Brain-Controlled Exoskeleton Robot for BMI Rehabilitation

Transcription

1 th IEEE-RAS International Conference on Humanoid Robots Nov.29-Dec.1, Business Innovation Center Osaka, Japan Brain-Controlled Exoskeleton Robot for BMI Rehabilitation Tomoyuki Noda ATR Computational Neuroscience Labs Seika-cho, Hikaridai, Souraku-gun Kyoto, Japan t noda@atr.jp Norikazu Sugimoto NICT, Center for Information and Neural Networks xsugi@nict.go.jp Junichiro Furukawa ATR Computational Neuroscience Labs, Graduate school of Frontier Biosciences, Osaka University, furukawa@atr.jp Masa-aki Sato Dept. of Computational Brain Imaging, ATR Dynamic Brain Imaging Labs masa-aki@atr.jp Sang-Ho Hyon ATR Computational Neuroscience Labs, Faculty of Science and Engineering, Ritsumeikan University, gen@fc.ritsumei.ac.jp Jun Morimoto ATR Computational Neuroscience Labs xmorimo@atr.jp Abstract In this paper, we introduce our attempt to develop an assistive robot system which can contribute to Brain- Machine Interface (BMI) rehabilitation. For the BMI rehabilitation, we construct a Electroencephalogram(EEG)-Exoskeleton robot system, where the exoskeleton robot is connected to the EEG system so that the users can control the exoskeleton robot by using their brain activities. We use a classification method which considers covariance matrices of measured EEG signals as inputs to decode brain activities. The decoded brain activities are used to control exoskeleton movements. In this study, we consider assisting the stand-up movement which is one of the most frequently appeared movements in daily life and also a standard movement as a rehabilitation training. To assist the stand-up movement, we develop a force control model which takes dynamics of tendon string into account for the pneumaticelectric hybrid actuation system used in our exoskeleton robot. The results show that the exoskeleton robot successfully assisted user stand-up movements, where the assist system was activated by the decoded brain activities. I. INTRODUCTION Since many countries are facing the aging population problem, development of an exoskeleton robot which can be used to assist user movements is becoming important research topic [1], [2]. In particular, these exoskeleton robot can be used as prosthetic devices for patients such as stroke patients and spinal cord injury patients in rehabilitation programs [3], [4], [5], [6], [7]. In recent years, it has been found that using brain activity to control robotic assistive system can be useful for stroke patients to enhance recovery of motor functions [8]. Therefore, there would be also a possibility to enhance recovery of lower limb motor functions if an exoskeleton robot to assist lower body can be controlled by brain activities. This rehabilitation approach is called Brain-Machine Interface (BMI) rehabilitation [9]. (Switch assist mode) Fig. 1. EEG-Exoskeleton System. EEG signals are detected by active electrodes (BIOSEMI) and the detected brain activities are amplified and converted to digital signals with 2048 Hz sampling rate. The sampled data are transmitted to computers and the received data are decoded to generate control command for the exoskeleton robot. The generated command is sent to the control system of the exoskeleton robot. According to the control command, the exoskeleton robot activates the assist control system. For BMI rehabilitation, we develop a Electroencephalogram(EEG)-Exoskeleton robot system, where the exoskeleton robot is connected to the EEG system so that the users can control the exoskeleton robot by using their brain activities (See Fig. 1). We use a classification method which considers covariance matrices of measured EEG signals as inputs to decode brain activities. The decoded brain activities are used to control exoskeleton movements. In this study, we consider assisting the stand-up movement which is one of the most frequently appeared /12/$ IEEE 21

2 movements in daily life activities and also a standard task movement as a rehabilitation training. To assist the stand-up movement, we develop a force control model which takes dynamics of tendon string into account for the pneumatic-electric hybrid actuation (PEHA) system used in our exoskeleton robot (See Fig. 2) [10]. To test the EEG-robot system and to evaluate the tendon-spring model, we also developed an one degree-of-freedom (DOF) pneumatic-electric hybrid actuation system (See Fig. 3). We show that a user was able to control the EEG-oneDOF system using subject s brain activities. Furthermore, the results show that the exoskeleton robot successfully assisted the stand-up movements, where the assist system was controlled by the decoded brain activities. We also show that the torque control performance can be greatly improved if we explicitly consider