Supervised by: Peter Corke Niko Sünderhauf Feras Dayoub

Awards

Outstanding Doctoral Thesis Award for 2023 The decision-making process, undertaken by QUT's Research Degrees Committee, placed this thesis in the top 5% of successful doctoral candidates for 2022. This award is presented in recognition of the outstanding contribution made to your chosen discipline and the standard of excellence demonstrated in higher degree research practice.

Siganto Foundation Medal 2023 The Siganto medal is annually awarded to one distinguished PhD graduate from the QUT Science and Engineering Faculties to support and encourage multidisciplinary research.

Abstract

The ultimate goal for robots - particularly in the manipulation field - is to operate reliably and robustly in real-world environments. For example, such environments could be a home, an office, a factory, or a field. These environments are dynamic, unstructured, and unknowable.

Consequently, a robot's controller must be closed-loop and reactive to environmental and task changes. When choosing an action to take, the controller must re-consider the state of the robot, environment, and goal and, if necessary, take corrective action to ensure successful and safe task completion. However, motion planning is the prevailing technique to drive a robot to a goal. Contradictory to the requirements outlined previously, the planning time associated with motion planners precludes their use as reactive controllers. This thesis addresses this gap and contributes to the reactive manipulation field through two avenues.

Firstly, we provide several advances to state-of-the-art reactive motion control. Our proposed controllers are closed-loop, meaning they can react to environmental changes to provide robust task execution in real-world dynamic environments. Using these controllers, we can maximise the manipulability of a manipulator - this improves the manipulator's ability to move in any direction, avoid joint position and velocity limits, dodge both static and moving obstacles, and operate on non-redundant, redundant, and mobile manipulators. These contributions bridge the capability gap between motion planners and reactive motion controllers and enable fullfeatured reactive motion control.

Secondly, we explain the underlying principles and provide the necessary theoretical and educational tools to make it trivial to construct a robot model, create a reactive motion controller, or formulate a numerical inverse kinematics solver. From this point, a roboticist can improve and extend algorithms to create new and innovative approaches, advancing the field. Additionally, we present an extensive open-source robotics ecosystem for the Python language called Python Robotics. The ecosystem provides fundamental robotics capabilities, including manipulator kinematics and dynamics, differential kinematics, robot modelling, control, mobile robotics, simulation, and spatial mathematics. The Toolbox enables education in key robotics algorithms and provides the tools necessary for innovation to the state-of-the-art.