Intelligent soft shape sensing from sparse to dense in real time

Abstract

Soft sensor, as a key component in wearable devices and soft robots, has been prevalent recently in various research fields such as smart prosthetics, surgical manipulators and intelligent robotic system. The inherent mechanical compliance and adaptability of soft sensor allow data collection over a deformed medium without any physical mismatches and motion artifacts commonly occurring in its conventional rigid counterparts. It opens up many new challenges and opportunities to enhance the sensing capability of human perception as well as robotic proprioception. The primary motivation of this thesis initiated from the challenge of designing a high-dimensional soft sensor, which is capable of sensing 3-D morphological changes. The research gap in modeling a soft shape sensor is addressed based on literature reviews on the recent advances of flexible electronics, optical-based sensing, as well as a machine learning method. Upon developing shape sensing prototypes, this thesis also investigates how to scale up the sensing area without compromising sensitivity and update frequency.

In the first study, a data-driven modeling approach was explored to reconstruct the morphological changes of a thin, A4-sized (210 × 297 × 1 mm) sensor comprising an optical fiber and silicone rubber. Strain responses in various sensor designs were simulated using finite element analysis (FEA) to safeguard 28 fiber Bragg gratings distributed along the single-core fiber. The simulated environment was further utilized to approximate sensor shape based on finite ground truths, where sparse 3-D positions were enriched to a denser nodes array used. An ensemble learning model that utilized fiber strains as inputs and nodal displacements as outputs were validated with 2.28 mm RMSe at 100 Hz. The hybrid modeling approach was further generalized to a shape sensing framework in a following study. A self-contained optical waveguide sensor was developed on top of optomechanical simulations. Light transmission from simple embedded light-emitting diodes (LEDs) and photodetectors (PDs) was analyzed in advance of modeling. The spatiotemporal characteristic in light intensity variations (input) and 3-D nodal displacements (output) was regarded in an autoregressive model. Underwater experiments demonstrated by a fish-liked shape sensor showcased enhanced prediction accuracy of RMSe 0.27 mm. Real-time shape detection was verified at a update frequency of 150 Hz.