dynamics of the tendon connected to the air muscle. This paper is organized as follows. In Section II, the classification method to decode brain activities to control robots is introduced. In Section III, the tendon string modeled as a spring to improve the force control performance is presented. In Section IV, we show the control performance of the developed EEG-oneDOF and EEG-Exoskeleton systems. Fig. 2. Exoskeleton Robot (XoR)[10]. Height: 1.5 m, Weight: 30 kg. XoR has ten degrees of freedom and six active joints. Each active joint uses a hybrid actuator composed of air muscle and an electric motor. XoR is designed to assist lower-limb movements in humans. Fig. 3. One DOF System. (Left) Upward state (Right) Downward state. II. DECODING BRAIN ACTIVITIES To decode brain activities, we use a classification method proposed by [11], [12] because the method classifies matrices with spectral l 1 -norm regularization which can lead to good generalization performance. In addition, the decoder can be efficiently derived by solving a convex optimization problem [11]. As suggested in [11], we use the covariance matrices of the measured EEG signals C as the input variables. In this classification method, The output probabilities of the twoclass classification problem are represented as: P (q t = +1 C t ) = P (q t = 1 C t ) = exp( a t ), (1) exp( a t ) 1 + exp( a t ), (2) where q t denotes the class label. The log odds or logit [13] is modeled as a liner function of the input C: a t = ln P (y = +1 X) P (y = 1 X) = tr [ W C t ] + b (3) Here W is a parameter matrix and b is a bias. A. Learning Classifier To construct the classifier, we consider minimizing a objective function: n min l(z t ) + λ W 1, (4) t=1 z t = q t a t, (5) where λ is the regularization constant and each term of the objective function is represented as: l(z t ) = ln(1 + exp( z t )) (6) r W 1 = σ i [W], (7) i=1 where σ i [W](i = 1,..., r) is the i-th singular value of a matrix W and r is the rank of W. This optimization problem can be efficiently solved by considering an equivalent linear matrix inequality (LMI) problem[11]. B. Online Decoding We decode brain activities by using the classification method introduced above after preprocessing measured EEG data. Figure 4 shows the processing procedure. A band-pass filter with the the frequency 7 30 Hz are applied to the measured EEG signals. The filtered signals are down-sampled with 128 Hz. Laplace filter and common average subtraction are applied for removing voltage bias. The covariance matrix of the processed data is used as the input variable for the classifier. we update a covariance matrix C t of filtered EEG signal at every time step (t = 1, 2, ): { x C t = t x t (t = 1) 1 N x t x t + N 1 N C. (8) t 1 (t 2) Here, x t R 1 D is a filtered EEG signal at time t (D = 64). By using estimated weight matrix W R D D and bias constant b R, we estimate a probability P (q t = +1 C t ) and P (q t = 1 C t ) as in Eqs. (1) and (2). Next, we explain how to select a control command g t = {up, down} which is given to the robot. Since input variables are based on measured EEG signals, the output 22

3 PC (ART Linux) AD/DA Energy Torque UDP Valve PAM1 + MFB MD Motor + - Valve PAM2 EC Quadrature encoded pulse τ θ Fig. 4. EEG signal processing procedure. A band-pass filter with the the frequency 7 30 Hz are applied to the measured EEG signals. The filtered signals are down-sampled with 128 Hz. Laplace filter and common average subtraction are applied for removing bias voltage on electrodes. The covariance matrix of the processed data is used as the input variable for the classifier. We used the classifier with the selected regularization parameter λ = 14. of the classifier can be fluctuated. Therefore, we employed following hysteresis: up (P (q t = +1 C t ) > P threshold ) g t = down (P (q t = 1 C t ) > P threshold ), (9) g t 1 otherwise where up denotes upward state of the onedof system or stand-up state of the exoskeleton robot, and down denotes downward state of the onedof system or sit-down state of the exoskeleton robot. We set the threshold as P threshold = 0.7. According to the output of the classification, upward/stand-up or downward/sit-down movements on the onedof system or the exoskeleton robot, respectively, are generated by using the torque control method introduced in the next section. III. MOVEMENT ASSIST STRATEGY A. Torque controller for Pneumatic-Electric Hybrid Actuator While many exoskeletons are controlled by position based controller, the torque based control is suitable to our application because our exoskeleton eventually control the interaction force to assist human user movements. In this section, we introduce torque controller to our EEG-exoskeleton system as the feed-forward torque model, and fit parameters in a system identification phase. While the Pneumatic Artificial Muscle (PAM) is very light-weight, it can generate large force by converting pressured gas energy into the contraction force thought its rubber tubes. The force generation principal is the path contraction of the spiral fiber expansion embedded by the pneumatic bladder. 