Learning-based Methods for Concentric Tube Continuum Robots Modeling and Shape Estimation
University of Toronto, 2021
This thesis addresses the modeling and sensing of concentric tube continuum robots from a data-driven perspective. Firstly, the shape-to-joint inverse kinematics problem for concentric tube continuum robots is investigated, by both data-driven and numerical-analysis approaches for the first time. Particularly for the learning-based method, a neural network is proposed to approximate the mapping from robot shape represented by up to 20 discrete points to its joint parameters, with errors of 2.22° and 1.45 mm obtained. This method is an enabler for future shape control algorithms. Secondly, we propose a learning-based method for estimating the robot shape from multi-view images. Using a convolutional neural network and end-to-end learning, this method enables a direct mapping from images of the robot taken from multi-views to robot shape or tip position, with sub-millimetre accuracy achieved in simulation. This allows for a calibration-free, markerless, and minimum image-processing 3d shape reconstruction of concentric tube continuum robots.