IN several countries, the increase in the aged population and

Transcription

1 1064 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS PART B: CYBERNETICS, VOL. 42, NO. 4, AUGUST 2012 An EMG-Based Control for an Upper-Limb Power-Assist Exoskeleton Robot Kazuo Kiguchi and Yoshiaki Hayashi Abstract Many kinds of power-assist robots have been developed in order to assist self-rehabilitation and/or daily life motions of physically weak persons. Several kinds of control methods have been proposed to control the power-assist robots according to user s motion intention. In this paper, an electromyogram (EMG)-based impedance control method for an upper-limb power-assist exoskeleton robot is proposed to control the robot in accordance with the user s motion intention. The proposed method is simple, easy to design, humanlike, and adaptable to any user. A neurofuzzy matrix modifier is applied to make the controller adaptable to any users. Not only the characteristics of EMG signals but also the characteristics of human body are taken into account in the proposed method. The effectiveness of the proposed method was evaluated by the experiments. Index Terms Electromyogram (EMG) based control, neurofuzzy modifier, power-assist robot. I. INTRODUCTION IN several countries, the increase in the aged population and the decrease in the working proportion are serious problems. In order to assist self-rehabilitation and/or daily life motions of physically weak persons such as elderly, disabled, or injured persons, many kinds of power-assist robots have been developed [1] [21]. It is important for the power-assist exoskeleton robot to be controlled in accordance with the user s motion intention. Since upper-limb motion is involved in various daily activities, it is particularly important for the upper-limb powerassist exoskeleton robot. In order to activate the robot according to the user s motion intention in real time, the motion of the user or the force between the user and the robot is sometimes observed to understand the motion intention of the user [16], [18], [22], [23]. The inverse model of the exoskeleton robot can be used to derive the joint torques [24]. In these control methods, the user has to be able to initiate the movement before his/her motion is assisted [10]. For rehabilitation, predefined or designated motion or force is basically generated with impedance controller for the user [15], [25]. Skin surface electromyogram (EMG), which is one of the biological signals, is often used in order to control the power- Manuscript received June 14, 2011; revised September 22, 2011 and December 21, 2011; accepted December 26, Date of publication February 10, 2012; date of current version July 13, This paper was recommended by Associate Editor T. H. Lee. K. Kiguchi is with the Department of Advanced Technology Fusion, Saga University, Saga , Japan ( kiguchi@ieee.org). Y. Hayashi is with Saga University, Saga , Japan. Color versions of one or more of the figures in this paper are available online at Digital Object Identifier /TSMCB assist robot according to the user s intention since it directly reflects the user s muscle activity level in real time [2] [11]. The robot can perform the effective power assist if the motion intention of the user is precisely estimated based on the EMG signals in real time. However, realizing the precise motion estimation for the power-assist based on EMG signals is not very easy because of the following: 1) Obtaining the same EMG signals for the same motion is difficult even with the same person since the EMG signal is a biologically generated signal. 2) The activity level of each muscle and the way of using each muscle for a certain motion are a little different between persons. 3) Real-time motion prediction is not very easy since many muscles are involved in a joint motion. 4) One muscle is not only concerned with one motion but also with other kinds of motion. 5) The role of each muscle for a certain motion varies in accordance with joint angles (i.e., in accordance with the upperlimb posture). 6) The activity level of some muscles such as biarticular muscles are affected by the motion of the other joint. 