Locomotion Control of MEMS Microrobot Using Pulse-Type Hardware Neural Networks

Transcription

1 Electrical Engineering in Japan, Vol. 186, No. 3, 2014 Translated from Denki Gakkai Ronbunshi, Vol. 132-C, No. 7, July 2012, pp Locomotion Control of MEMS Microrobot Using Pulse-Type Hardware Neural Networks KEN SAITO, KAZUTO OKAZAKI, TATSUYA OGIWARA, MINAMI TAKATO, KATSUTOSHI SAEKI, YOSHIFUMI SEKINE, and FUMIO UCHIKOBA Nihon University, Japan SUMMARY This paper presents the locomotion control of a microelectromechanical system (MEMS) microrobot. The MEMS microrobot demonstrates locomotion control by pulse-type hardware neural networks (P-HNN). P-HNN generate oscillatory patterns of electrical activity like those of living organisms. The basic component of P-HNN is a pulse-type hardware neuron model (P-HNM). The P-HNM has the same basic features as biological neurons, such as the threshold, the refractory period, and spatiotemporal summation characteristics, and allows the generation of continuous action potentials. P-HNN has been constructed with MOSFETs and can be integrated by CMOS technology. Like living organisms, P-HNN has realized robot control without using software programs or A/D converters. The size of the microrobot fabricated by MEMS technology was mm. The frame of the robot was made of a silicon wafer, equipped with rotary actuators, link mechanisms, and six legs. The MEMS microrobot emulated the locomotion method and the neural networks of an insect by rotary actuators, link mechanisms, and the P-HNN. We show that the P-HNN can control the forward and backward locomotion of the fabricated MEMS microrobot, and that it is possible to switch its direction by inputting an external trigger pulse. The locomotion speed was 19.5 mm/min and the step size was 1.3 mm Wiley Periodicals, Inc. Electr Eng Jpn, 186(3): 43 50, 2014; Published online in Wiley Online Library (wileyonlinelibrary.com). DOI /eej Key words: pulse-type hardware neuron model; neural networks; MEMS; microrobot. Contract grant sponsor: Nihon University Academic Research Grant (11-002) JSPS KAKENHI ( ). We also received support from the Research Center for Micro Functional Devices, Nihon University. 1. Introduction There has been extensive research on microrobots, and their use is expected in many fields, such as medical treatment and the operation of micro components. Further miniaturization and sophistication of microrobots requires the development of small-size locomotion mechanisms, small-size energy sources, and control systems that can flexibly handle unforeseen situations. In addition, there is a limit to the fabrication of components by conventional mechanical machining technologies. There are reports of microrobots using microelectromechanical system (MEMS) technologies based on IC production line [1, 2]. Insects and other small animals have very compact systems with excellent functions, but at the current level of program control using microcontrollers and other devices, it is difficult to implement such flexible advanced control as that in living organisms. Aiming at utilization of the superior information processing mechanisms of biological neural networks, researchers have tried to imitate neurons by mathematical models and to implement their functions by networking. In particular, biological information processing is realized by synchronous firing among neurons [3, 4], and consequently synchronous phenomena occurring in coupled neurons have been attracting attention [5]. Kawakami and colleagues thoroughly analyzed synchronous phenomena in coupled systems using models proposed by Bonhoeffer van der Pol and Hodgkin Huxley [6 8]. There have also been studies of hardware neuron models intended for engineering applications. Specifically, hardware neuron models implemented by nonlinear oscillators have been networked to analyze their synchronous phenomena, just as in the case of mathematical models. In such networks, nonlinear oscillators are coupled via resistances, and various phase patterns can be generated by varying the initial values of the oscillators and control parameters (coupling resistance). However, the phase pat Wiley Periodicals, Inc.

