Data-driven control strategies for precise manipulation of continuum robots

Abstract

Continuum robots are designed to inherently possess dexterity and adaptability, as bio-inspired by elephant trunks and octopus tentacles. This compliance with surroundings endows them the advantage of safe operation in confined space, even in medical applications. However, the tradeoff between manipulator flexibility and precise control for the infinite degrees of freedom requires particular approaches concerning modeling and sensing. Different from conventional rigid-link robots whose kinematics/dynamics could be certainly determined, modeling uncertainties due to intrinsic and extrinsic factors would apparently deteriorate the model-based control performances. Under such circumstances, data-driven strategies making use of learning algorithms would be a promising way out for continuum robot control.

The main focus of this thesis is to propose appropriate control schemes and sensing modalities for continuum robots, utilizing data-driven algorithms to resolve or enhance the precise manipulation under various environments. Visual servoing control is a significant focus of this thesis, where the self-contained camera on the robot end-effector can provide intuitive perception and positional feedback. A novel visual servoing framework utilizing localized online learning enables stable path following in the camera view, even under external forces and varying tip loading. Optical fiber inscribed with fiber Bragg gratings (FBGs) can enhance the feature tracking accuracy in the 2D visual servoing. Integrated with the flexible fiber, the vision-based sensing can possess higher reliability under poor or even extreme visual conditions (e.g., full shielding, absolute darkness). The proposed online learning-based sensing fusion of camera and FBG measurements can also provide robust 6D pose estimation of the end-effector, expanding the task space filed for continuum robot control. Referring to the convergence guaranteed by kinematic model-based control and the accommodation accredit to learning-based control, a hybrid controller is designed to incorporate their strengths. A kinematic model without the need for fine tuning and an online-updated error compensator can increase the control accuracy, as well as circumvent prior data exploration and training at the same time. The assistance of FBG is also investigated on shape tracking and learning-based modeling of slim continuum robots, which are validated on a robotic catheterization platform. Continuum robot visual servoing has excellent potential in robotic endoscopy, while the optical fiber’s compatibility to magnetic resonance imaging (MRI) also indicates the future clinical application of the proposed shape sensing and control.