7) The effect of the antagonist muscle must be taken into account. In order to cope with these problems, neurofuzzy EMGbased control (i.e., the combination of adaptive neurocontrol and flexible fuzzy control) has been proposed for upper-limb exoskeleton robots [7] [9]. Although the neurofuzzy control method is effective, fuzzy control rules become complicated if the number of the degrees of freedom (DOF) of the exoskeleton robot is increased. In this paper, an effective EMG-based impedance control method for an upper-limb power-assist exoskeleton robot is proposed. The proposed method is simple, easy to design, humanlike, and adaptable to any user in a short time. The user s joint torque required for the intended motion is estimated based on EMG signals in the proposed method. The proposed method is applicable to any user who generates EMG signals for the intended motion. A neurofuzzy matrix modifier is applied to make the controller adaptable to every upper-limb posture of any users. Not only the characteristics of EMG signals but also the characteristics of human body are taken into account in the proposed method. In order to generate natural and comfortable motion for the user, the user s hand force vector is calculated based on the estimated joint torque. Then, the user s hand acceleration is obtained based on the calculated hand force vector to estimate the user s hand trajectory. Finally, the impedance control is applied to the user s hand trajectory to realize humanlike motion [26], [27]. Since impedance parameters of the human upper limb are changed based on the upper-limb posture [28] [32] and the relationship between agonist and antagonist muscles [32], impedance parameters of the exoskeleton robot are also changed based on them in the proposed method /$ IEEE

2 KIGUCHI AND HAYASHI: EMG-BASED CONTROL FOR AN UPPER-LIMB EXOSKELETON ROBOT 1065 Fig. 2. Location of the EMG electrodes. TABLE II MUSCLES FOR EACH EMG CHANNEL Fig. 1. Structure of the robot. TABLE I MOVABLE RANGE OF EACH JOINT The proposed control method has been applied to a 7DOF upper-limb power-assist exoskeleton robot to assist all upperlimb joint motions of a human user according to the user s motion intention in real time. The effectiveness of the proposed method was evaluated by experiments. II. 7DOF UPPER-LIMB EXOSKELETON ROBOT The 7DOF exoskeleton power-assist robot for the upper limb used in this paper is shown in Fig. 1. The robot consists of seven dc motors with encoders or potentiometers to measure each joint angle. In addition, force/torque sensors (PD3-32, Nitta Corporation) are set in the forearm and wrist parts to measure the force between the user and the robot. This robot is able to assist the most of each human upper-limb joint motion, as shown in Table I. Motors 1, 2, 3, and 4 are the motors for shoulder vertical, shoulder horizontal, shoulder rotation, and elbow motions, respectively. The generated torque by these motors is transferred to the robot joints by pulleys and cable drives. Considering the fact that many physically weak persons use wheelchairs, the robot was designed to be installed on a wheelchair. For the safety and the convenience of users, these motors have been fixed to the back frame of the wheelchair. The other four motors are connected to the robot directly by spur gear. There are three hook-and-loop-fastener parts for the user to wear the robot. One is the upper-arm cover on biceps, another is the forearm cover, and the other is the cover on the palm holder. For shoulder vertical and horizontal motions, a moving center rotation (CR) mechanism has been applied. The main advantage of the CR mechanism is the ability to adjust the distance between the upper-arm holder and the CR of the shoulder joint of the robot in accordance with the shoulder motion automatically with a link-work mechanism, in order to cancel out the ill effects caused by the position difference between the CR of the robot shoulder and that of the human shoulder since the CR of the human shoulder itself is moved with respect to the human body according to the upper-limb posture [7], [9]. The highest priority must be given to safety of the user. Safety components are considered both in the software and the hardware to prevent sudden unexpected motion. In the software, maximum torque and maximum velocity are limited. In the hardware, there are physical stoppers for each joint to limit the joint motion within the movable range of human upper limb. The movable range of the robot and that of the human upper limb (average) is shown in Table I.

