No abstract

Actuators using soft materials feature a large number of degrees of freedom. This tremendous flexibility allows a soft actuator to passively adapt its shape to the objects under interaction. In this paper we propose a novel proprioception method for soft actuators during real-time interaction with priorly unknown objects. Firstly, we design a color-based sensing structure that instantly translates the inflation of a bellow into changes in color, which are subsequently detected by a miniaturized color sensor. The color sensor is small and thus multiple of them can be integrated into soft pneumatic actuators to reflect local deformations. Secondly, we make use of a Feedforward Neural Network (FNN) to reconstruct a multivariate global shape deformation from local color signals. Our results demonstrate that deformations of the actuator during interaction, including sigmoid-like shapes, can be accurately reconstructed. The accurate shape sensing represents a significant step towards closed-loop control of soft robots in unstructured environments.