In this work, we established multi-dimensional wrist musculoskeletal model to be used for suppressing the wrist joint's tremor by functional electrical stimulation. Often the wrist model for FES control is based on 1DOF biomechanical model for the simplicity and convenience to develop the controller. However, wrist motion is generated by complex interactions of multiple muscles spanning the wrist joint. Here, we have tried to have 3DOF wrist model considering main musles involved in flex-extension, radial-ulnar deviation, and pron-supination as in the left of Fig. (opensim model). Inertia of 3DOF joint model was obtained based on the work of de Leva which modified Zatsiorsky parameters. Joint dynamics was formed by inertia, gravitational torque, and passive visco-elastisity. As for the mapping between torque and muscles, the moment arm matrix $(3\times 4)$was obtained from opensim software, thus anatomical information could be considered along with muscle parameters such as muscle-length, isometric maximal forces. For this first trial, 4 muscles were considered based on Hill-type model. In order to confirm the generated motion by muscle activation, we activated the ECRB (extensor carpi radialis brevis) and FCU (flexor carpi ulnaris) in antiphase, the corresponding 3DOF wrist angles were obtained as in Fig. . Qualitatively, we know that ECRB can generate negative flexion angle, negative deviation angle and positive pronation when it is activated; while FCU can produce positive flexion angle, positive deviation angle and positive pronation angle. When both ECRB and FCU are activated in order, the reasonable wrist angles could be generated. We will work on the development of tremor controller in the future work.

EMG signal is widely used to control a limb protheses such as exoskeleton. It would be also useful to control prosthetic hand for upper-extremity amputees. Especially for upper limb control, hybrid control both for position and torque is required for dynamic motions. In this work, we aimed at establishing the EMG-joint angle model for dynamic motions, to effectively decode EMG signals to reproduce the corresponding motion in different velocity. In the experiment, the subject performed the wrist flextion-extension with different speeds along with EMG measurements on flexor and extensor muscles. ARX model, of which the parameters are adaptively identified by extended kalman filter, was applied to represent 1 DoF (degree of freedom) EMG-joint angle model. The result shows that the EMG-angle model could produce good angle tracking as shown in Fig. . We have observed that three principal parameters a1, a2, a3 for autoregressive term are almost constant at the same angular speed and different angular speeds correspond to different model parameters. With this regular pattern, SR (Switch Rregime) model would be possible to be used to switch the model to reproduce the motion in different conditions. In the further work, we work on creating a SR EMG-angle model using learning and classification technique to dynamicaly decode the EMG signals into the corresponding motions only based on EMG without angle information.

In this work, we explore skeletal muscle volumetric deformation. The current available simulation for musculoskeletal model is basically using wire-type muscle model which considers only principal longitudinal path of the muscle-tendon units. If we aim at the simulation of the interaction between muscles and objects like orthesis, exoskeleton keeping classical biomechanical property, wire-type modeling is not sufficient. In addition, muscle modeling in volumetric way gives another advantage to reflect microscopic muscle fiber direction and function. We have tried to implemente real-time volumetric skeletal muscle deformation as in Fig. using the INRIA SOFA environment. The idea is making more realistic musculoskeletal simulation from the current approximation of muscle model as wire element to physically and functionally detailed simulation as volumetric element.

In current biomechanics approach, the assumptions are commonly used in body-segment parameters and muscle strength parameters due to the difficulty in accessing those subject-specific values. Especially in the rehabilitation and sports science where each subject can easily have quite different anthropometry and muscle condition due to disease, age or training history, it would be important to identify those parameters to take benefits correctly from the recent advances in computational musculoskeletal modeling. In this paper, Mass Distribution Identification to improve the joint torque estimation and Muscle Strength Identification to improve the muscle force estimation were performed combined with previously proposed methods in muscle tension optimization. This first result highlights that the reliable muscle force estimation could be extracted after these identifications. Fig. shows the estimated muscle forces of Rectus Femoris, Vastus Lateralis and Vastus Medialis with different speeds (first two series are normal speed, second two are slow and last two are fast). The corresponding visualizations of estimated muscle tensions at the indicated time instant are depicted in the bottom. The proposed framework toward subject-specific musculoskeletal modeling would contribute to a patient-oriented computational rehabilitation .