1) PEHA onedof system: The Figure 5 shows the essential onedof Pneumatic-Electric Hybrid Actuator (PEHA) system. The PEHA is one implementation of the distributed Fig. 5. Essential PEHA onedof system: the interface board, Multi Function Board (MFB) connected to the PC, controls actuators (Valves, Motor Driver (MD), and PAMs) and reads sensors (encoder and Load Cells (LC) as the arm angle and force). In the picture, the upper PAM is the flexor muscle (PAM1) and the lower is the extensor muscle (PAM2). Note that PAM2 is antagonistic and generates opposed force. Micro-Mini Actuation (DM 2 ) developed in [14]. However, in the large force operation, we found that the torque control is difficult because the wire extension causes large error of PAM force generation. The motor consequently cannot generate large torque to cover it. This problem is detailed later section, and we propose the improved PAM force model and develop better PEHA/DM 2 torque controller. We newly developed the ARM-based Multi Function Board (ARM cortex-m3 processor core equipped with 16bit AD, 16bit DA, Quadrature Decoder, and IOs) to deal with all interfaces of PEHA by only one Ethernet cable. The Figure 6 shows how the each actuator (PEHA)[10] effects to the joint. The PAMs force f is converted to the torque through the wires and the pulley. τ P AMs = (f P AM1 f P AM2 ) r 0, (10) where r 0 is the pulley radius and is constant setup in this system. 2) Motor torque: The motor torque can be transmitted in parallel, e.g., small torque is transferred through a mechanical belt. The PAMs excel at generating DC or low frequency torque, and additionally small electric motor covers error of τ P AMs as quick and high frequent torque but can be small. τ = τ P AMs + τ motor. (11) 3) Light weight tendon: From the PAM, the wire transfers the force into the pulley and drives the joint toward one direction. Vectran, the manufactured fiber (made of liquid crystal polymer fiber), was adopted because it is light-weight, strong, and flexible compared to the other method (metal wire, chain, or mechanical belt). Thus, the PEHA coordinates two different specific actuators through these mechanism. B. PAM model The pneumatic actuator, including PAMs, wires, and pulleys, have a lot in common with human s muscle and is frequently called air muscle. During our exoskeleton 23

4 supports human weight, we found the fiber is also similar in nature to a tendon. To discriminate from other mechanical terminology such as wires embedded in PAMs, we use this force transmitter as tendon in this section. Thus, we propose the tendon-spring equilibrium model for torque controller. Note that the tension is generated toward only one direction. We derive the torque controller for τ P AMs with the equilibrium assumption. Suppose that, using proportional pressure control valves, the pressure p can be controlled by closed feedback loop and enough stable. Although valve pressure and PAMs pressure has pneumatic dynamics, at a curtain time constant later, the PAM contracts until the force equilibrates to the external kinetic constraints such as weight of robot and human. Air circuit dynamics is small and can be ignored with quasi-static movements. At this equilibrium point, the force generation depends on the internal pressure and contraction rate, and the PAM force model[15], [16] is given by f = πd2 0p 4 { 3 tan 2 ψ 0 (1 ε) 2 1 sin 2 ψ 0 }, (12) where ε is contraction rate of PAM, and D 0 and ψ 0 are the PAM diameter and the angle of the embedded spiralwire at the atmosphere pressure. Unlike an air cylinder, the remarkable fact is the nonlinear torque altered by the joint angle changes. Under the assumption that the kinetic constrain is always unchanged, PAMs pressure gives always same equilibrium point ε (p). Inverse models can be learned by machine leaning methods. For example, Hartmann et. al. proposed kinetic dynamics learning [17]. However, this assumption can t be kept in our exoskeleton application because different interaction forces alter kinetic constrain dynamically and equilibrate as task-dependent interaction force F (e.g., ε (p, F)). Thus, in exoskeleton application, the interest is to control PAM force. C. Tendon-spring equilibrium model To sustain human (such as 60[kg] weight), the PAM generate its tension 3000[N] typically (5000[N] maximum theoretically in our system). We found this large forces bring an another problem that the tendon is