Shape Recognition of a Tensegrity With Soft Sensor Threads and Artificial Muscles Using a Recurrent Neural Network
Wen-Yung Li, Atsushi Takata, Hiroyuki Nabae, Gen Endo, Koichi Suzumori
Abstract
This letter proposes a novel method to accomplish shape recognition by utilizing a tensegrity structure with a soft sensor via a recurrent neural network (RNN). The combination of soft tensegrity and soft sensors make it capable of recognizing the deformation to reflect the shape of its surroundings. As the first step to this goal, we build a three-bar tensegrity prism with nine separate soft sensors in which the resistance value of the sensors changes with length variation. The prism is actuated by thin McKibben muscles and deforms when the pressure inside a muscle. The positions of the six nodes in the prism are obtained using a motion-capture system. The measured resistance and position data are used as training data for the RNN to build a prediction model that can reflect the shape variation during the period of deformation of the tensegrity. If several prisms are connected, they can be used with this approach to recognize the shape of a three-dimensional environment that is difficult to observe directly.