3 1066 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS PART B: CYBERNETICS, VOL. 42, NO. 4, AUGUST 2012 Fig. 3. Controller structure. III. EMG-BASED IMPEDANCE CONTROL Sixteen channels of EMG signals are used as main input signals to estimate the user s upper-limb motion intention. The locations of EMG electrodes are shown in Fig. 2. Each channel mainly corresponds to one muscle, as shown in Table II. The forearm force (i.e., generated force between the robot and the forearm of the user), the hand force (i.e., generated force between the robot and the hand of the user), and the forearm torque (i.e., generated torque between the wrist holder of the robot and the forearm of the user) are used as subordinate input signals for the controller. The structure of the controller is shown in Fig. 3. In this control method, the user s upperlimb joint torque vector is estimated based on EMG signals and/or force/torque sensor signals at first. Then, the user s hand force vector is calculated using the estimated joint torque vector. Finally, the impedance control is applied to realize the user s hand force vector. In the case of the force/torque sensorbased control (i.e., when the user does not activate the muscles actively), the robot only follows a user s upper limb (without the power assist) so that the robot does not encumber the user s motion. Since raw EMG signals are not suitable as input signals to the controller, the feature of the raw signal must be somehow extracted. In order to extract the feature of the raw EMG signal, the root mean square (RMS) of the EMG signal is calculated and used as an input for the controller in this paper. The RMS calculation is written as RMS = 1 N N vi 2 (1) i=1 where N is the number of the segments (N = 400) and v i is the voltage at ith sampling. The sampling frequency in this paper is 2 khz. The joint torque of the user is estimated based on EMG signals and/or force/torque sensor signals. The signals used for the joint torque estimation are automatically switched according to the user s EMG signal levels [5]. When the user s EMG signal level is high, EMG signals are applied to estimate the user s motion to be assisted by the robot. On the other hand, force/torque sensor signals are applied to estimate the Fig. 4. Membership function. user s motion when the user s muscle activation levels are low since the noise ratio in the EMG signals is increased. When the user s EMG signal levels are intermediate, both the EMG signals and the force/torque sensor signals are simultaneously applied. The EMG levels of corresponding muscles for each joint motion are used to switch the signals for each joint motion, as shown in Fig. 3. The membership functions used to switch the signals are shown in Fig. 4. Thus, as a monitored muscle increases the activity level, the signals used for the joint torque estimation are gradually switched from the force/torque sensorbased estimation to the EMG-based estimation. The relationship between the 16 EMG RMSs and the joint torque vector used for the EMG-based estimation is modeled as τ 1 τ 2 w 11 w 12 w 115 w 116 ch 1 τ 3 w 21 w 22 w 215 w 216 ch τ est = τ 4 = τ 5 w 61 w 62 w 615 w 616 ch (2) 15 τ 6 w 71 w 72 w 715 w 716 ch 16 τ 7 where τ est is the joint torque vector and τ 1 τ 7 are the joint torques for each joint motion, as shown in Table I. w ij is the weight value for the jth EMG signal to estimate the joint torque τ i, and ch j represents the RMS value of the EMG signal measured in channel j. The weight matrix (i.e., the muscle-model matrix) in (2) can be roughly defined using the knowledge of human upper-limb anatomy or the results of preliminal experiments. In this paper, the initial weight matrix was set considering the relationship between the muscle and the direction of the rotation of the joint. Therefore, the joint torque

