UC Berkeley

Brain-machine interface (BMI) systems attempt to restore motor function lost due to injury or neurodegenerative disease by bypassing natural motor pathways and allowing direct neural control of a movement actuator. Such systems also hold promise for investigating questions about learning and motor control in a highly controlled and observable system. Here we utilize a BMI paradigm in which single unit neural spiking activity recorded from motor cortical areas in non-human primates is used to control the movement of a virtual on-screen movement actuator. A Kalman filter is used to decode intended movements from neural signals.

This work presents a newly designed set of hardware and software components for running BMI and other real-time control experiments that provides a flexible and extendible framework for synchronized data collection, stimulus presentation, and behavioral task control. This framework is Python based and soon to be released open source to the neuroscience community.

The feasibility of chronically recording from bimodal sensory neurons in macaque ventral pre-motor cortex (PMv) was tested. Sensory response properties of individual PMv neurons were measured through the presentation of visual and tactile stimuli on and around the primate subject’s arm, while electrophysiology recordings were collected from the brain. The existence of “bimodal” PMv neurons with linked visual-tactile receptive fields was confirmed, and groundwork was laid for a future study observing changes in sensory responses in PMv neurons as skilled BMI control develops.

Inter-subject and inter-task variability in adaptation strategies to a novel feedback perturbation were explored in both natural motor control and BMI control contexts. The feedback perturbation involved the addition of a constant velocity vector to the control signal generated by the subject, which both improved and hampered performance on a target hitting task depending on the movement direction within the workspace. No consistent adaptation strategy across all subject-context combinations was observed, indicating that neither a local nor a global adaptation strategy was consistently applied in the presence of such a perturbation.

Finally, a novel, virtual, kinematically redundant BMI actuator consisting of a 4-link kinematic chain was developed and tested. Macaque subjects were shown to be capable of actively controlling redundant degrees of freedom when available. Removing redundant degrees of freedom hindered performance on a control task. In contrast to previously published observations for kinematic control in the natural motor system, both task-relevant and null movements increased over time.

In summary, this work addresses a range of questions related to the design and implementation of BMI systems, as well as the control strategies employed by the brain during various types of BMI control. It lays the groundwork for several future studies to investigate the performance effects of characteristics such as PMv control and redundancy, and establishes new paradigms for studying learning in the context of BMI control and potentially motor control in general.