Data-driven strategies for soft sensing, robot modelling and control

Abstract

Soft/continuum robots are developed with high dexterity, flexibility and adaptability, as inspired by biologies like octopus, elephant trunk and vines. This high compliance endows them with the advantage of safe interaction with their surroundings or humans, which indicates a wide spectrum of applications. However, unlike conventional rigid-link robots with identified kinematics/dynamics, high flexibility also brings great challenges for their sensing, modelling and control. The infinite degree-of-freedoms (DoFs) and high non-linearity make it hard to find solutions with conventional analytical methods. In such situations, data-driven strategies using learning algorithms are a promising way out to solve these concerns for soft/continuum robots. The main work of this thesis is to propose appropriate shape sensing approaches, kinematic modelling methods and control schemes for flexible continuum robots, utilizing data-driven algorithms to improve the robotic performance. To perform accurate proprioception for soft and flexible structures, a real-time shape sensing framework is proposed, which is applicable to various sensor shapes and sensing element types. Optical fiber inscribed with fiber Bragg gratings (FBGs) has been used to measure local strain changes of the soft robots, aiming to estimate the shape information of the flexible structures. The framework integrates finite element methods with learning-based approaches, requiring only sparsely distributed sensing elements, to predict dynamic and complicated morphology changes of the sensors or robots. This sensing strategy has been experimentally validated on continuum robots and flexible skin-like surface sensors. A soft robot with a camera attached on the tip has been designed, where the eye-in-hand camera can provide intuitive visual feedback. Integrated with a helically embedded FBG fiber, the shape estimation technology can improve the accuracy of feature tracking for visual servoing. By experimental validation, the vision-based sensing can be more reliable in poor or even extreme visual conditions (e.g., large-area shielding and complete darkness). Not limited to the tube-like soft robot, the data-driven sensing method has been validated on a thin A4-sized surface shape sensor, which enables high-frequency and accurate shape sensing with the aid of computational mechanics. In this sensing framework, a learning-based model is obtained using finite element (FE)-enriched data, which enables application-focused customization and production. Additionally, a two-segment continuum robot has been designed for precise and fast laser sweeping in a narrow environment, with a multi-core optical fiber in the central channel for real-time shape sensing feedback. A Koopman-based dynamic model has been built to enable linear representation of a non-linear system, utilizing a data-driven strategy to identify the system parameters. This enables the implementation of various closed-loop control strategies on the high-performance robotic laser sweeping systems.