4 KIGUCHI AND HAYASHI: EMG-BASED CONTROL FOR AN UPPER-LIMB EXOSKELETON ROBOT 1067 Fig. 6. Membership functions in the fuzzifier layer. Fig. 5. Neurofuzzy modifier. vector generated by the muscle force can be calculated if every weight for the EMG signals is properly defined. It is not easy, however, to define the proper weight matrix for the user from the beginning because of personal difference. Furthermore, the posture of the upper limb affects the relationship between the EMG signals and the generated joint torques because of anatomical reasons such as the change of the moment arm. Consequently, the effect of the posture difference of the upper limb must be taken into account to estimate the precise upperlimb motion for the power assist. Therefore, a neurofuzzy muscle-model matrix modifier is applied to take into account the effect of the upper-limb posture difference of the user. The neurofuzzy modifier outputs the coefficient for each weight of the muscle-model matrix shown in (2) to modify the weight matrix in real time based on the upper-limb posture of the user. The coefficient is used to adjust the weight matrix in an online manner by multiplying the weight by the coefficient in accordance with the upper-limb posture of the user, so that the effect of upper-limb posture difference can be effectively compensated. It also makes the same effect of adjusting the weight matrix itself to be suitable for each user. The structure of the neurofuzzy modifier is the same as a neural network, and the process of the signal flow in the neurofuzzy modifier is the same as that in fuzzy reasoning. The neurofuzzy modifier consists of five layers (input, fuzzyfier, rule, defuzzifier, and output layers), as shown in Fig. 5. Input variables to the neurofuzzy modifier are joint angles of the user s upper limb. Three fuzzy linguistic variables are prepared for each joint angle in the fuzzifier layer. The sigmoidal functions f s and the Gaussian functions f g are used as membership functions in the fuzzifier layer, as shown in Fig. 6. Output variables from the neurofuzzy modifier are coefficients for components of the weight matrix. In the rule layer, the coefficient for each weight is reasoned in any combination of upper-limb joint angles. The neurofuzzy modifier is trained to adjust its output for each user during the training period before operation. Initially, the output of the neurofuzzy modifier is set to be 1.0 for every weight of the muscle-model matrix. During the training, the robot is supposed to generate the same motion as the estimated user s motion in real time. The error-backpropagation learning algorithm is applied to minimize the squared error functions in order to eliminate the error of the muscle-model matrix. If the muscle-model matrix with the neurofuzzy modifier correctly works, the generated robot motion and the user s motion are supposed to be the same. The amount of the error can be provided by the force/torque sensors, since the force/torque sensors measure force/torque caused by the motion difference between the user and the robot. Therefore, the squared error function used in this paper is written as E = 1 2 f 2 err (3) where E is the error function to be minimized and f err is the measured force/torque between the user and the robot. The output from the force/torque sensors becomes zero if the generated robot motion and the user s motion are the same. After the neurofuzzy modifier is properly trained, the joint torque vector of the user can be estimated using the musclemodel matrix with the neurofuzzy modifier. If the user cannot move his/her upper limb properly, although he/she can generate EMG signals, a motion indicator, which is manipulated by the movable part of the user, can be used to prepare the amount of motion error for the learning [33]. In order to estimate the user s motion intention, hand force vector calculation is carried out based on the estimated joint torque vector. The estimated joint torque vector is transferred to the hand force vector of the user using the Jacobian matrix as follows: F hand = J T τ est (4) F h,avg = 1 N f N f F hand (k) (5) k=1