2 terns of coupled oscillator networks often depend on the initial settings, and it may be difficult to generate the desired patterns solely by varying the control parameters; in addition, nonlinear oscillators include multiple inductors, which complicates IC chip implementation [9 13]. We built and investigated coupled oscillator networks using a pulse-type hardware neuron model (P-HNM) in which the nonlinear oscillators did not include inductors [14 17]. For example, we built a model of a CPG (Central Pattern Generator) for quadrupeds that makes possible the generation of locomotion patterns [16]. In this study, we used MEMS technologies to fabricate a mm microrobot that imitates the locomotion of an insect, and realized locomotion control of the robot with a pulse-type hardware neural networks (P- HNN). 2. The MEMS Microrobot Fig. 2. Schematic diagram of rotary actuator. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.] 2.1 Main components The main components of the microrobot were fabricated on a silicon wafer using MEMS technologies. The robot has six legs to simulate the motion patterns of an insect. The robot is composed of a frame, rotary actuators, and a link mechanism. The MEMS microrobot is shown in Fig. 1. The frame, rotary actuators, and link mechanism were implemented on a silicon wafer. Components with thicknesses of 100, 200, 385, and 500 µm were created by photolithography, patterning, and ICP dry etching. Except for the artificial muscle wires [18] to drive the rotary actuators and the shafts to connect the legs, link mechanism, and frame, all parts of the mm microrobot were fabricated by MEMS technologies [19]. The rotary actuator is shown schematically in Fig. 2. The actuators, consisting of a rotor and four artificial muscle wires, were arranged on the left and right sides of the microrobot. The four artificial muscle wires were installed in through holes provided in the rotor, with one end connected to the power supply and the other end grounded. Thus, the microrobot in Fig. 1 has a total of eight artificial muscle wires. Eight input terminals and eight ground terminals are arranged on the top portion of the robot. The artificial muscle wires are made of shape-memory alloy (Ti-Ni alloy); the wire contracts by about 200% at high temperature (70 C), and relaxes when it gives off heat. In this study, the artificial muscle wires were contracted by the Joule heat produced by current of 75 ma, and were relaxed when the current was interrupted. The motion of the rotary actuators is determined by the order in which the artificial muscle wires are contracted. The link mechanism is shown schematically in Fig. 3. The mechanism transmits the rotation of the rotary actuators to the robot legs. The center leg is connected to Fig. 1. Fabricated MEMS microrobot. Fig. 3. Schematic diagram of link mechanism. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.] 44

3 actuators in opposite phase. That is, A and D, B and C, C and B, D and A were coupled so that the six legs were driven by four-channel locomotion generation patterns. 3. Pulse-Type Hardware Neural Network 3.1 Pulse-type hardware neuron model Fig. 4. Waveforms to actuate MEMS microrobot. the rotary actuator and rotates in phase with the actuator. The front and rear legs are connected to the central leg by links and move with a phase shift of π/2 with respect to the central leg. Using this link mechanism, three legs are driven simultaneously by a single actuator to simulate insect locomotion. P-HNM, a structural element of P-HNN, is an analog electronic circuit model that reproduces the pulse waveforms output by biological neurons. The basic circuit diagram of P-HNM is given in Fig. 6. P-HNM consists of a synaptic model and cell body model. The basic circuit diagram of a synaptic model with J inputs is shown in Fig. 6(a). The synaptic model has 2.2 Locomotion generation mechanism Figure 4 shows the locomotion generation waveforms that operate the rotary actuators of the MEMS microrobot. The direction of locomotion of the MEMS microrobot depends on the order in which current is passed through the artificial muscle wires. Let A, B, C, D and A, B, C, D denote the artificial muscle wires on the right and left side of the microrobot, respectively. When