1. Introduction

Many types of multilegged robots have been developed for various applications [1–3]. Most of these robots were bioinspired by the structures, features, and excellent functionalities of living organisms [4, 5]. Living organisms autonomously operate under the control of small-sized neural networks. Therefore, researchers have begun studying artificial neural networks (ANNs) for robot control [6–10]. Biological neural networks are universally characterized by oscillatory patterns of electrical activity. These patterns govern several functions of living organisms, such as heart rhythms, movements, and swallowing [11, 12]. The oscillatory patterns

© The Author(s). Licensee InTech. This chapter is distributed under the terms of the Creative Commons © 2018 The Author(s). Licensee InTech. This chapter is distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/3.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Attribution License (http://creativecommons.org/licenses/by/3.0), which permits unrestricted use, distribution, and eproduction in any medium, provided the original work is properly cited.

of living organisms can be clarified through coupled neuron models, which have two categories: class I and class II [13]. Given that the class II model is more easily synchronized than the class I model, the class II model is applied in studies of synchronization phenomena. Famous class II neuron models include the Hodgkin-Huxley model [14] and Bonhoeffer-van der Pol model [15], mathematical neuron models that form the basis of bioinspired oscillatory pattern generation [16–18]. Most of the central pattern generators (CPGs) designed for the synchronized locomotion control of multilegged robots [6–8] are also constructed by mathematical neuron models. A CPG model using mathematical neuron models can be implemented on a field programmable gate array (FPGA). However, an FPGA board cannot be mounted on a millimeter-sized robot system because of its size. Instead, oscillatory patterns for very small robots can be generated by hardware neuron models. Hardware rings of coupled oscillators, which can generate various oscillatory patterns by using the synchronization phenomena [19, 20], have been employed as the structural elements of ANNs. However, given that most of the hardware neuron models contain inductors in their circuit architectures [19–22], they are difficult to implement in an integrated circuit (IC); thus, the use of such models is disadvantageous on the circuit scale [23]. In particular, ICs can be combined with mechanical parts of the robot by using microelectromechanical system (MEMS) technology, which can reduce the robot size to the millimeter scale.

The authors are studying hardware ANNs based on a pulse-type hardware neuron model [24– 27] with the same basic features as biological neurons. Specifically, this model possesses spatiotemporal summation characteristics, a threshold period, and a refractory period and generates oscillating patterns of electrical activity. Furthermore, the pulse-type hardware neuron model requires no inductors; therefore, the system is easily implemented in an IC.

Previously, the authors proposed two types of prototype multilegged robots: a quadruped robot approximately 10 cm in size [26] and a hexapod robot approximately 5 mm in size [27]. Both multilegged robots move their limbs by stepping motions. A multilegged robot usually needs actuators for each joint. In our multilegged robots, the number of actuators is reduced by a link mechanism, and the gait is controlled by a hardware ANN. The hardware ANN consists of 4 excitatory synaptic models, 16 inhibitory synaptic models, and 8 cell body models for the quadruped robot [26], and 12 inhibitory synaptic models and 4 cell body models for the hexapod robot [27–29].

This chapter describes the basic characteristics of the hardware ANNs that generate the gait of multilegged robots. After briefly introducing both types of multilegged robots, it discusses the hardware ANNs and mathematically describes the characteristics of the pulse-type hardware neuron model. The oscillation characteristic of the model requires a negative resistance and is described in a phase plane. The synchronization characteristics of connected hardware ANNs are also discussed. Finally, the hardware ANNs are validated in locomotion tests of the multilegged robot.
