Visual Servoing Control Approach for Tendon-driven Continuum Joints

Kurzfassung

Bioinspired soft robots have increasingly garnered attention in the academic world due to their various advantages compared to rigid robots, such as ameliorated efficiency, compliance and agility. At the same time, the technological utility of another scientific field, called visual servoing (VS), has become more apparent with recent progress in computer vision. In attempts to link these two disciplines with each other, several methods are proposed to control the end effector of a soft robot using visual sensor information. However, to the best of the authors knowledge, no VS control approach has been described for a tendon-driven continuum joint (TDCJ) representing the neck of a humanoid robot. This thesis addresses the adaptation and implementation of a model-based VS approach onto a human neck-inspired TDCJ in an eye-in-hand setup. Within that scope, an object detection model for AprilTags is implemented with OpenCV providing the controller with one marker to track. Having linked image-based visual servoing (IBVS) with the kinematic model of the continuum, the overall system Jacobian projects the control errors from the image to motor velocities, which are integrated to positional motor commands. As part of the research, the underlying pseudo-inverse model of the controller is examined in simulative unit tests in terms of feasibility. The control is implemented on a real tendon-driven continuum platform and is able to stabilize the closed loop using different combinations of controller gains. To find feasible controller gain combinations, a gain tuning methodology is proposed based on pre-defined key performance indicators (KPIs), which monitor the achieved closed loop performance. The experimental investigations comprise step excitations and trajectory tracking objectives occuring inside a circle within the image space. Since the set of feasible gains corresponding to a satisfactory performance does not overlap for both step excitations and trajectory tracking, the controller is extended with a feedforward velocity estimation of the observed object. As a result, the performance of the controller is improved and a single set of gains is obtained, complying with the specified KPIs. Besides the successful validation of the feedforward-extended controller, the experimental results indicate for all controllers direction-dependent performance differences in the image space, which can be narrowed to not well-identified model parameters.