Insect Type MEMS Micro Robot Controlled by CMOS IC of Hardware Neural Networks

M. Takato1, S. Yamasaki1, S. Takahama1, J. Tanida1, K. Saito2 and F. Uchikoba2

1 Precision Machinery Engineering, Graduate School of Science and Technology, Nihon University 2Dept. of Precision Machinery Engineering, Collage of Science and Technology, Nihon University takato@eme.cst.nihon-u.ac.jp, kensaito@eme.cst.nihon-u.ac.jp, uchikoba@eme.cst.nihon-u.ac.jp

Abstract This paper describes insect type micro robots controlled by a CMOS IC of hardware neural networks. The micro robot is fabricated by the micro electro mechanical systems (MEMS) technology using a silicon wafer, and the actuator is composed of artificial muscle wires on the basis of shape memory alloy. Insect-like walking is achieved by link mechanisms that transform the actuator’s rotational motion to locomotive motion. The CMOS IC generates the driving waveform of the micro robot and realizes insect-like walking. The hardware neural networks are built as cell body models and inhibitory synaptic models. The output signal ports of the hardware neural networks are connected to the artificial-muscle-wire-driving circuit. This robot system does not require specialized software programs and A/D converters. The developed neural networks are composed of self-functioning, interconnected plural unit neurons. For proper driving, each neuron in the developed neural network control must be synchronized, as occurs in the neural networks of living organisms. In this study, the motion of the MEMS micro robot is controlled by non-synchronization and anti-phase synchronization driving waveforms. When the non-synchronization driving waveform is input, the micro robot ceases walking motion, but resumes walking upon receipt of the anti-phase synchronization driving waveform. The sideways, endways, and height dimensions of the fabricated micro robot are 4.0 mm, 2.7 mm and 2.5 mm, respectively. The obtained locomotion speed is 26.4 mm/min and the step width is 0.88 mm.

Introduction Insect-like micro robots have been increasingly applied in medicine and other fields that require precise manipulation (Shibata et al., 1997; Takeda, 2001; Habib et al., 2007, 2011; Baisch et al., 2010; Yan et al., 2007). However, to date, very few insect robots are operational at the level of living organisms. Developmental obstacles include the miniaturization of the mechanism, small energy requirements for ensuring long lifetime, and realizing the flexibility of living organisms.

In conventional robot mechanisms, the dominant actuator is an electromagnetic motor, and robotic components are manufactured by mechanical machining. However, very small robots cannot be fabricated by conventional technologies. To overcome this limitation, researchers have developed the micro electro mechanical systems (MEMS) technology on the

basis of the IC production process (Donald et al., 2006; Edqvist et al., 2009). The miniature actuators reported to date include electrostatic actuators (Tang et al., 1989; Sniegowski et al., 1996), electromagnetic actuators (Asada et al., 1994), piezoelectric actuators (Suzuki et al., 1999), and shape memory alloy (SMA) actuators (Surbled et al., 2001). However, the movement of micro robots built with these actuators is impeded on uneven surfaces by small gaps or dust particles. Therefore, micro robots that can walk on an uneven surface are highly sought. Conventional robot control is implemented by digital systems based on microprocessors and software programs. While the pre-programmed digital system exerts adequate control in the specified environment, it may not respond appropriately to unpredictable events. As a means of realizing flexible control, artificial neural networks have attracted considerable attention (Nakada et al., 2003; Delcpyn, 1980). For example, an organism such as an insect realizes walking motion by combining simple neural networks. Moreover, living organisms are flexible and respond sensitively to accidental events. For these reasons, artificial neural network control is a preferable choice in micro robot design. Neural networks have mostly been investigated by software approaches based on mathematical calculations (Tsumoto et al., 2003, 2006; Tsuji et al., 2007; Hodgkin et al., 1952; FitzHugh, 1961). However, using this approach, even simple neural networks consume vast physical and temporal computer resources. Therefore, several researchers have implemented neural networks in hardware (Endo et al., 1978; Kitajima et al., 2001; Yamauchi et al., 1999, 2003; Nagumo et al., 1962; Maeda, 2008). Hardware networks can process nonlinear operations continuously at high speed. Moreover, because the circuit can be embodied in an integrated circuit (IC) (Lewis et al., 2000), extreme size reduction is expected even for large circuits.