5 1068 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS PART B: CYBERNETICS, VOL. 42, NO. 4, AUGUST 2012 where F hand is the hand force vector (6-D vector) of the user, J is the Jacobian matrix, and F h,avg is the average of F hand in N f number of samples. Then, the desired hand acceleration vector is calculated from (5), i.e., Ẍ d = M 1 F h,avg (6) where Ẍd is the desired hand acceleration vector (6-D vector) and M is the mass matrix of the robot and the user s upper limb. To realize the user s intended motion, the following impedance control is applied to obtain the resultant hand force vector F : F = MẌd + B(Ẋd Ẋ)+K(X d X) (7) Fig. 7. Experimental setup. TABLE III CONSTANT IMPEDANCE PARAMETER VALUES where Ẋd and X d are the desired hand velocity vector and position vector, which are calculated from (6), respectively. B and K are the viscous coefficient matrix and the spring coefficient matrix, respectively. Since impedance parameters of the human upper limb are changed based on the upper-limb posture and the relationship between agonist and antagonist muscles, impedance parameters of the exoskeleton robot are also changed based on them in the proposed method in order to realize natural and comfortable power assist. Therefore, the impedance parameter matrix B and K in (7) are changed depending on the upper-limb posture and the activity levels of activated upper-limb antagonist muscles in real time. Conclusively, the joint torque command vector for the seven dc motors is calculated as follows: τ motor = κj T F (8) where τ motor is the joint torque command vector and κ is the power-assist rate. The impedance parameters B and K are adjusted as follows: B = m B l B B 0 (9) K = m K l K K 0 (10) where K 0 and B 0 are initial spring and viscous coefficient diagonal matrices for the initial position when EMG levels of muscles used to estimate impedance are in the defined level. l K and l B are the diagonal matrices effected from the upper-limb posture. The results of other researches [28] [32] have been applied to define the matrices. m K and m B are the coefficients affected by the EMG levels of agonist and antagonist muscles. The effect from EMG signals can be written as m K = λ K rate ch6 rate ch8 (11) m B = λ B rate ch6 rate ch8 (12) where λ K and λ B are the coefficients of the EMG effects, rate ch6 and rate ch8 are ratios between RMS values of channels 6 (bicep-long head) and 8 (tricep-lateral head), and initial RMS values. Thus, the amount of impedance parameters B and K is increased when the activity level of both agonist and antagonist muscles is simultaneously increased [32]. IV. EXPERIMENT RESULTS In order to evaluate the effectiveness of the proposed control method, experiments have been carried out. In the experiments, three healthy young male subjects (A: 23 years old; B: 27 years old; and C: 28 years old) performed the same upper-limb motions with the assist of the exoskeleton robot. For the first and second experiments, the effect of the adjustment of impedance parameters was evaluated. In the first experiment, the subjects were instructed to grasp a spoon and move the spoon between trays 1 and 2, as shown in Fig. 7(a). The experiments were performed with three kinds of constant impedance parameters (lowest, middle, and highest), as shown in Table III, and with adjusting impedance parameters. In these experiments, κ is set to be 1.5. Therefore, the muscle activity levels (i.e., the amount of RMS values) are supposed to be reduced for the same motion if the impedance parameters are properly defined. The experiment results of subject A are shown in Fig. 8. Fig. 8(a) (d) show the elbow angle, the wrist palm flexion/extension angle, and the RMS values of channels 6 (bicep-long head) and 13 (flexor carpi radialis) in each impedance parameters. In addition, examples of the change of weight values in (2) are shown in Fig. 9. These weight values are the results in the case of Fig. 8. One can see that the weight values were changed in real time depending on the posture of the subject s upper limb. In the experiment with the adjusting impedance parameters [see Fig. 8(d)], the parameters are adjusted based on the proposed method, as shown in Fig. 8(e) and (f). Here, the impedance parameters are adjusted relatively close to the lowest constant parameters in the experiment shown in Fig. 8(d). It can be observed that the experimental results with the adjusting impedance parameters show the best results among them. The average amounts of the RMS values in Fig. 8(a) (c) are 107%, 114%, and 123% with respect to that in Fig. 8(d), respectively. Similar results were obtained with another subject. In the second experiment, the subjects lifted a dumbbell (5 kg), as shown in Fig. 7(b). The experiments were also performed with three kinds of constant impedance parameters (lowest, middle, and highest) and with adjusting impedance

