SoftTex: Soft Robotic Arm With Learning-Based Textile Proprioception

Abstract

Soft robots are promising in biomedical applications thanks to their inherent structural compliance and distributed large deformations. However, integrating a sensory system that maintains the robot's dexterity while offering accurate state estimation remains an open challenge for their widespread adoption. This letter presents SoftTex, a small-scale soft robotic arm built with textile fabrics. SoftTex redesigns the STIFF-FLOP soft manipulator to favor affordable, rapid, and repeatable fabrication methods, facilitating the integration of a proprioceptive system based on piezoresistive textile strips preserving the soft arm compliance. First, we characterized the bending and stretching capabilities of the soft robotic arm and its workspace. The force tests demonstrated effectiveness for potential biomedical applications, revealing pulling forces ranging from 3.4–7.4 N and pushing forces from 2–7.5 N. Finally, we leveraged actuation, motion, and proprioceptive data collected with an open-loop controller to develop a position state estimator using a parallel recurrent neural network trained with supervised curriculum learning. The proprioceptive network achieves an average prediction error of ${\text{2.0}} \pm {\text{1.8}}$ mm ($3.4 \pm 2.9\%L$, where $L$ is module length). The findings are promising for closed-loop control, addressing the demand for low-cost, sensor-equipped soft robotic arms in the medical field and enhancing their potential for confined space exploration.