The previous micro robot system reported by the authors (Suematsu et al., 2009; Saito et al., 2010; Okazaki et al., 2011) possessed two legs controlled by the MEMS technology and a plastic body. The actuator was fabricated by SMA. The micro robot was controlled by pulse-type hardware networks (P-HNNs), functioning similarly to the central pattern generator of living organisms. The P-HNN was composed of discrete components. In the next step of the project, the six-legged structure and walking motion of the insect were developed. The rotational actuator used SMA-based artificial

ECAL - General Track

ECAL 2013 546

muscle wires, and walking motion was realized by link mechanisms. Motion was controlled by a hardware neural network circuit connected to the actuator. Moreover, the micro robot dimensions were reduced to 4.0 mm, 2.7 mm, 2.5mm in the sideways, endways and height directions, respectively (Uchikoba et al., 2012). In this study, the robot was again controlled by discrete P-HNN circuits.

Recently, we built hardware neural networks into CMOS IC. The IC was connected to the insect-type MEMS micro robot, and its performance was evaluated. In this paper, we explain the mechanism and system of the micro robot and P-HNN IC, and discuss the control characteristics of the P-HNN IC. The driving waveform is provided by synchronization of several hardware neural networks.

System of Micro Robot

CMOS IC Controlled MEMS Micro Robot

In future applications, the developed micro robot will observe and possibly lead swarms of similar micro robots. Therefore, the robot should resemble a live insect as much as possible. For this purpose, we consider hexapod walking motion as a primary objective. Figure 1 shows the mechanism of the designed micro robot.

Figure 1: Mechanism of the Designed Micro Robot

The robot is fitted with six legs on either side. Three of the legs are connected by the link mechanism, while the central leg is held on the rotational actuator built into the body parts. The rotor of the actuator is suspended by artificial muscle wires extending in four directions. The artificial muscle wires, composed of SMA, shrink when heated above the transition temperature of the alloy, and revert to their original length by cooling below the transition temperature. The temperature is raised by flowing an electrical current directly into the wire. The wire is coiled inside the actuator to enable a larger displacement than is possible with a linear wire. The specifications of the artificial muscle wire are shown in Table 1 (http://www.toki.co.jp).

Coil diameter (mm) 0.2 Wire diameter (mm) 0.05 Drive current (mA) 50–120 Resistance (m−1) 3600

Force (gf) 3-5 Displacement (%) 50

Table 1: Specifications of Artificial Muscle Wire

Rotational motion is realized by shrinking the artificial muscle wires in rotational order. A schematic of the actuator’s rotational motion is shown in Figure 2. As wire A shrinks under heating, the rotor is pulled toward the A side. In the next step, wire A is extended by cooling, and wire B is shrunk, dragging the rotor toward the B side. The heating and cooling processes are repeated for wires C and D. These activities induce a clockwise rotation of the rotor.

Figure 2: Schematic of Clockwise Rotational Motion of the Actuator

The driving waveform of the neural networks for generating hexapod walking motion is shown in Figure 3. The walking motion is generated by four pulses. The input voltage and pulse width are 3 V and 0.5 s, respectively.

The central leg is connected to the rotor shaft. The front and rear legs float, while the central leg touches the ground surface. Governed by the rotational motion of the actuator, the middle leg kicks the ground, propelling the micro robot body forward. As the middle leg rises, the front and rear legs touch the ground surface. The outer and central legs are out of phase by 180°. During these motions, the touched points always form a triangle, and stable walking motion, mimicking that of an insect, is realized. Moreover, backward locomotion is generated by reversing the actuator motion. The forward and backward motions are controlled by the electrical current pulses shown in Figure 3 (a) and 3 (b), respectively.