artificial muscle wire A contracts, the rotor moves rightward, and the front and rear legs push off against the ground, while the central leg floats. When the current in wire A is interrupted and wire B contracts, the central leg touches the ground and the front and rear legs float. When wire C contracts, the central leg pushes off against the ground and the front and rear legs move forward. When wire D contracts, the front and rear legs touch the ground while the central leg leaves the ground. Thus, the MEMS microrobot advances as current is applied successively from wire A to wire D; conversely, the robot retrogresses as current is applied successively from wire D to wire A (see Fig. 5). We simulated insect locomotion by driving the right and left rotary Fig. 5. Locomotion of MEMS microrobot. Fig. 6. Basic circuit diagram of P-HNM. 45

4 spatiotemporal summation characteristics, like a living organism; the j-th inputs v inj (t) are summed spatiotemporally to obtain the synapse output voltage v s (t) as follows: The basic circuit configuration of the cell body model is shown in Fig. 6(b). The cell body model is composed of a voltage-controlled negative resistance circuit [20], an equivalent inductance circuit, a membrane capacitance C M, and a leakage resistance R M. The voltage-controlled negative resistance circuit and equivalent inductance include an n-channel enhancement mode MOSFET, a p-channel enhancement mode MOSFET, a voltage source V A, a resistance R G, and a capacitance C G. Like biological neurons, the cell body model is provided with refractory period analog output pulses, and a time-varying negative resistance. Forced oscillations are generated by DC or AC voltage input. That is, the cell body model is an oscillator model with a voltage-controlled negative resistance element. When the synaptic model is added to the cell body model, v s (t) is input from the latter and v M (t) is output at the terminals of capacitance C M. The output equations for v M (t) and the terminal voltage v G (t) of the capacitance C G are the following two differential equations: The current i Λ (v M (t), v G (t)) flowing in the voltagecontrolled negative resistance (the time-varying nonlinear term in Eq. (2) related to the Λ-shaped negative resistance characteristic) is defined as follows, where k n (k p ) and V Tn (V Tp ) denote the conductance coefficient and the threshold voltage of the n-channel (p-channel) MOSFET. When v G (t) + V A + V Tn + V Tp < v M (t) V A + V Tp : (1) (2) (3) Here A = v G (t) + V Tn, B = V Tp v M (t); k n and k p are approximated by Eq. (7) as follows (where µ is the carrier mobility, ε is the dielectric constant of the gate oxide film, t is the thickness of the oxide film, W is the channel width, and L is the channel length): An example phase plane of the cell body model is shown in Fig. 7. The phase plane is usually used in analysis of entrainment phenomena in oscillators and for stability instability discrimination. The horizontal axis represents v M (t) and the vertical axis represents v G (t); the dotted line and the dashed line show nullclines of v M (t) in Eq. (2) and v G (t) in Eq. (3), respectively, and the solid line shows the output attractor of the cell body model. The circuit constants were set at β = 0.23 ma/v 2, V Tn = 0.8 V, V Tp = 1.5 V, C M = 270 nf, C G = 39 µf, R M = 10 kω, R G = 300 kω; the source voltage was V A = 3.3 V. In the diagram, the attractor is drawn to a limit cycle. In addition, when dv M /dv G > 0 (a positive gradient at the equilibrium point where the two nullclines intersect), the cell body model becomes unstable and self-excited oscillations occur. When dv M /dv G < 0, the equilibrium point of the cell body model is stable and self-excited oscillations do not occur. In this study, the circuit constants were set so that the gradient at the equilibrium point was positive, the cell body model was unstable, and self-excited oscillations occurred. That is, no matter where the initial values of the cell body model lie on the phase plane, the output attractor is drawn to a limit cycle along the v M (t) nullcline, and oscillations continue steadily. In addition, the nullclines in Fig. 7 have the properties of Class II neurons, just as in the Bonhoeffer van der Pol model, the Hodgkin Huxley model, and other widely used neuron models. (7) (4) When V A + V Tp < v M (t) v G (t) + V A + V Tn : (5) When v G (t) + V A + V Tn < v M (t) V A : (6) Fig. 7. Example of phase plane of cell body model. 46