extended, which also alters the equilibrium point to different ε by this large force. It is mechanically difficult to measure this contraction directly. The measurable is only the apparent contraction rate ε = r 0 θ/l 0, where L 0 is initial PAM length. Here we consider linear tendon-spring model. f = k ε, (13) where ε is the tendon extension corresponding to the extra contraction of the PAM caused by force f, and k is the spring constant. From Eq.12, without tendon-spring model, the force is modeled by quadratic expression with the three parameters (while the two dependent): where f = g(ε, p) = p(aε 2 + bε + c), (14) a = 3πD2 0 4 tan 2, b = 3πD2 0 ψ 0 2 tan 2, ψ ( 0 ) c = πd tan 2 ψ 0 sin 2. (15) ψ 0 At the equilibrium point, the force f was decreased by the extension f = g (p; ε ) = g (p; (ε + ε)), (16) where g ( ) is force model with tendon-spring assumption, ε is actual contraction of PAM, and ε is additional to the PAM contraction. From Equ. (13), ε = f k. Note that the actual contraction ε is difficult to be measured directory. Thus, we use instead estimated contraction ε est (= ε + ε). If the desired force f is needed at ε, the desired valve pressure p derived from inverse model g 1 ( ) p = g 1 (f ; ε ) = g 1 (f ; ε est ), (17) To install light weight PAM to our exoskeleton, we decomposed commercially available PAM (from FESTO Inc. as Fluidic Muscle,) to get the bladder part and combined it with original edge parts (light-weight mechanical structure made by a engineer plastic such technique used in [18]) achieved almost half weight reduction. Using least square algorithm, we estimated PAM parameters a, b, and c by data took from the public data-sheet virtually, and hand-tuned k by the calibration experiment (measuring equilibrium force by onedof system such as Figure 6 and also described below sections). The major torque for movement assist (low frequency torque) is covered by PAM, and actual torque can be measured from LC (τ P AMs ). The high frequency torque (high frequency torque) is generated by motor. The desired motor torque is τ motor = τ τ P AMs. (18) Figure 7 (a) and Figure 7 (b) show typical PAM operation range and posture stand-up/sit-down posture of XoR. Figure 7 (c) shows force measured force in calibration experiment compared with the torque controller performance between the proposed PAM torque model (with tendon-spring model) and the conventional model. Figure 8 shows the torque controller response which is observed using the onedof system. The desired input is sign wave and the two comparisons (original model and proposed model) as two plots. D. Generating Vertical Force to Assist Stand-up movements In this study, the vertical assistive force for the stand-up movements is generated as: τ = J T F, (19) 24

5 f f motor Fig. 6. PEHA architecture: In the upper figure, two PAMs antagonistically generate large force transmitted to a joint through the wires. Because the wire extend by the PAMs forces, the contraction at the force equilibrium changes. The problem is the difficulty to measure true PAM contraction rate. Only the apparent contraction rate calculated from joint angle is measurable. The small motor additionally generates torque to the joint through mechanical belt. (C) p=0.5 p=0.4 p=0.2 p=0.1 p=0.6 p=0.3 Desired torque as input Tendon-compliance equilibrium model Original PAM model Fig. 8. Improved frequency response of the torque control: The antagonistic muscle pressure was kept by maximum (0.6[MPa]) and we input the sign curve as desired torque to the flexor muscle. Thus, the joint angle is moved. Because the tendon-spring model is the pressure to torque model provided the change of angle is quasi-static, the frequency response has a certain error while the static force model (shown in Figure 7(c)) well fits to the measured force. This frequency response error can be covered by motor torque. generated by upper body movements of a user. IV. RESULTS Here we show the results of the developed EEG-robot systems. p=0 A. EEG-oneDOF system Air muscle F (a) (a) Z (top) Wire X (front) HFE h KFE h (b) (b) HFE h KFE Figure 9 shows the experiment setup. The subject watched the display during all the experiment to recognize classifier output (EEG decoder output) in realtime. Figures 10 and 11 show the control performance of the EEG-oneDOF system. The subject dose not wear the onedof system in this example. The subject tries to control onedof system to follow the up/down direction indicated on the display by using motor imagery. The white/gray region in Figure 11 shows the up/down target direction respectively. Fixed edge AFE AFE Realtime EEG Decoded feedback (EEG Decoder output) Fig. 7. PAMs implementation in our exoskeleton: (a) At the lower squat posture, the PAM generate tension large. (b) At the upper