ECAL - General Track

547 ECAL 2013

Figure 3: Schematic of Driving Waveform: (a) Clockwise Pulse (b) Reverse Rotation Pulse

Structure of Micro Robot The fabricated micro robot comprises the body parts, the rotational actuator, and the legs. The body, legs, and actuator rotor are made from single crystal silicon wafers fabricated by the MEMS technology. The four frames of the body structure are assembled by jutting the parts into the grooves. For this purpose, four pieces of artificial muscle wires are connected to the rotor section of the rotational actuator, which is modified by zigzag slits. An electrical ground line is led from the rotor. The other end of the artificial muscle wire is connected to the frame. Figure 4 shows a schematic of the micro robot body and the actuator part, while the shape and dimensions of the rotor part are shown in Figure 5.

Figure 4: Schematic of Micro Robot Body and Actuator

Figure 5: Shape and Dimensions of Rotor Part

The three leg parts are connected by the link mechanism. The front and rear legs move in the counter-direction relative to the center leg. To ensure that the robot moves parallel to the ground, the central leg is made shorter than the other legs. Figure 6 shows the link mechanism of the legs. The central legs are connected to the center of the rotor by a tungsten carbide shaft, while the outer legs are held by shafts attached to the body frame.

Figure 6: Link Mechanism of Legs

The silicon part was fabricated by MEMS photolithography. After washing the silicon wafer, aluminum was deposited by physical vapor deposition, followed by coating with a photoresist. The aluminum film was approximately 0.1 m thick. The designed pattern was exposed to the resist film, and developed by soaking in the developer. The aluminum film on the specimen was then etched chemically, leaving an imprint of the designed pattern. The washed and dried specimen was dry-etched by high-aspect-ratio induced coupled plasma etching combined with a Bosch process (Bhardwaji et al., 1995). The rotor part was obtained after removing the aluminum film and washing. The other robot parts were obtained by repeating this process on both surfaces of the wafer. Hand assembly of the fabricated parts yielded the micro robot.

ECAL - General Track

ECAL 2013 548

Pulse-Type Hardware Neural Networks The P-HNN was built from the cell body and synaptic model circuits. These circuits reproduce the functions of biological neurons. The cell body model circuit is configured as a voltage control negative resistance circuit, an equivalent inductance circuit, membrane capacitor CM, and leak resistor MC4. The cell body model circuit is characterized by a firing threshold, refractory period, and a continuous pulse waveform, similar to the characteristics of biological neurons. Figure 7 is a diagram of the cell body model circuit, powered by source VA.

Figure 7: Diagram of the Cell Body Model Circuit

The synaptic model circuit comprises the excitatory and inhibitory synaptic model circuits, which differ only in the direction of their current. The synaptic model circuit has spatio-temporal summation characteristics mimicking those of biological synapses. That is, the synaptic model circuit sums the output pulses of the cell body model circuit. A diagram of the synaptic model circuit is provided in Figure 8. This circuit is powered by source VDD. The synaptic weight (connection strength between the cell body models) is adjustable by changing the ration of the channel length L and the channel width W in the gate of MIS1 in Figure 8 (b). For example, iISout is increased by increasing the W/L of MIS1.

Figure 8: Diagram of the Synaptic Model Circuit: (a) Excitatory Synaptic Model Circuit (b) Inhibitory Synaptic Model Circuit

The P-HNNs are synchronized by the excitatory and inhibitory synaptic model circuits. Excitatory and inhibitory mutual coupling generates in-phase and anti-phase synchronization, respectively. During walking, we consider that the neural networks of the micro robot might consist solely of inhibitory synaptic models and anti-phase synchronization. If the excitatory synaptic model can be removed without loss of functionality, the number of elements in the model can be reduced. The P-HNN connections are schematized in Figure 9.

Figure 9: Connection Diagram of P-HNN

The four cell body models are mutually coupled by 12 inhibitory synaptic models. Four output ports, extracted from the P-HNN, are connected to the actuator of the MEMS micro robot. In addition, the trigger pulse input ports are extracted to P-HNNs. By accepting different input timing of the external trigger pulse, the P-HNN can alter the sequence of the output waveform.

CMOS IC Figure 10 displays the layout pattern of the P-HNN. The process line is a double-metal single-poly CMOS 0.35 m rule. The area of the IC chip is 1.93 mm square. Because capacitors CG and CM are too large to position on the CMOS IC chip, they are externally connected to the chip; this arrangement also enables straightforward dynamic frequency adjustment. Figure 11 shows a diagram of the artificial muscle wire driving circuit that generates the walking motion of the MEMS micro robot.

