Neuromorphic meets neuromechanics, part I: the methodology and implementation

Abstract

Objective One goal of neuromorphic engineering is to create “realistic” robotic systems that interact with the physical world by adopting neuromechanical principles from biology. Critical to this is the methodology to implement the spinal circuitry responsible for the behavior of afferented muscles. At its core, muscle afferentation is the closed-loop behavior arising from the interactions among populations of muscle spindle afferents, alpha and gamma motoneurons, and muscle fibers to enable useful behaviors. Approach We used programmable Very-Large-Scale-Circuit (VLSI) hardware to implement simple models of spiking neurons, skeletal muscles, muscle spindle proprioceptors, alpha-motoneuron recruitment, gamma motoneuron control of spindle sensitivity, and the monosynaptic circuitry connecting them. This multi-scale system of populations of spiking neurons emulated the physiological properties of a pair of antagonistic afferented mammalian muscles (each simulated by 1,024 alpha- and gamma-motoneurones) acting on a joint via long tendons. Main results This integrated system was able to maintain a joint angle, and reproduced stretch reflex responses even when driving the nonlinear biomechanics of an actual cadaveric finger. Moreover, this system allowed us to explore numerous values and combinations of gamma-static and gamma-dynamic gains when driving a robotic finger, some of which replicated some human pathological conditions. Lastly, we explored the behavioral consequences of adopting three alternative models of isometric muscle force production. We found that the dynamic responses to rate-coded spike trains produce force ramps that can be very sensitive to tendon elasticity, especially at high force output. Significance Our methodology produced, to our knowledge, the first example of an autonomous, multi-scale, neuromorphic, neuromechanical system capable of creating realistic reflex behavior in cadaveric fingers. This research platform allows us to explore the mechanisms behind healthy and pathological sensorimotor function in the physical world by building them from first principles, and it is a precursor to neuromorphic robotic systems.

Neuromorphic meets neuromechanics, part II: the role of fusimotor drive

“…Here we have observed and documented some of these emergent properties. Examples include phasic and tonic reflex responses to a ramp-and-hold perturbation, and signal dependent noise as shown in this paper; as well as maintaining a stable posture following a transient force perturbation to the joint as shown in the companion paper [12]. In future, we will explore other emergent properties of the system (see scientific and clinical implications below).…”

“…We have begun investigating other muscle models that are capable of explaining more complicated physiological phenomena as demonstrated in our companion paper [12]. By implementing a library of muscle models, our neuro-mechano-morphic setup can provide a benchmarking system to compare them in closed-loop (i.e.…”

“…We extended a computational neuromorphic system emulating spinal and transcortical stretch reflex loops [11], and created a hardware-in-the-loop neuro-mechanomorphic system capable of controlling robotic and cadaveric fingers in the physical world as discussed in Part I of our work [12]. In this work — Part II — we present a systematic application of this neuro-mechano-morphic system to explore the effect of different combinations of fusimotor drive on the nuances of the reflex force response.…”

On neuromechanical approaches for the study of biological and robotic grasp and manipulation

“…Know how to solve every problem that has been solved.' 3 In our context, it can be taken to mean that, if we have over one hundred years of sensorimotor neuroscience since Sir Charles Sherrington [74], and if the principles we have deduced are sound, then we should be able to build components that embody those mechanisms in such a way that when assembled they behave like biological systems [75,76]. One example of such a neuromorphic approach uses ultra-fast computer processors to simultaneously implement populations of autonomous, interconnected spiking neurons in real time that follow Hodgkin-Huxley rules of how action potentials in neurons are initiated and propagated [77].…”

“…Bayesian inference [186,187] or stochastic control [143] are formal approaches to describe the emergence of probability density functions of the mapping from perception/intention to action in the presence of unavoidable uncertainty, noise, risk and variability in the real world. Moreover, there are probabilistic approaches that can be used with populations of spiking neurons to produce physical behavior in a cost-agnostic, emergent, model-free way [188,189,75,76].…”