Robot manipulator skill learning and generalising through teleoperation

Robot manipulators have been widely used for simple repetitive, and accurate tasks in industrial plants, such as pick and place, assembly and welding etc., but it is still hard to deploy in human-centred environments for dexterous manipulation tasks, such as medical examination and robot-assisted healthcare. These tasks are not only related to motion planning and control but also to the compliant interaction behaviour of robots, e.g. motion control, force regulation and impedance adaptation simultaneously under dynamic and unknown environments. Recently, with the development of collaborative robotics (cobots) and machine learning, robot skill learning and generalising have attained increasing attention from robotics, machine learning and neuroscience communities. Nevertheless, learning complex and compliant manipulation skills, such as manipulating deformable objects, scanning the human body and folding clothes, is still challenging for robots. On the other hand, teleoperation, also namely remote operation or telerobotics, has been an old research area since 1950, and there have been a number of applications such as space exploration, telemedicine, marine vehicles and emergency response etc. One of its advantages is to combine the precise control of robots with human intelligence to perform dexterous and safety-critical tasks from a distance. In addition, telepresence allows remote operators could feel the actual interaction between the robot and the environment, including the vision, sound and haptic feedback etc. Especially under the development of various augmented reality (AR), virtual reality (VR) and wearable devices, intuitive and immersive teleoperation have received increasing attention from robotics and computer science communities. Thus, various human-robot collaboration (HRC) interfaces based on the above technologies were developed to integrate robot control and telemanipulation by human operators for robot skills learning from human beings. In this context, robot skill learning could benefit teleoperation by automating repetitive and tedious tasks, and teleoperation demonstration and interaction by human teachers also allow the robot to learn progressively and interactively. Therefore, in this dissertation, we study human-robot skill transfer and generalising through intuitive teleoperation interfaces for contact-rich manipulation tasks, including medical examination, manipulating deformable objects, grasping soft objects and composite layup in manufacturing. The introduction, motivation and objectives of this thesis are introduced in Chapter 1. In Chapter 2, a literature review on manipulation skills acquisition through teleoperation is carried out, and the motivation and objectives of this thesis are discussed subsequently. Overall, the main contents of this thesis have three parts: Part 1 (Chapter 3) introduces the development and controller design of teleoperation systems with multimodal feedback, which is the foundation of this project for robot learning from human demonstration and interaction. In Part 2 (Chapters 4, 5, 6 and 7), we studied primitive skill library theory, behaviour tree-based modular method, and perception-enhanced method to improve the generalisation capability of learning from the human demonstrations. And several applications were employed to evaluate the effectiveness of these methods.In Part 3 (Chapter 8), we studied the deep multimodal neural networks to encode the manipulation skill, especially the multimodal perception information. This part conducted physical experiments on robot-assisted ultrasound scanning applications.Chapter 9 summarises the contributions and potential directions of this thesis. Keywords: Learning from demonstration; Teleoperation; Multimodal interface; Human-in-the-loop; Compliant control; Human-robot interaction; Robot-assisted sonography