ECAL - General Track

549 ECAL 2013

Figure 10: Layout Pattern of P-HNN

Figure 11: Artificial Muscle Wire Driving Circuit

Results and Discussion A photograph of the fabricated micro robot is shown in Figure 12. The sideways, endways, and height dimensions of the micro robot are 4.0 mm, 2.7 mm, and 2.5 mm, respectively. The dimensional error of the actuator component was measured by an optical con-focal microscope, and was found to be always within ±3 m. Moreover, the leg and link parts were connected with adequate clearance fit (measured as 20–30 m).

Figure 12: Photograph of the Fabricated MEMS Micro Robot

Figure 13 compares the discrete circuit board of the artificial neural networks with the packaged IC. This figure shows that the packaged IC was miniaturized to 3.5% of the discrete circuit area, while the bare die realized an area reduction to 0.05%. Thus, considerable size reduction may be achieved by the IC construction.

Figure 13: Size Comparison between the Discrete Circuit and the Packaged IC Discrete Circuit: 10 cm × 8 cm, Packaged IC: 14 mm × 20 mm, Bare Die IC: 1.93 mm × 1.93 mm

Figure 14 shows the output waveform of the IC chip when VDD = 5 V. At this voltage level, the inhibitory synaptic circuit is turned on. In other words, the inhibitory synaptic model circuit is connected, generating an output waveform that inhibits the other cell body model circuits. In response, the cell

ECAL - General Track

ECAL 2013 550

body model circuit comprising I1, I2, I3, and I4 outputs an anti-phase synchronization pattern. Since P-HNNs can generate driving waveforms such as those shown in Figure 3, we have demonstrated that the fabricated IC chip generates correct walking motion of the silicon micro robot.

Figure 14: Output Waveform of the IC Chip

Figure 15 illustrates the walking motion of the MEMS micro robot. The driving waveform that moves the robot is output by the P-HNN. The rotary actuator is then activated and the link mechanism is converted to a walking motion. The locomotion speed is 26.4 mm/min with a step width of 0.88 mm.

Figure 15: Walking Motion of MEMS Micro Robot

Figure 16: Output Waveform of IC (Non-Synchronization Mode)

Figure 17: Output Waveform of IC (Anti-Phase Synchronization Mode)

The motion of the MEMS micro robot was examined under non-synchronization and anti-phase synchronization P-HNN driving waveforms, displayed in Figs. 16 and 17, respectively. In Figure 16, four waveforms are randomly generated, while in Figure 17, they are alternately generated. In the former case, the actuator of the micro robot did not perform rotational motion, hence the robot did not achieve walking function. By contrast, when the anti-phase synchronization driving waveform was connected to the micro robot, walking motion was initiated (Figure 18). To easily visualize the leg movements, these motional comparisons were conducted on a larger micro robot than the fabricated one. The width, length, and height of this micro robot were 8.1 mm, 8.9 mm, and 9.0 mm, respectively. Each figure was observed at 0.5 s intervals. The locomotion speed of this micro robot was 96.0 mm/min with a step width of 4.0 mm. In this test, the MEMS micro robot was controlled by the CMOS IC of hardware neural

ECAL - General Track

551 ECAL 2013

networks, whose synchronization phenomena mimicked those of brain networks.

Figure 18: Motion Response of MEMS Micro Robot to an Anti-Phase Synchronization Driving Waveform

Conclusion This paper proposes an insect-type MEMS micro robot controlled by CMOS IC of hardware neural networks. The micro robot contains a rotational actuator using SMA-based artificial muscle wires and possesses the hexapod structure of an insect. The dimensions of the micro robot are 4.0 mm (sideways), 2.7 mm (endways), and 2.5 mm (height). The micro robot walks when supplied with a synchronized output driving waveform generated by the IC chip of pulse-type hardware neural networks. Thus, as in living organisms, walking motion is realized by the synchronization of artificial neural networks. Moreover, the movement of the micro robot can be controlled by pulse-type hardware neural networks. The locomotion speed is 26.4 mm/min and the step width is 0.88 mm. In the future, other motions exhibited by living organisms will be incorporated into the insect-type MEMS micro robot, by integrating functional elements such as sensors onto a single chip. When the sensor has been integrated, the walking motion systems will form a closed loop, enabling more lifelike behavior of the micro robot. Such micro robots are not only useful in medical fields but will also assist biologists in understanding the movements of living organisms. Moreover, since the all-in-one package installed in the micro robot requires a built-in battery, the micro robot may be used to investigate electromagnetic induction-type wireless power transfer or solar cells.

