From insect to robot walking: A biorobotics investigation into network topology

The ability to walk in legged animals requires the coordination of many degrees of freedom. This coordination can be driven by sensory feedback, network architecture, or most likely, a combination of the two. Rhythm­generating circuits called central pattern generators (CPGs) are commonly accepted to be the neural mechanism driving oscillating output to antagonistic muscle pairs. The output from these circuits can be shaped by sensory feedback, adapting movement to enable walking across varied environments. However, the exact network architecture of these neural circuits and how they communicate is not well­defined. Here we present a network topology for CPGs capable of replicating biological behaviors and a mechanism for loosely coupling these circuits to promote coordination. In order to maintain biological plausibility, we adhere to a structured methodology of developing our network using a bottom­up approach constrained by biological parameters and comparing the output to biological measurements. Based on this approach, we find that non­spiking interneurons (NSIs) that communicate through a graded signal play a pivotal role within spiking neural networks (SNNs), increasing biological plausibility while providing benefits to the control of legged robots. This differs from the current control paradigm that uses homogenous neural networks consisting of only one neuronal type. We observe that NSIs can be used to continuously manipulate network output as well as connect spiking populations while isolating their dynamics. This introduces a potential fundamental principle of neurophysiology that NSIs are required for connecting spiking populations to remove instabilities created by competing network dynamics. This separation of dynamics is also useful in robotics, presenting a new method for building controllers of distributed and modular networks. From a broader perspective, the addition of NSIs to SNNs increases biological fidelity so that biological hypotheses can be tested with neural network simulations, providing a controlled and reproducible test environment.