Unique Solution for Feed-forward Control of Neuroprosthetic Systems Characterized by Redundant Muscles Acting on Multiple Degrees of Freedom

INTRODUCTION We previously developed a method for implementing feedforward neuroprosthetic controllers for musculoskeletal systems with multiple degrees of freedom and complex mechanical interactions [1]. These controllers employed inverse models of the musculoskeletal systems under control. Experimental tests showed poorer performance than expected, which we attribute to redundancy of the data used to develop the inverses. The inverse relationship between muscle output and electrical stimulation is not unique (most joints have redundant actuation with non-stationary input-output muscle properties and coupled degrees of freedom) and if left unrestricted, neural networks trained to model the inverse may produce undesirable solutions. To overcome this problem, we automated a method to choose a single optimal inverse prior to network training [2]. Our present work involves obtaining this unique inverse solution and training a controller capable of independently controlling coupled degrees of freedom. We evaluate our solution method by testing the controllers in simulation and experimentally with able-bodied and spinal cord injured human subjects.