On Human to Robot Skill Transfer for Robust Grasping and Dexterous Manipulation

Abstract

With increasing rates of automation in industry, service, and home environments, robots are required to perform progressively more sophisticated tasks and rapidly adapt to dynamic environments. Many of these tasks are simple and intuitive for humans, but have proven to be incredibly difficult to reliably implement on a robot system, particularly in the context of grasping and dexterous manipulation. The issue of efficiently transferring human skill to robot systems is therefore becoming increasingly more important, with approaches ranging from traditional robot programming to task learning with minimal human guidance. This thesis progresses through methods that require varying degrees of human involvement, proposing and evaluating approaches that facilitate manual robot teaching, teleoperation, programming by demonstration, high-level process supervision, and crowd participation. Beginning with manual robot teaching, an open-source, generic robot teaching interface is proposed and compared with alternative devices in terms of usability and efficiency. The work then focuses on human to robot motion mapping, introducing methods that enhance robot teleoperation through intuitive motion capture, interface design, mapping, and control. On a higher level of system autonomy, a method of enhancing programming by demonstration is proposed, utilising path optimisation and local replanning to allow for efficient teaching and execution of assembly tasks. The work then advances to flexible robotic assembly that requires minimal human involvement, proposing a framework that relies on compliance control, CAD based localisation, and a multi-modal gripper to facilitate rapid adaptation to different task requirements. Finally, the thesis proposes a framework that leverages human perception by combining crowdsourcing and gamification, employing it to enhance the grasping and manipulation capabilities of assistive and autonomous robotic platforms. To evaluate the efficiency of the developed methods, numerous experiments with different robot systems in both structured and dynamic environments have been conducted.