Sensorimotor learning can be defined as a process by which an organism benefits from its experience, such that its future behaviour is better adapted to its environment. Humans are sensorimotor learners par excellence, and neurologically intact adults possess an incredible repertoire of skilled behaviours. Nevertheless, despite the topic fascinating scientists for centuries, there remains a lack of understanding about how humans truly learn. There is a need to better understand sensorimotor learning mechanisms in order to develop treatments for individuals with movement problems, improve training regimes (e.g. surgery) and accelerate motor learning in tasks such as handwriting in children and stroke rehabilitation. This thesis set out to improve our understanding of sensorimotor learning processes and develop methodologies and tools that enable other scientists to tackle these research questions using the power of recent developments in computer science (particularly immersive technologies). Errors in sensorimotor learning are the specific focus of the experimental chapters of this thesis, where the goal is to address our understanding of error perception and correction in motor learning and provide a computational understanding of how we process different types of error to inform subsequent behaviour. A brief summary of the approaches employed, and tools developed over the course of this thesis are presented below.
Chapter 1 of this thesis provides a concise overview of the literature on human sensorimotor learning. It introduces the concept of internal models of human interactions with the environment, constructed and refined by the brain in the learning process. Highlighted in this chapter are potential mechanisms for promoting learning (e.g. error augmentation, motor variability) and outstanding challenges for the field (e.g. redundancy, credit assignment).
In Chapter 2 a computational model based on information acquisition is developed. The model suggests that disruptive forces applied to human movements during training could improve learning because they allow the learner to sample more information from their environment. Chapter 3 investigates whether sensorimotor learning can be accelerated through forcing participants to explore (and thus acquire more information) a novel workspace. The results imply that exploration may be a necessary component of learning but manipulating it in this way is not sufficient to accelerate learning. This work serves to highlight the critical role of error correction in learning.
The process of conducting the experimental work in Chapters 2 and 3 highlighted the need for an application programme interface that would allow researchers to rapidly deploy experiments that allow one to examine learning in a controlled but ecologically relevant manner. Virtual reality systems (that measure human interactions with computer generated worlds) provide a powerful tool for exploring sensorimotor learning and their use in the study of human behaviour is now more feasible due to recent technological advances. To this end, Chapter 4 reports the development of the Unity Experiment Framework - a new tool to assist in the development of virtual reality experiments in the Unity game engine.
Chapter 5 builds on the findings from Chapters 2 & 3 on learning by addressing the specific contributions of visual error. It utilises the Unity Experiment Framework to explore whether visually increasing the error signal in a novel aiming task can accelerate motor learning. A novel aiming task is developed which requires participants to learn the mapping between rotations of the handheld virtual reality controllers and the movement of a cursor in Cartesian space. The results show that the visual disturbance does not accelerate the learning of skilled movements, implying a crucial role for mechanical forces, or physical error correction, which is consistent with the findings reported in Chapter 2. Uncontrolled manifold analysis provides insight into how the variability in selected solutions related to learning and performance, as the task deliberately allowed a variety of solutions from a redundant parameter space.
Chapter 6 extends the scope of this thesis by examining how error information from the sensorimotor system influences higher order action selection processes. Chapter 5 highlighted the loose definition of “error” in sensorimotor learning and here, the goal was to advance our understanding of error learning by discriminating between different sources of error to better understand their contributions to future behaviour. This issue is illustrated through the example of a tennis player who, on a given point, has the options of selecting a backhand or forehand shot available to her. If the shot is ineffective (and produces an error signal), to optimise future behaviour, the brain needs to rapidly determine whether the error was due to poor shot selection, or whether the correct shot was selected but just poorly executed.
To examine these questions, a novel ‘action bandit’ task was developed where participants made reaching movements towards targets, with each target having distinct probabilities of execution and selection error. The results revealed a significant selection bias towards a target that produced a higher frequency of execution errors (rather than a target associated with more selection error) despite no difference in expected value. This behaviour may be explained by a gating mechanism, where learning from the lack of reward is discounted following sensorimotor errors. However, execution errors also increase uncertainty about the appropriateness of a selected choice and the need to reduce uncertainty could equally account for these results. Subsequent experiments test these competing hypotheses and show this putative gating mechanism can be dynamically regulated though coupling of selections and execution errors. Development of models of these processes highlighted the dynamics of the mechanisms that drive the behaviour. In Chapter 7, the motor component of the task was removed to examine whether this effect is not unique to execution errors, but a feature of any two-stage decision-making process with, multiple error types which are presumed to be dissociated. These observations highlight the complex role error plays in learning and suggest the credit assignment process is guided and modulated by internal models of the task at hand.
Finally, Chapter 8 closes this thesis with a summary of the key findings and arising from this work in the context of the literature on motor learning and decision making.
It is noted here that this thesis sought to cover two broad research topics of motor learning and decision making that have, until recently, been studied by separate groups of researchers, with very little overlap in literature. A key goal of this programme of research was to contribute towards bringing together these hitherto disparate fields by focussing on breadth to establish common ground. As the experimental work developed, it became clear that the processing of error required a multi-pronged approach. Within each experimental chapter, the focus on error was accordingly narrowed and definitions refined. This culminated in developing and testing how individuals discriminate between errors in the sensorimotor and cognitive domains, thus presenting a framework for understanding how motor learning and decision making interact.