6 KIGUCHI AND HAYASHI: EMG-BASED CONTROL FOR AN UPPER-LIMB EXOSKELETON ROBOT 1069 Fig. 8. Experimental results when subject A grasped a spoon and moved the spoon between trays 1 and 2, as shown in Fig. 7(a), with (a) low parameters, (b) middle parameters, (c) high parameters, (d) adjusting parameters, (e) spring coefficient, and (f) viscous coefficient. Fig. 9. Examples of the change of weight values in (2). parameters. In these experiments, κ is set to be 1.5. The muscle activity levels are supposed to be reduced for the same motion if the impedance parameters are properly adjusted. Since the subjects have a heavy object, the impedance parameters are supposed to be increased because of the effect of the antagonist muscles. The experimental results of subject A are shown in Fig. 10. Fig. 10(a) (d) show the elbow angle and the RMS value Fig. 10. Experimental results when subject A lifted a dumbbell (5 kg), as shown in Fig. 7(b), with (a) low parameters, (b) middle parameters, (c) high parameters, (d) adjusting parameters, (e) spring coefficient, and (f) viscous coefficient. of channel 6 (bicep-long head) in each impedance parameter. In the experiment with the adjusting impedance parameters [see Fig. 10(d)], the parameters are adjusted based on the proposed method, as shown in Fig. 10(e) and (f). Here, the impedance parameters are adjusted between the middle and highest constant parameters. It can be observed that the experimental results with the adjusting impedance parameters show the good results among them. The average amount of the RMS values in Fig. 10(a) (c) are 126%, 117%, and 129% with respect to that in Fig. 10(d), respectively. The similar results were obtained with another subject. Thus, the first and second experimental results show that the power assist with adjusting impedance parameters can reduce the amount of the RMS values for the same motion compared with that with constant impedance parameters. For the third experiment, the effect of power-assist rate was verified. In this experiment, the subjects were instructed to grasp a cup and move the cup to the mouth, as shown in Fig. 7(c). The experiments were performed with four kinds of power-assist rates (κ =0.5, 1.0, 1.5, and 2.0). If the power

7 1070 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS PART B: CYBERNETICS, VOL. 42, NO. 4, AUGUST 2012 Fig. 12. Experimental results when subject B grasped a cup and moved the cup, as shown in Fig. 7(c), with (a) κ =1.0and (b) κ =1.5. Fig. 11. Experimental results when subject A grasped a cup and moved the cup, as shown in Fig. 7(c), with (a) κ =0.5, (b)κ =1.0, (c)κ =1.5, and (d) κ =2.0. assist is properly performed, the amount of RMS values is supposed to be reduced for the same motion. The experimental results of subject A are shown in Fig. 11. In Fig. 11(a) (d), the experimental results with the power-assist rate of 0.5, 1.0, 1.5, and 2.0 are depicted, respectively. In the case of κ =1.0, the robot just follows the user s motion and does not perform the power assist. In the cases of κ =1.5 and 2.0, the robot performs power assist. On the other hand, in the case of κ =0.5, the robot moves slower than the user. That means the robot rather disturb the user s motion. In the experimental results with κ = 1.5 [see Fig. 11(c)], the amount of the RMS values are smaller than that with κ = 1.0 [see Fig. 11(b)]. This means that proper power assist is carried out by the robot. In the case of κ = 2.0 [see Fig. 11(d)], the amount of RMS values are almost the same as that with κ =1.0 [see Fig. 11(b)]. That means that the power assist with the overlarge power-assist rate might make the negative effect. Comparing Fig. 11(a) and (b), the amount of the RMS values in Fig. 11(a) is larger than that in Fig. 11(b). That means that the power assist with the too-small powerassist rate disturbs the user s motion. The average amount of the RMS values in Fig. 11(a), (c), and (d) are 143%, 67%, and 89% with respect to that in Fig. 11(b), respectively. These results show that proper power assist can be realized with the proper amount of power-assist rate. In the case of κ =1.0, the robot provides the user with the load for the training. On the other