of a twisted string actuator. A twisted string actuator is difficult to model-based control because of the nonlinear relationship between the angle of twisting and the linear contractile force of the string. The deep learning-based force co...

In this paper, we propose deep learning-based force control of a twisted string actuator. A twisted string actuator is difficult to model-based control because of the nonlinear relationship between the angle of twisting and the linear contractile force of the string. The deep learning-based force control method can implement a force controller with high tracking performance through the input and output of the twisted string actuator without model identification of the twisted string actuator. Among the learning-based control methods, a direct control method using an inverse model was used. The input and output of the inverse model of the twisted string actuator consist of the contractile force, the angle of twisting, and the voltage of the electric motor. In order to learn the nonlinearity of the twisted string actuator and implement accurate force tracking performance, the inverse model structure of the twisted string actuator was designed as a deep neural network structure. Training of the deep neural network inverse model was performed offline, and the output performance of the trained deep neural network inverse model was evaluated with test data. We propose a control framework for implementing deep learning-based force control of the twisted string actuator using a trained inverse model. We introduce a hardware system of the twisted string actuator to implement deep learning-based force control and a method for collecting training data in an experimental environment. The proposed deep learning-based force controller showed better force control performance than the PI controller through experiments.

Keyword