5 3.2 Locomotion control A schematic diagram of the P-HNN that outputs locomotion control waveforms for the MEMS microrobot is shown in Fig. 8. In the diagram, E denotes the excitatory neuron model, I denotes the inhibitory neuron model, and the excitatory and inhibitory connections are shown by white and black circles, respectively. In this model composed of four excitatory inhibitory neural pairs, all inhibitory neuron models are interconnected via inhibitory synapses. The locomotion control waveforms output by P-HNN are the output waveforms of excitatory neuron models E 1, E 2, E 3, E 4. Therefore, excitatory neuron models E 1, E 2, E 3, E 4 are connected to artificial muscle wires A, B, C, D, A, B, C, D of the MEMS microrobot. A circuit diagram of the P-HNN element, the excitatory inhibitory neural pair model, is shown in Fig. 9. The circuit constants of the synaptic models were set at R S1E-S4E = 1 MΩ, R ine = R ini = 100 kω, C SE = C SI = 1 pf. The circuit constants of the cell body models were set at C M = 270 nf, C G = 39 µf, R M = 10 kω, R G = 300 kω; the source voltage was V A = 3.3 V. The P-HNN was surface-mounted on an FR4 substrate measuring about 10 cm on a side. The output waveforms of the P-HNN are shown in Fig. 10. The diagram shows the waveforms needed for driving the MEMS microrobot, as in Fig. 4; thus, the newly developed P-HNN can control forward and backward locomotion of the robot. Switching between the forward and backward directions is possible by inputting a trigger pulse. In locomotion experiments, output control waveforms from the P-HNN were input to the MEMS microrobot placed on a cutting board. The robot moved at a speed of 19.5 mm/min with a step size of 1.3 mm. Thus, we demonstrated that the proposed P-HNN provides effective locomotion control of the MEMS microrobot. The walking speed of the robot depends on the P-HNN output waveforms; in order to Fig. 9. Circuit diagram of excitatory inhibitory neural pair model. increase the locomotion speed, one must reduce the width and period of the output pulses of the cell body model. The width and period of the output pulse can be reduced by setting C G and C M smaller in the cell body model; however, contraction and relaxation of the artificial muscle wires takes 1 s, which governs C G and C M. 3.3 CMOS-based P-HNN With the cooperation of the VLSI Design and Education Center (VDEC), we designed an IC chip measuring 2.3 mm on a side with two polysilicon layers and two metal interconnection layers (ON Semiconductor). In a singlechip implementation for installation on a microrobot, the network circuit must be simplified, and enhancement mode MOSFETs must be used as shown in Fig. 6. A simplified configuration of the circuit in Fig. 8 is shown in Fig. 11. An excitatory inhibitory neural pair model is needed to generate various motion patterns of a quadruped animal [16]; however, when only forward and Fig. 8. Schematic diagram of P-HNN. Fig. 10. Output waveforms of P-HNN. 47

6 Fig. 11. Schematic diagram of simplified P-HNN. backward motion patterns are required, the excitatory connections can be simplified and a network can be configured using only inhibitory connections [22]. The inhibitory connections suppress oscillations of the cell body models, and therefore when the source voltage is supplied simultaneously to four cell body models interconnected by inhibitory synaptic models, they first oscillate randomly but then eventually synchronize due to mutual inhibition. In this study, the oscillation order of the cell body models was controlled by a time-shifted voltage supply for simplicity. However, the oscillation order can be controlled arbitrarily by inputting trigger pulses to simulate external stimuli. The basic circuit diagram of the CMOS-based P- HNN is shown in Fig. 12. The CMOS-based synaptic model in Fig. 12(a) has spatiotemporal summation characteristics, as in Fig. 6(a). Here inputs from three cell body models are summed. The circuit constants of the CMOS-based synaptic model were set so as to obtain a connection weight of 