Learning-based control and sensing for soft robotic systems

Abstract

The recent emergence of soft robotics has led to exciting developments in the robotics field. Soft robots are inherently compliant and manoeuvrable manipulators that can passively adapt to their environment - traits previously unseen in rigid-link robots. Researchers saw potential applications in areas such as minimally invasive surgery, rehabilitation, search and rescue, and underwater robotics. However, as with any emergent field, many new challenges arise in the investigation of their unique traits and how to best apply them. This thesis aims to investigate methods to tackle challenges specifically in the control and sensing of soft and flexible robotic systems. Soft robots excel in unstructured and unpredictable environments, in thanks to their compliant and adaptable nature, however, this often comes at the cost of precise control and predictability of the system. In order to fully make use of their unique properties, accurate control should still be maintained while affected by the environment. Commonly used model-based approaches often have low tolerance to unmodelled loading, resulting in significant error when acted on by them. Therefore, in this first study we employ a nonparametric learning-based method that can approximate and update the inverse model of a redundant two-segment soft robot in an online manner. The primary contribution of this work is the application and evaluation of the proposed framework on a redundant soft robot. With the addition of redundancy, a constrained optimisation approach is taken to consistently resolve null-space behaviour. Through this control framework, the controller can continuously adapt to unknown external disturbances during runtime and maintain end-effector accuracy. The performance of the control framework was evaluated by tracking of a 3D trajectory with a static tip load, and a variable weight tip load. The results indicate that the proposed controller could effectively adapt to the disturbances and continue to track the desired trajectory accurately. In a following study, a general framework for the design, fabrication, and modelling of soft and flexible surface shape sensor was presented. A prototype surface shape sensor using a single FBG optical fibre was developed and its performance experimentally validated. The single-core optical fibre with fibre Bragg gratings (FBGs) is capable of detecting sparse local strains at high bandwidth using the wavelength-division multiplexing (WDM) interrogation method. The fibre was embedded into an elastomeric substrate to reconstruct its global surface morphology. Finite element analysis (FEA) was used to determine the design parameters and to validate the unique mapping from sparse strain measurements to the continuum shape of the sensor. To simplify the fabrication and error compensation process without precise/prior knowledge of the FBG locations in the sensor, machine learning-based modelling was applied. This enables real-time, robust and reliable shape reconstruction. It is demonstrated to outperform various applications of electronics-based sensors which require sophisticated electrode wiring and calibration. Experiments were performed to evaluate the sensing accuracy and repeatability. In a final study, 3D shape reconstruction of a multi-core optical fibre was conducted. Preliminary experimental evaluation of bending curvature and bending direction was performed.