Learning singularity avoidance from constrained motion

Student thesis: Doctoral ThesisDoctor of Philosophy


With the continual growth of technology and resulting shift towards the development of versatile and autonomous robots, robotic systems are able to perform increasingly more complex tasks. In turn, expanding their application to other fields. Programming by demonstration is particularly suitable in tasks which require expert knowledge. As novice users can proficiently demonstrate a task in which they are an expert with little or no understanding of the system being taught. The promise of introducing collaborative robots for automation outside traditional manufacturing settings is their usability by novice users, i.e., those potentially with domain-expertise but little knowledge of robotic engineering. It is envisaged that such users would teach task-oriented skills through demonstration, thereby avoiding the need for formal training in the techniques and concepts familiar to roboticists. It has long been established, for example, that differences in embodiment of the human musculoskeletal system and most robotic actuation systems mean that direct imitation of human motor behaviour is suboptimal for robots. With this in mind, when designing interfaces and approaches to the programming by demonstration of systems, there is a need to take account of the natural motor behaviour of novices, as well as, manipulability of the taught robot.
This dissertation explores the optimisation of redundancy in robots by learning task-oriented behaviour through programming by demonstration; for the transfer of demonstrated skills to robotic systems. The research is split over three stages. First, looking at how the constraint learning system copes with demonstration data, namely how it is processed with respect to data dimensionality. A new method is proposed to quickly handle the ever increasing complexity of data provided to systems. It works by reducing the constraint learning system's search space by applying gradient descent. This enables the system's use in learning tasks within shorter periods of time and of greater dimensionality. Now, with the system's propagated applicability to complex problems, the second stage looks at the output of the system. Specifically, how data is prepared for use by the taught robot. Thus, a method is developed to resolve redundancy in learnt task-oriented behaviour. It works by taking the robot's own structure into account subject to a learnt constraint and uses this information to avoid singularities through learnt manipulability maximisation. Finally, the third stage looks at improving how data can be extrapolated when teaching comes from human demonstrators. Thus, a method is proposed which takes stereotypical behaviours in natural human movements into account. Thereby, making use of assumptions on how the demonstrator resolves redundancy, which makes it easier to learn the constraints contained within a task. A pipeline is presented for transferring behaviour optimised for humans, such that these are adapted according to the robotic system's own embodiment. Various experiments are conducted including in the real world which demonstrate the system's applicability to different tasks such as in reaching objectives as well as closing drawers.
Date of Award1 Jan 2021
Original languageEnglish
Awarding Institution
  • King's College London
SupervisorMatthew Howard (Supervisor)

Cite this