1, namely, C S1 = C S2 = C S3 = 1 pf, M S11 13, M S21 23, M S31 33 : W/L = 1, M S4,5 : W/L = 1; the source voltage was V DD = 5 V. The CMOS-based cell body model in Fig. 12(b) has the Fig. 13. Output waveforms of P-HNN using CMOS (PSpice simulation). same characteristics as that in Fig. 6(b). The circuit constants of the CMOS-based cell body model were set so as to obtain an oscillation period of 4 s and a pulse width of 1 s, namely, C G = 39 µf, C M = 270 nf, M 1, M 2 : W/L = 10, M 3 : W/L = 0.1, M 4 : W/L = 0.3; the source voltage was V A = 3.3 V. The output waveforms of the CMOS-based P-HNN are shown in Fig. 13. These output waveforms were obtained by PSpice simulations. As can be seen from the diagram, the waveforms are identical to those required for driving the MEMS microrobot (Figs. 4 and 10); thus, the proposed P-HNN can control forward and backward locomotion of the robot. Based on the results of the PSpice simulations, the waveforms can be transmitted via a GPIB connected to a waveform generator, thus driving the MEMS microrobot. The capacitances C G, C M in the cell body model are too large to be implemented on the designed CMOS IC chip, and thus external chip capacitors had to be connected by wire bonding (Fig. 14). Fig. 12. Basic circuit diagram of P-HNN using CMOS. Fig. 14. MEMS microrobot with CMOS IC chip. 48

7 4. Conclusions We fabricated a mm microrobot using MEMS technologies. In addition, we developed a P-HNN and used it for locomotion control of the MEMS microrobot. The robot s locomotion was controlled by the P- HNN output waveforms without using any programs, A/D converters, and the like. The fabricated microrobot performed forward and backward motion patterns at a speed of 19.5 mm/min and with a step size of 1.3 mm. In the future, we plan to mount the developed CMOS IC chip on the robot. Acknowledgment This study was assisted by a Nihon University Academic Research Grant (11-002) JSPS KAKENHI ( ). We also received support from the Research Center for Micro Functional Devices, Nihon University. The VLSI chip in this study has been fabricated in the chip fabrication program of VLSI Design and Education Center (VDEC), the University of Tokyo in collaboration with Cadence Design Systems, Inc., On-Semiconductor, HOYA Corp., and KYOCERA Corp. REFERENCES 1. Edqvist E, Snis N, Mohr RC et al. Evaluation of building technology for mass producible millimetersized robots using flexible printed circuit boards. J Micromech Microeng 2009;19(7): Donald BR, Levey CG, McGray CD, Paprotny I, Rus D. An untethered, electrostatic, globally controllable MEMS micro-robot. J Microelectromech Syst 2006;15: Delcomyn F. Neural basis of rhythmic behavior in animals. Science 1980;210: Arbib M (editor). The handbook of brain theory and neural networks, 2nd ed. MIT Press; Yamaguchi Y. Recognition and memory oppressed by oscillatory neural circuits. Science Publ.; p (in Japanese) 6. Tsumoto K, Yoshinaga T, Aihara K, Kawakami H. Bifurcations in synaptically coupled Hodgkin Huxley neurons with a periodic input. Int J Bifurcation Chaos 2003;13: (in Japanese) 7. Tsuji S, Ueta T, Kawakami H, Aihara K. Bifurcation analysis of current coupled BVP oscillators. Int J Bifurcation Chaos 2007;17: (in Japanese) 8. Tsumoto K, Yoshinaga T, Iida H, Kawakami H, Aihara K. Bifurcations in a mathematical model for circadian oscillations of clock genes. J Theor Biol 2006;239: Moro S, Matsumoto T. Effect of coupling resistors on steady patterns in coupled oscillators networks. Trans IEICE 2003;J86-A: (in Japanese) 10. Endo T, Mori S. Mode analysis of a ring of a large number of mutually coupled van der Pol oscillators. IEEE Trans Circuits Syst 1978;25: Kitajima H, Yoshinaga T, Aihara K, Kawakami H. Burst firing and bifurcation in chaotic neural networks with ring structure. Int J Bifurcation Chaos 2001;11: Yamauchi M, Wada M, Nishino Y, Ushida A. Wave propagation phenomena of phase states in oscillators coupled by inductors as a ladder. IEICE Trans Fundam 1999;E82-A: Yamauchi M, Okuda M, Nishino Y, Ushida A. Analysis of phase-inversion waves in coupled oscillators synchronizing at in-and-anti-phase. IEICE Trans Fundam 2003;E86-A: Sasano N, Saeki K, Sekine Y. Short-term memory circuit using hardware ring neural networks. Artif Life Robotics 2005;9: Nakabora