Closed-loop prosthetic hand : understanding sensorimotor and multisensory integration under uncertainty.
To make sense of our unpredictable world, humans use sensory information streaming through billions of peripheral neurons. Uncertainty and ambiguity plague each sensory stream, yet remarkably our perception of the world is seamless, robust and often optimal in the sense of minimising perceptual variability. Moreover, humans have a remarkable capacity for dexterous manipulation. Initiation of precise motor actions under uncertainty requires awareness of not only the statistics of our environment but also the reliability of our sensory and motor apparatus. What happens when our sensory and motor systems are disrupted? Upper-limb amputees tted with a state-of-the-art prostheses must learn to both control and make sense of their robotic replacement limb. Tactile feedback is not a standard feature of these open-loop limbs, fundamentally limiting the degree of rehabilitation. This thesis introduces a modular closed-loop upper-limb prosthesis, a modified Touch Bionics ilimb hand with a custom-built linear vibrotactile feedback array. To understand the utility of the feedback system in the presence of multisensory and sensorimotor influences, three fundamental open questions were addressed: (i) What are the mechanisms by which subjects compute sensory uncertainty? (ii) Do subjects integrate an artificial modality with visual feedback as a function of sensory uncertainty? (iii) What are the influences of open-loop and closed-loop uncertainty on prosthesis control? To optimally handle uncertainty in the environment people must acquire estimates of the mean and uncertainty of sensory cues over time. A novel visual tracking experiment was developed in order to explore the processes by which people acquire these statistical estimators. Subjects were required to simultaneously report their evolving estimate of the mean and uncertainty of visual stimuli over time. This revealed that subjects could accumulate noisy evidence over the course of a trial to form an optimal continuous estimate of the mean, hindered only by natural kinematic constraints. Although subjects had explicit access to a measure of their continuous objective uncertainty, acquired from sensory information available within a trial, this was limited by a conservative margin for error. In the Bayesian framework, sensory evidence (from multiple sensory cues) and prior beliefs (knowledge of the statistics of sensory cues) are combined to form a posterior estimate of the state of the world. Multiple studies have revealed that humans behave as optimal Bayesian observers when making binary decisions in forced-choice tasks. In this thesis these results were extended to a continuous spatial localisation task. Subjects could rapidly accumulate evidence presented via vibrotactile feedback (an artificial modality ), and integrate it with visual feedback. The weight attributed to each sensory modality was chosen so as to minimise the overall objective uncertainty. Since subjects were able to combine multiple sources of sensory information with respect to their sensory uncertainties, it was hypothesised that vibrotactile feedback would benefit prosthesis wearers in the presence of either sensory or motor uncertainty. The closed-loop prosthesis served as a novel manipulandum to examine the role of feed-forward and feed-back mechanisms for prosthesis control, known to be required for successful object manipulation in healthy humans. Subjects formed economical grasps in idealised (noise-free) conditions and this was maintained even when visual, tactile and both sources of feedback were removed. However, when uncertainty was introduced into the hand controller, performance degraded significantly in the absence of visual or tactile feedback. These results reveal the complementary nature of feed-forward and feed-back processes in simulated prosthesis wearers, and highlight the importance of tactile feedback for control of a prosthesis.