Patients with damage to the cerebellum make reaching movements that are uncoordinated, or ataxic — movements can be misdirected, over- or under-shoot targets, and more variable than the movements of healthy people. In addition, it is thought that cerebellar patients cannot adapt to (learn) new dynamics. This dissertation explores models of cerebellar ataxia and the use of robotic tools to improve human motor performance. We began by developing a dynamic model of the human arm and a robot exoskeleton. This model was populated with parameter values obtained through direct measurement, system identification, and use of Dempster's anthropometric table [1]. This system was then used to conduct three studies.

In the first study, we used an exoskeleton robot to record the movements of cerebellar patients performing a targeted reaching task. These data and our dynamic model were used to test hypotheses proposed in the literature about the role of the cerebellum. We found evidence supporting the general hypothesis that the cerebellum functions as an internal model for planning movements and that damage to the cerebellum results in a biased (inaccurate) model, although the model does not completely explain cerebellar behavior. Computer simulations found optimal, patient-specific perturbations to actual limb dynamics that are predicted to reduce root-mean-squared directional reaching errors by an average of 41%.

In the second study, we used the same robot to implement a change in dynamics predicted to assist each patient. This required a controller design, the addition of accelerometers to the robot, and characterization of the robot's ability to render acceleration-dependent forces. We found that this approach may improve the reaching of some cerebellar patients and not for others, and there was no evidence of motor learning.

In the third study, we investigated how a different assistive controller can be used to improve reaching behavior. We used the robot to render force channels to constrain arm movements during reaching, and explored how the ordering of null (robot passive) and channel (robot active) trial blocks affected motor learning. We found that force channels resulted in significant improvements in reaching performance in the directions both parallel and orthogonal to the force channel. However, we found no evidence that reaching practice in these channels results in improved reaching performance after the channels are removed.

The use of tools from the field of robotics allows precise measurements and interventions to understand and treat human motor deficits. We believe that model-based and patient-specific robot assistance and rehabilitation paradigms will lead to an increased understanding of the brain and improved patient outcomes.