Acknowledgments Specimen fabrication was supported by Research Center for Micro Functional Devices of Nihon University. This study was supported by the CST research project of Nihon University, and by JSPS KAKENHI (25420226, 23760243).

References Doi, S. Kumagai, S. editors (1962). Nonlinear dynamics of small-scale

biophysical neural networks. Foundations of Integrative Neuroscience, Mary Ann Liebert, page 261-301.

Shibata, T. Aoki, Y. Otsuka, M. Idogaki, T. Hattori, T. (1997). Microwave Energy Transmission System for Microrobot. IEICE transactions on electronics, E80-C(2):303–308.

Habib, M. K. (2011). Biomimetcs: Innovations and. Robotics. International Journal of Mechatronics and. Manufacturing Systems, 4:113–134.

Donald, B. R. Levey, C. G. McGray, C. D. Paprotny, I. Rus, D. (2006). An Untethered, Electrostatic, Globally Controllable MEMS Micro-Robot. Journal of Microelectromechanical Systems, 15:1–15.

Edqvist, E. Snis, N. Mohr, R. C. Scholz, O. Corradi, P. Gao, J. Diéguez, A. Wyrsch, N. Johansson, S. (2009). Evaluation of Building Technology for Mass Producible Millimeter-Sized Robots using Flexible Printed Circuit Boards. Journal of Micromechanics and Microengineering, 19:1–11.

Asada, N. Matsuki, H. Minami, K. Esashi, M. (1994). Silicone Micromachined Two-Dimensional Galvano Optical Scanner. IEEE Transactions on Magnetics, 30:4647–4649.

Surbled, P. Clerc, C. Le Pioufle, B. Ataka, M. Fujita, H. (2001). Effect of the Composition and Thermal Annealing on the Transformation Temperature Sputtered TiNi Shape Memory Alloy Thin Films. Thin Solid Films, 401:52–59.

Nakada, K. Asai, T. Amemiya, Y. (2003). An Analog CMOS Central Pattern Generator for Interlimb Coordination in Quadruped Locomotion. IEEE Transaction on Neural Networks, 14:1356–1365.

Delcomyn, F. (1980). Neural Basis of Rhythmic Behavior in Animals. Science, 210:492–498.

Tsumoto, K. Yoshinaga, T. Aihara, K. Kawakami, H. (2003). Bifurcations in Synaptically Coupled Hodgkin-Huxley Neurons with a Periodic Input. International Journal of Bifurcation and Chaos, 13:653–666.

Tsuji, S. Ueta, T. Kawakami, H. Aihara, K. (2007). Bifurcation Analysis of Current Coupled BVP Oscillators. International Journal of Bifurcation and Chaos, 17:837–850.

Tsumoto, K. Yoshinaga, T. Iida, H. Kawakami, H. Aihara, K. (2006). Bifurcations in a Mathematical Model for Circadian Oscillations of Clock Genes. Journal of Theoretical Biology, 239:101–122.

Hodgkin, A. L. Huxley, A. F. (1952). A Quantitative Description of Membrane Current and its Application to Conduction and Excitation in Nerve. The Journal of Physiology, 117:500–544.

FitzHugh, R. (1961). Impulses and Physiological States in Theoretical Models of Nerve Membrane. Biophysical Journal, 1:445–466.

Endo, T. Mori, S. (1978). Mode Analysis of a Ring of a Large Number of Mutually Coupled van del Pol Oscillators. IEEE Transactions on Circuits Systems, 25:7–18.

Kitajima, H. Yoshinaga, T. Aihara, K. Kawakami, H. (2001). Burst Firing and Bifurcation in Chaotic Neural Networks with Ring Structure. International Journal of Bifurcation and Chaos, 11:1631–1643.