Y, Saeki K, Sekine Y. Synchronization of coupled oscillators using pulse-type hardware neuron models with mutual coupling International Technical Conference on Circuits/Systems, Computers and Communications (ITC-CSCC 2004), p 8D2L-3-1 8D2L Hata K et al. A pulse-type hardware CPG model for generation and transition of quadruped locomotion pattern. Trans IEEJ 2007;127C: (in Japanese) 17. Sekine Y, Saeki K. CMOS implementation of pulsetype hardware neuron model and its application. JNNS J 2008;5: (in Japanese) 18. Homma D. Metal artificial muscle biometal fiber. RSJ 2003;21: Okazaki K, Ogiwara T, Yang D, Sakata K, Saito K, Sekine Y, Uchikoba F. Development of pulse control type MEMS microrobot with hardware neural network. Artif Life Robotics 2011;16: (in Japanese) 20. Sekine Y. A Λ-type neuron model using enhancement mode MOSFETs. Trans IEICE 2001;J84-C: (in Japanese) 21. Saito K, Matsuda A, Saeki K, Uchikoba F, Sekine Y. Synchronization of coupled pulse-type hardware neuron models for CPG model. The relevance of the time domain to neural network models. Springer Series in Cognitive and Neural Systems, Vol. 3, p , Saito K, Okazaki K, Sakata K, Ogiwara T, Sekine Y, Uchikoba F. Pulse-type hardware inhibitory neural networks for MEMS micro robot using CMOS technology. Proc 2011 International Joint Conference on Neural Networks, p

8 AUTHORS (from left to right) Ken Saito (member) received a bachelor s degree in electronic engineering from Nihon University in 2001, completed the M.E. and doctoral programs in 2004 and 2007, and became a research associate there. He is now a professor of precision mechanical engineering in the College of Science and Technology. His research interests are hardware neural networks. He holds a D.Eng. degree. Kazuto Okazaki (student member) received a bachelor s degree in precision mechanical engineering from Nihon University in 2010 and is now in the M.E. program in precision mechanical engineering in the Graduate School of Science and Technology. His research interests are microrobots. Tatsuya Ogiwara (nonmember) received a bachelor s degree in precision mechanical engineering from Nihon University in 2010 and is now in the M.E. program in precision mechanical engineering at the Graduate School of Science and Technology. His research interests are insect-type MEMS microrobots with artificial brain. Minami Takato (nonmember) received a bachelor s degree in precision mechanical engineering from Nihon University in 2010 and is now in the M.E. program in precision mechanical engineering at the Graduate School of Science and Technology. Her research interests are miniature ceramics-based energy systems for microrobots. Katsutoshi Saeki (member) completed the first stage of the doctoral program at Nihon University (Graduate School of Science and Technology) in 1989 and joined Toshiba Corporation. He was appointed a research associate at Nihon University in 1993, and is now an associate professor in the Department of Electronics and Computer Science, College of Science and Technology. His research interests are analog IC, hardware implementation of neural networks. He received an IEEJ Excellent Paper Presentation Award in 1999 and IEEJ Electronics, Information and Systems Society Contribution Award in He holds a D.Eng. degree, and is a member of IEICE and JNNS. Yoshifumi Sekine (fellow) received a bachelor s degree in electrical engineering from Nihon University in 1966, completed the M.E. program in 1968, and joined the faculty as a research associate. He is now a professor in the Department of Electronics and Computer Science. His research interests are negative resistance elements, oscillation circuits, hardware neuron models, neural coding, development and application of pulse-type hardware neural networks. He holds a D.Eng. degree. Fumio Uchikoba (member) received a bachelor s degree in applied physics from Waseda University in 1983, completed the M.E. program at the University of Electro-Communications in 1985 and joined TDK Corp. He was a visiting researcher at Massachusetts Institute of Technology in He was appointed an associate professor at Nihon University in 2003, and is now a professor of precision mechanical engineering in the College of Science and Technology. He holds a D.Eng. degree. 50