standing posture, necessary torque for balance and sustain is small. (c) The plot shows model comparison (proposed tendon-spring model and original PAM model) and measured the force generation at each contraction. where J is the COM Jacobian matrix, F is the desired virtual forces to assist stand-up movements [10], and τ is the desired torque at each joint of the exoskeleton robot. Here we only consider the vertical force and assume that horizontal force applied to COM of user-robot system, and vertical force is Subject Display Fig. 9. EEG realtime visual feedback: The subject watches the display during experiment to get visual EEG decoded feedback. The visual feedback is displayed as the red bar (upper/lower) in the picture as probability of EEG decoder output (P (q t = +1 C t ) and P (q t = 1 C t ) respectively). The direction queue is displayed at the middle of the bars. The subject task is motor imaginary of left arm movement and of right arm movement, corresponding to up/down direction. Figures 12 and 13 show, again, control performance of the EEG-oneDOF system. However, in this case, the subject wears the onedof system. The subject tries to control 25

6 Fig. 10. EEG-oneDOF experiment. The subject dose not wear the onedof system in this example. The subject tries to control onedof system to follow the up/down direction indicated on the display the subject by using motor imagery. (Left) Up state. (Right) Down state. time [s] Fig. 11. Control performance of the EEG-oneDOF system. The subject dose not wear the onedof system in this example. The white/gray region shows the up/down target direction respectively. The performance (0.5 thresholded correct rate) was in the last half of the session from 150 [sec] to 300 [sec]. (Top) Decoded brain activities. (Middle) Control command generated from the decoded brain activities. (Bottom) Joint angle trajectory of the onedof robot. onedof system to follow the up/down direction indicated on the display by using motor imagery. The white/gray region in Fig. 13 shows the up/down target direction respectively. In both cases, i.e., 1) the case in which the subject dose not wear the robot and 2) the case in which the subject wears the robot, the indicated robot movements are correctly generated in most cases. time [s] Fig. 13. Control performance of the EEG-oneDOF system. The subject wears the onedof system in this example. The white/gray region shows that the target direction is up/down respectively. (Top) Decoded brain activities. (Bottom) Control command generated from the decoded brain activities. tion indicated on the display by using motor imagery. The white/gray region in Fig. 15 shows the stand-up/sit-down target direction respectively. The user was successfully controlled the exoskeleton robot by using his brain activities. Fig. 14. EEG-Exoskeleton experiment. The subject tries to control EEG- Exoskeleton system to follow the stand-up/sit-down direction indicated on the display by using motor imagery. The gravity compensation of lower limbs model (total weight of XoR and human lower limbs) was activated when the EEG decoding is up. The PAM torque controller was implemented with tendon-spring equilibrium model. When this system used in rehabilitation, self-balancing torque controller[10] and another safety systems can be also installed. The performance (0.5 thresholded correct rate) was in the last half of the session from 60 [sec] to 120 [sec]. (Left) Sit-down state. (Right) Stand-up state. time [s] Fig. 12. EEG-oneDOF experiment. The subject wears the onedof system in this example. The subject tries to control onedof system to follow the up/down direction indicated on the display by using motor imagery. The performance (0.5 thresholded correct rate) was in the last half of the session from 150 [sec] to 300 [sec]. (Left) Up state. (Right) Down state. B. EEG-exoskeleton system Figures 14 and 15 show control performance of the EEG- Exoskeleton System. The subject tries to control EEG- Exoskeleton system to follow the stan-up/sit-down direc- Fig. 15. Control performance of the EEG-Exoskeleton System. The white/gray region shows that the target direction is stand-up/sit-down respectively. (Top) Decoded brain activities. (Bottom) Control command generated from the decoded brain activities. V. CONCLUSION In this study, we developed the EEG-Exoskeleton system in which the exoskeleton robot were controlled by using decoded EEG signals. Brain activities of the users are decoded by using a classification method which can consider 26