hand, in the case of κ =1.0, the robot provides the power assist for the user. Fig. 12 shows the experimental results of another subject (subject B). These results show that the proposed control method can be applicable to any users. V. C ONCLUSION In this paper, an effective and adaptable EMG-based impedance control method has been proposed for an upperlimb power-assist exoskeleton robot. The user s joint torque required for the intended motion has been estimated based on EMG signals in the proposed method. A neurofuzzy musclemodel matrix modifier has been applied to make the controller adaptable to every upper-limb posture of any users. In order to generate natural and comfortable motion for the user, the user s hand force vector has been calculated based on the estimated joint torque, and then the estimated user s hand motion has been realized with the impedance control. Since impedance parameters of the human upper limb are changed based on the upper-limb posture and the relationship between agonist and antagonist muscles, impedance parameters of the exoskeleton robot have been also changed based on them in the proposed method. Experimental results have shown the effectiveness of the proposed control method. REFERENCES [1] E. Guizzo and H. Goldstein, The rise of the body bots, IEEE Spectr., vol. 42, no. 10, pp , Oct [2] C. J. Yang, J. F. Zhang, Y. Chen, Y. M. Dong, and Zhang Y., A review of exoskeleton-type systems and their key technologies, Proc. IMechE, J.Mech.Eng.Sci.,C, vol. 222, no. 8, pp , [3] B. Dellon and Y. Matsuoka, Prosthetics, exoskeletons, and rehabilitation (grand challenges of Robotics), IEEE Robot. Autom. Mag.,vol. 14,no.1, pp , Mar [4] M. A. Oskoei and H. Hu, Myoelectric control systems A survey, Biomed. Signal Process. Control, vol. 2, no. 4, pp , Oct [5] K. Kiguchi, S. Kariya, K. Watanabe, K. Izumi, and T. Fukuda, An exoskeletal robot for human elbow motion support Sensor fusion, adaptation, and control, IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 31, no. 3, pp , Jun [6] J. Rosen, M. Brand, M. Fuchs, and M. Arcan, A myosignal-based powered exoskeleton system, IEEE Trans. Syst., Man, Cybern. A, Syst., Humans, vol. 31, no. 3, pp , May [7] K. Kiguchi, T. Tanaka, and T. Fukuda, Neuro-fuzzy control of a robotic exoskeleton with EMG signals, IEEE Trans. Fuzzy Syst., vol. 12, no. 4, pp , Aug [8] K. Kiguchi, R. Esaki, and T. Fukuda, Development of a wearable exoskeleton for daily forearm motion assist, Adv. Robot., vol. 19, no. 7, pp , Aug [9] K. Kiguchi, M. H. Rahman, M. Sasaki, and K. Teramoto, Development of a 3DOF mobile exoskeleton robot for human upper limb motion assist, Robot. Autonom. Syst., vol. 56, no. 8, pp , Aug [10] C. Fleischer and G. Hommel, A human -exoskeleton interface utilizing electromyography, IEEE Trans. Robot., vol. 24, no. 4, pp , Aug

8 KIGUCHI AND HAYASHI: EMG-BASED CONTROL FOR AN UPPER-LIMB EXOSKELETON ROBOT 1071 [11] E. E. Cavallaro, J. Rosen, J. C. Perry, and S. Burns, Real-time myoprocessors for a neural controlled powered exoskeleton arm, IEEE Trans. Biomed. Eng., vol. 53, no. 11, pp , Nov [12] H. Kobayashi, T. Shiiban, and Y. Ishida, Realization of all 7 motions for the upper limb by a muscle suit, J. Robot. Mechatron., vol. 16, no. 5, pp [13] D. Sasaki, T. Noritsugu, and M. Takaiwa, Development of pneumatic soft hand for human friendly robot, J. Robot. Mechatron., vol. 15, no. 2, pp , [14] J. L. Pons, E. Rocon, A. F. Ruiz, and J. C. Moreno, Upper-limb robotic rehabilitation exoskeleton: Tremor suppression, in Rehabilitation Robotics, S. S. Kommu, Ed. Rijeka, Croatia: InTech, 2007, pp [15] N. G. Tsagarkis and D. G. Caldwell, Development and control of a Soft- Actuated exoskeleton for use in physiotherapy and training, Autonom. Robots, vol. 15, no. 1, pp , Jul [16] K. Kong and D. Jeon, Design and control of an exoskeleton for the elderly and patients, IEEE/ASME Trans. Mechatron., vol. 11, no. 4, pp , Aug [17] S. Lee and Y. Sankai, Power assist control for walking aid with HAL-3 based on EMG and impedance adjustment around knee joint, in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 2002, pp [18] K. Nagai and I. Nakanishi, Force analysis of exoskeletal robotic orthoses for judgment on mechanical safety and possibility of assistance, J. Robot. Mechatron., vol. 