WaveLeNet: Transfer Neural Calibration for Embedded Sensing in Soft Robots

Soft robots have exhibited excellent compatibility with functional and physical requirements of intraluminal procedures such as bronchoscopy and cardiovascular intervention [1]. Despite their favourable mechanical compliance and scalable design, integrating miniature force and shape sensors on them is cumbersome [2]. Also, large mechanical deformation of such robots, i.e., flexures, may push traditional rigid sensors out of their linear range [3]. As an alternative approach, the authors have recently introduced a novel soft sensing method and soft embedded sensors for flexures that exhibited less than 10mN error in measuring external 3D tip forces on soft robots for bronchoscopy and cardiovas- cular applications [4], [5], [6]. Fig. 1(a –c) depict the conceptual design, the prototyped sensor developed in [5], and a representative interventional application. Their soft sensor was comprised of a gelatin-based matrix filled with graphite nano-particles that exhibited stable piezoresistivity under extremely large deformation. De- spite its accuracy, the accuracy of the proposed sensor was adversely affected in noisy environments, e.g., op- eration rooms. The reason was that the rate-dependent features used in its neural calibration would amplify the peripheral noise which would diminish the accuracy. In this study, we have proposed and validated an alterna- tive deep-learning-based method for calibration of the proposed soft sensor that is derivative-free thus does not amplify the peripheral noise and is versatile. Con- ceptually, the proposed calibration methods can be used to assemble an array of sensor readings for distributed sensing on soft robots. Our proposed method is based on generating a scalogram from the temporal-frequency content of the measured voltages using real-time wavelet transform and using transfer learning technique to infer rate-dependent and deformation-dependent features from the voltages’ scalogram.