7 covariance matrix of the observed EEG signals as input variables. Then, we connected the EEG system to our one DoF test system. The subject tried to control the robot with two different situations: 1) the subject is not wearing the robot and 2) the subject is wearing the robot. We showed that EEG-oneDoF system was successfully controlled by using the decoded brain activities in the two different situations. Finally, we connected the EEG system to the exoskeleton robot. We showed that even when the EEG system was used with the exoskeleton robot, the user was able to decode the brain activities and control the exoskeleton robot. In these application, the proposed tendon-spring equilibrium model improved the torque controller performance of PEHA approach in large force operation of PAM. We are planning to increase the number of subjects to verify generalization performance of the developed EEG- Exoskeleton system. As a future study, we will investigate how the assist torque generated by the robot influence the brain activities. ACKNOWLEDGMENTS We thank Tatsuya Teramae, Nao Nakano and Akihide Inano for hardware maintainance and helping our experiments. We also thank Yusuke Takeda and Okito Yamashita for helpful comments on EEG measurement. This research is the result of Brain Machine Interface Development carried out under the Strategic Research Program for Brain Sciences by the Ministry of Education, Culture, Sports, Science and Technology of Japan. T.N. was partially supported by Grantin-Aid for Scientific Research J.M. was partially supported by Strategic International Cooperative Program, Japan Science and Technology Agency (JST) and by Grantin-Aid for Scientific Research on Innovative Areas: Prediction and Decision Making [9] D. Kamatani, T. Fujiwara, J. Ushiba, K. Shindo, A. Kimura, and L. Meigen, Study for evaluation method of effect of bmi rehabilitaiton by using near infrared spectoscopy, in Neuro 2010, 2011, pp. P2 f16. [10] S. Hyon, J. Morimoto, T. Matsubara, T. Noda, and M. Kawato, XoR: Hybrid Drive Exoskeleton Robot That Can Balance, in IEEE/RSJ International Conference on Intelligent Robots and Systems, [11] R. Tomioka and K. Aihara, Classifying matrices with a spectral regularization, in Proceedings of the 24th international conference on Machine learning (ICML 07). ACM Press, 2007, pp [12] R. Tomioka and K. R. Müller, A regularized discriminative framework for EEG analysis with application to brain-computer interface, NeuroImage, vol. 49, pp , [Online]. Available: [13] C. M. Bishop, Pattern Recognition and Machine Learning. NY, USA: Springer, [14] I. Sardellitti, J. Park, D. Shin, and O. Khatib, Air muscle controller design in the distributed macro-mini (DM 2 ) actuation approach, Intelligent Robots and Systems, IROS IEEE/RSJ International Conference on, pp , [15] K. Inoue, Rubbertuators and applications for robots, in Proceedings of the 4th international symposium on Robotics Research. MIT Press, 1988, pp [16] D. Caldwell, A. Razak, and M. Goodwin, Braided pneumatic muscle actuators, in Proceedings of the IFAC Conference on Intelligent Autonomous Vehicles, 1993, pp [17] C. Hartmann, J. Boedecker, O. Obst, S. Ikemoto, and M. Asada, Realtime inverse dynamics learning for musculoskeletal robots based on echo state gaussian process regression, in Accepted at RSS 2012, [18] K. Kawashima et al., Development of robot using pneumatic artificial rubber muscles to operate construction machinery, Journal of Robotics and Mechatronics, vol. 16, no. 1, REFERENCES [1] S. Jacobsen, On the Development of XOS, a Powerful Exoskeletal Robot, in 2007 IEEE/RSJ International Conference on Intelligent Robots and Systems, Plenary Talk, [2] H. Kazerooni and A. C. nad R. Steger, That Which Does Not Stabilize, Will Only Make Us Stronger, International Journal of Robotics Research, vol. 26, no. 1, pp , [3] K. Suzuki, M. G. Kawamoto, H. Hasegarwa, and Y. Sankai, Intensionbased walking support for paraplegia patients with Robot Suit HAL, Advanced Robotics, vol. 21, no. 12, pp , [4] S. K. Au, P. Dilworth, and H. Herr, An ankle-foot emulation system for the study of human walking biomechanics, in IEEE International Conference on Robotics and Automation, 2006, pp [5] H. Kobayashi, A. Takamitsu, and T. Hashimoto, Muscle Suit Development and Factory Application, International Journal of Automation Technology, vol. 3, no. 6, pp , [6] G. Yamamoto and S. Toyama, Development of Wearable-Agri-Robot- Mechanism for Agricultural Work, in IEEE/RSJ International Conference on Intelligent Robots and System, 2009, pp [7] T. Kagawa and Y. Uno, Gait pattern generation for a power-assist device of paraplegic gait, in The 18th IEEE International Symposium on Robot and Human Interactive Communication, 2009, pp [8] K. Shindo, K. Kawashima, J. Ushiba, N. Ota, M. Ito, T. Ota, A. Kimura, and M. Liu, Effects of neurofeedback training with an electroencephalogram-based brain-computer interface for hand paralysis in patients with chronic stroke: a preliminary case series study, J. Rehabil. Med., vol. 43, pp ,