Design of control algorithms for redundant neuroprosthetic brain-machine interfaces
Dexterous control of upper arm movements is integral to activities of daily living. This important ability may be lost due to neurological injury or disease, yet brain-machine interfaces (BMIs) or neuroprosthetics may offer the ability to restore motor function by bypassing damaged neural circuitry. Though tremendous progress has been made in BMI performance and robustness, additional gains are needed to improve the clinical utility of BMI systems. Of particular importance is the need to scale BMI performance beyond the simple 2-D reaching paradigm typically studied and extend to the multi-degree-of-freedom (multi-DOF) case. We explored methods and design considerations for redundant multi-DOF BMIs and examine the neuroprosthetic control mechanisms engaged by subjects.
In this work, healthy non-human primate subjects controlled a variety of virtual and physical actuators using neural signals recorded from electrode arrays implanted in the motor cortex. We analyzed 2-D natural arm movements and developed a fast method to extract sparse components of hand kinematics during point-to-point reaching. We then analyzed neuroprosthetic control of a computer cursor in which neural decoding is performed using a Kalman filter (KF). We discovered several links between KF parameters and control performance of the overall BMI system, including the ability to manipulate the speed-accuracy tradeoff of neural cursors. Building from hypotheses of model based control in natural motor control, we examine BMI cursor control for evidence of model-based and feedforward control.
We then proceeded to develop an architecture for BMI control of a redundant actuator. Subjects learned BMI control of a redundant four-link virtual “arm” which was confined to move in a two-dimensional plane. When controlling this redundant arm, performance improved when subjects made use of the redundant control dimensions. Furthermore, redundant control signals were critical to escaping manipulator singularities, similar to how redundant control signals are used in robotic systems. These results suggest the utility of using neural signals in conjunction with standard robotic path planning methods to control endpoint-redundant aspects of a robotic prosthesis. Altogether, the work of this dissertation contributes progress toward BMI control of highly redundant prosthetic systems for clinical use.