Yamauchi, M. Wada, M. Nishino, Y. Ushida, A. (1999). Wave Propagation Phenomena of Phase States in Oscillators Coupled by Inductors as a Ladder. IEICE Trans. Fundamentals, E82-A:2592–2598.

Yamauchi, M. Okuda, M. Nishino, Y. Ushida, A. (2003). Analysis of Phase-Inversion Waves in Coupled Oscillators Synchronizing at In-and-Anti-Phase. IEICE Trans. Fundamentals, E86-A:1799–1806.

Okazaki, K. Ogiwara, T. Yang, D. Sakata, K. Saito, K. Sekine, Y. Uchikoba, F. (2011). Development of Pulse Control Type MEMS Micro Robot with Hardware Neural Network. Artificial Life and Robotics, 16:229–233.

Uchikoba, F. Takato, M. Saito, K. (2012). Hardware Neural Networks Controlled MEMS Rotational Actuators and Application to Micro

ECAL - General Track

ECAL 2013 552

Robot. Journal of Mechanics Engineering and Automation, 2:499–506.

Sniegowski, J. J. Garcia, E. J. (1996). Surface-micromachined Gear Trains Driven by an On-Chip Electrostatic Microengine. IEEE Electron Device Letters, 17:366–368.

Takeda, M. (2001). Applications of MEMS to Industrial Inspection. In Proceedings of MEMS 2001, pages 182–191.

Habib, M. K. Watanabe, K. Izumi, K. (2007). Biomimetcs robots: From Bio-inspiration to Implementation. In Proceedings of 33rd Annual Conference of the IEEE Industrial Electronics Society, pages 143–148.

Baisch, A. T. Sreetharan, P. S. Wood, R. J. (2010). Biologically-Inspired Locomotion of a 2g Hexapod Robot. In Proceedings of IEEE IROS 2010, pages 5360–5365.

Yan, G. Ye, D. Zan, P. Wang, K. Ma, G. (2007). Micro-Robot for Endoscope Based on Wireless Power Transfer. In Proceedings of 33rd Annual Conference of the 2007 IEEE International Conference on Mechatronics and Automation, pages 3577–3581.

Tang, W. C. Nguyen, T.-C. H. Howe, R. T. (1989). Laterally Driven Polysilicon Resonant Microstructure. In Proceedings of IEEE Micro Electro Mechanical Systems. An Investigation of Micro Structures, Sensors, Actuators, Machines and Robots, pages 53–59.

Suzuki, Y. Tani, K. Sakuhara, T. (1999). Development of a New Type Piezo Electric Micromotor. In Proceedings of Transduceres '99, pages 1748–1751.

Nagumo, J. Arimoto, S. Yoshizawa, S. (1962). An Active Pulse Transmission Line Simulating Nerve Axon. In Proceedings of IRE, pages 2061–2072.

Maeda, Y. (2008). A Hardware Neuronal Network Model of a Two-level Central Pattern Generator. Transactions of the Japanese Society for Medical and Biological Engineering, 46:496–504.

Lewis, A. M. Cummings, E. R. Cohen, H. A. Hartmann, M. (2000). Toward Biomorphic Control Using Custom a VLSI CPG Chips. In Proceedings of International Conference on Robotics & Automaton, pages 494–500.

Suematsu, H. Kobayashi, K. Ishii R. Matsuda, A. Skine, Y. Uchikoba, F. (2009). MEMS Type Micro Robot with Artificial Intelligence System. In Proceedings of International Conference on Electronics Packaging, pages 975–978.

Saito, K. Okazaki, K. Kawakami, T. Matsuda, A. Uchikoba, F. Sekine, Y. (2010). Pulse-Type Hardware CPG Model for MEMS Type Micro Robot. In Proceedings of 2010 IEEJ International Analog VLSI Workshop, pages 219–223.

Bhardwaj, J. K. Ashraf, H. (1995). Advanced Silicon Etching using High-Density Plasmas. In Proceedings of SPIE Micromachining and Microfabrication Process Technology, pages 224–233.

ECAL - General Track

553 ECAL 2013