16, no. 5, pp , [19] L. Lucas, M. DiCicco, and Y. Matsuoka, An EMG-controlled hand exoskeleton for natural pinching, J. Robot. Mechatron., vol. 16, no. 5, pp , [20] A. Tsukahara, R. Kawanishi, Y. Hasegawa, and Y. Sankai, Sit-to-stand and stand-to-sit transfer support for complete paraplegic patients with robot suit HAL, Adv. Robot., vol. 24, no. 11, pp , [21] E. A. Brackbill, Y. Mao, S. K. Agrawal, M. Annapragada, and V. N. Dubey, Dynamics and control of a 4-dof wearable cable-driven upper arm exoskeleton, in Proc. IEEE Int. Conf. Robot. Autom., 2009, pp [22] H. Kaminaga, T. Amari, Y. Niwa, and Y. Nakamura, Electro-hydrostatic actuators with series dissipative property and their application to power assist devices, in Proc. 3rd IEEE RAS EMBS Int. Conf. Biomed. Robot. Biomechatron., 2010, pp [23] H. Kazerooni and S. L. Mahoney, Dynamics and control of robotic systems worn by humans, Trans. ASME, J. Dyn. Syst., Meas. Control, vol. 113, no. 3, pp , Sep [24] H. Kazerooni and R. Steger, The Berkeley lower extremity exoskeleton, Trans. ASME, J. Dyn. Syst., Meas. Control, vol. 128, no. 1, pp , Mar [25] J. F. Veneman, R. Kruidhof, E. E. G. Hekman, R. Ekkelenkamp, E. H. F. Van Asseldonk, and H. van der Kooij, Design and evaluation of the LOPES exoskeleton robot for interactive gait rehabilitation, IEEE Trans. Neural Syst. Rehabil. Eng., vol. 15, no. 3, pp , Sep [26] T. Flash and N. Hogan, The coordination of arm movements: An experimentally confirmed mathematical model, J. Neurosci., vol. 5, no. 7, pp , Jul [27] L. P. J. Selen, D. W. Franklin, and D. M. Wolpert, Impedance control reduces instability that arises from motor noise, J. Neurosci., vol. 29, no. 40, pp , Oct [28] F. A. Mussa-Ivaldi, N. Hogan, and E. Blzzl, Neural, mechanical, and geometric factors subserving arm posture in humans, J. Neurosci., vol. 10, no. 5, pp , Oct [29] T. Tsuji, P. G. Morasso, K. Goto, and K. Ito, Human hand impedance characteristics during maintained posture, Biol. Cybern., vol. 72, no. 6, pp , [30] H. Gomi and M. Kawato, Equilibrium-point control hypothesis examined by measured arm stiffness during mutijoint movement, Science, vol. 272, no. 5258, pp , Apr [31] R. Osu and H. Gomi, Multijoint muscle relation mechanisms examined by measured human arm stiffness and EMG signals, J. Neurophysiol., vol. 81, pp , [32] F. Popescu, J. M. Hidler, and W. Z. Rymer, Elbow impedance during goal-directed movement, Exp. Brain Res., vol. 152, no. 1, pp , Sep [33] K. Kiguchi, Y. Imada, and M. Liyanage, EMG-based neuro-fuzzy control of a 4DOF upper-limb power-assist exoskeleton, in Proc. IEEE Int. Conf. Eng. Med. Biol. Soc., 2007, pp Kazuo Kiguchi received the B.Eng. degree in mechanical engineering from Niigata University, Niigata, Japan, in 1986, the M.A.Sc. degree in mechanical engineering from the University of Ottawa, Ottawa, Canada, in 1993, and the Dr.Eng. degree from Nagoya University, Nagoya, Japan, in He was a Research Engineer with Mazda Motor Company between 1986 and 1989, and with MHI Aerospace Systems Company between 1989 and Between 1994 and 1999, he was with the Department of Industrial and Systems Engineering, Niigata College of Technology, Niigata, Japan. He is currently a Professor with the Department of Advanced Technology Fusion, Graduate School of Science and Engineering, Saga University, Japan. His research interests include biorobotics, human-assist robots, rehabilitation robots, and application of robotics in medicine. Dr. Kiguchi is a member of the IEEE Robotics and AUtomation Systems, Man, and Cybernetics; and Engineering in Medicine and Biology Societies. He is also a member of the Japan Society of Mechanical Engineers (JSME), the Robotics Society of Japan, the Society of Instrument and Control Engineers, the Japan Society of Computer-Aided Surgery, the Virtual Reality Society of Japan, and the Japanese Society for Clinical Biomechanics and Related Research. He was a recipient of the J. F. Engelberger Best Paper Award at World Automation Congress 2000, the Toshio Fukuda Award at the IEEE International Conference on Advanced Mechatronics 2008, and the JSME Funai Award in Yoshiaki Hayashi received the Dr.Eng. degree from Kyushu University, Higashi, Fukuoka, Japan, in He is currently with Saga University, Saga, Japan.