2.1. Quadruped robot

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 neu-

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

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

The quadruped and hexapod robots have been described in previous works [26, 27]. This section briefly introduces the mechanical components of the fabricated multilegged robots.

ron model requires no inductors; therefore, the system is easily implemented in an IC.

hexapod robot [27–29].

302 Advanced Applications for Artificial Neural Networks Advanced Applications for Artificial Neural Networks

multilegged robot.

2. Multilegged robots

The width, length, and height dimensions of the quadruped robot are 130, 140, and 90 mm, respectively (see Figure 1). The quadruped robot is constructed from mechanical and electrical components. The mechanical components comprise the body frame, four servo motors, link mechanisms, and four legs. The electrical components consist of the control board, hardware ANNs, and battery. The limbs and body frame are made from aluminum base alloy 2017 and aluminum base alloy 5052, respectively. The mechanical parts are fabricated by a computerized numerical control machining system. The four legs of the quadruped robot system are actuated by four servo motors, and the stepping motion of each leg is generated by the link mechanisms. The servo motor is an HSR-8498HB (Hitec Multiplex Japan) model, which generates sufficient maximum torque to actuate the robot. The mechanical components of the quadruped robot are detailed in [26].

Figure 2 shows the relative phase difference of the quadruped gait pattern under a given driving rhythm of the actuators. The relative phase difference is referenced to the left forelimb (0). Under various actuation rhythms, the quadruped robot generates different gait patterns. Figure 2 displays five typical gait patterns: walk, trot, pace, bound, and gallop. The directional changes and turning of the quadruped robot are not realized at present. The robot gait is easily controlled by software programs implemented on a control board. However, in the proposed robot control, the software program for generating the locomotion rhythms is replaced by a hardware ANN.
