IEEE Transactions on Medical Robotics and Bionics , 2 (4), pp. 619-630, 2020
Concentric Tube Continuum Robots are among the smallest and most flexible instruments in development for minimally invasive surgery, thereby enabling operations in areas within the human body that are difficult to reach. Unfortunately, integrating state-of-the-art force sensors is challenging for these robots due to their small form factor, although contact forces are essential information in surgical procedures. In this work, we propose a novel data-driven approach based on Deep Direct Cascade Learning (DDCL) to create a virtual sensor for computing the tip contact force of Concentric Tube Continuum Robots. By exploiting the robot’s backbone’s inherent elasticity, deflection is used to estimate the respective external tip contact force. We evaluate our approach on different data representations for a single tube and apply it subsequently on a three-segment Concentric Tube Continuum Robot. Furthermore, we devise a novel transfer learning approach through DDCL to improve the estimation accuracy by pre-training a cascaded network with simulated data. Subsequently, we fine-tune the network based on a small real-world data set recorded from the physical robot.