Joint State Estimation for Interlocked Rigid Segments-Based Continuum Robots

This thesis investigates shape estimation in rigid segment-based continuum robots using magnetic sensors, focusing on applications in minimally invasive surgery. The goal is to determine joint angles in real time through Hall-effect sensors, eliminating the need for fluoroscopy and providing a more efficient and precise localization method. The study models the mechanical play of the ball joint, explores different approaches, including a lookup table (LUT) method that uses precomputed magnetic field values and machine learning models, specifically multi-layer perceptrons (MLP) and convolutional neural networks (CNN), to improve accuracy and lastly expands the set up from 1 Degree of Freedom to 2. The experimental setup consists of a magnetized ball joint structure where the male part is magnetized thanks to a coil, and the female part remains non-magnetic. A Hall sensor on the female joint measures the three-dimensional magnetic field variations, allowing for angle estimation. A motorized calibration system with piezo actuators ensures precise sensor positioning. Results show that the MLP approach outperforms both the LUT method and CNNs, achieving the lowest mean squared error (1.1°), while CNNs performed worse. The system also demonstrates robustness to external magnetic interference, ensuring reliable performance even if coupled with magnetic actuation. However, challenges remain, particularly regarding mechanical play or misalignment in the set up, which introduce estimation errors. The findings of this thesis contribute to the field of robotic catheter navigation, offering a viable alternative to fluoroscopy-based tracking systems and paving the way for more autonomous medical robotics.