In this work, we explore the combined use of inertial sensors and the Kinect for applications on rehabilitation robotics and assistive devices. In view of the deficiencies of each individual system, a new method based on Kalman Filtering was developed in order to perform online calibration of sensor errors automatically whenever measurements from Kinect are available. The method was evaluated on experi- ments involving healthy subjects performing multiple DOF tasks. Accelerometers and gyrometers are used to estimate joint angle, while the Kinect is used for initializing the inertial system and for enabling 3D visualization of the performed task as in Fig. .

Fibre type and diameter selective stimulation may allow to restore various motor and sensory functions of human body that have been lost due to disease or injury. For example in people unable to voluntary empty the bladder, selective stimulation of small fibres within the ventral branch of the sacral nerve roots (S2-S4/5) would induce detrusor contraction and those cause bladder empting closest to normal physiology. Currently, it is not possible to perform it in such a way, because stimulation of sacral nerve roots activates also bigger fibres innervating the urethral sphincter, which closest the outlet of the bladder.

Already many stimulation techniques have been proposed for fibre type and diameter selective stimulation. They were verified performing computer simulations and in some cases also by in vivo experiments on mammalian models. However, results of computer simulations still need to be confirmed by in vivo experiments, whereas experiments on mammalian models, due to high number of fibres within stimulated nerve, can be very complex to perform and obtained results difficult to interpret. As a result, it is still unclear which stimulation parameters may allow for selective stimulation of only particular group of fibers.

Therefore we propose the earthworm (Lumbricus terrestris) as a model for selective stimulation. The earthworm has three giant nerve fibres, with two distinctly different conduction velocities and diameters. Therefore it is very easy to distinguish between fibres that are firing at the moment. As a consequence the selectivity of stimulation may be immediately verified without application of sophisticated signal processing and averaging techniques.

We have investigated influence of various pulse amplitudes and durations on the selectivity of stimulation. Using a simple experimental set-up  shown in the fig. A, we were able to achieve selective activation of small (fig. C) and big (fig. D) fibers, as well as concurrent activation of both fibers (fig. B)  . For that purpose we have used so called "anodal block" technique.

Based on the results of the above experiments, the recommendations of the optimal parameters for selective stimulation of nerve fibres will be prepared. Afterwards we are going to verified these recommendations in mammalian models (rats).

Due to high nonlinearity and time variance in muscle dynamics under FES, the identification of muscle model is complex task. The time-variance of muscle response may come from muscle fatigue, but also the electrode attachment condition. Along with such long-term time-variance, short term time-variance may be created by reflex effect. In addition, the characteristics of such variance even may change in time in an unpredicted way. Reinforcement learning framework may be applied to bring the robustness in adaptive identification. Current work is focused on the usage of discrete-time recurrent neural network for model identification.

Hemiplegia is a condition where one side of the body is paretic or paralyzed; it is usually the consequence of a cerebro-vascular accident. One of the main consequences of hemiplegia is the drop-foot syndrome. Due to lack of controllability of muscles involved in flexing the ankle and toes, the foot drops downward and impede the normal walking motion. Today, there are commercially available assistive systems that use surface electrodes to stimulate Tibialis Anterior (TA) muscle and prevent drop-foot. The efficiency of drop-foot stimulators depends on the timing of stimulation and functionality of dorsiflexion motion. Classically, available stimulators use footswitches to detect foot on/off events. These discrete events allow only for triggering the stimulation and/or playing with the duration of the stimulation pattern, but does not allow for precise online modification of the pattern itself. We have developed algorithms to monitor the ongoing walking cycle by observing the valid limb movements. In order to ensure legs coordination during walking, we propose a robust phase estimation method based on the observer of a nonlinear oscillator. We have modified a commercial stimulator, ODSTOCK, in order to be able to trigger it using our own wireless sensors and algorithms. Agreement from Nîmes CPP (ethical committee) was obtained in June 2010 to run tests on patients. The protocol comprises 1) the control algorithm triggering the stimulator based on signals issued from one wireless inertial sensor placed on healthy shank, 2) a sensor setup including inertial sensors placed on deficient shank and foot, one goniometer measuring deficient ankle angle, one EMG sensor placed on stimulated TA and one instrumented carpet (GAITRITE) (fig. ). Several patients have been included in the study and data is being processed .

Electrical stimulation (ES) is one of the solutions for drop foot correction. Conventional ES systems deliver predefined stimulation pattern to the affected muscles. However, time-variant muscle response may influence the gait performance as they are difficult to be taken into account in advance. Therefore, closed-loop ES control is important to obtain desired gait in presence of muscle response variation. In this work, a dual predictive control, which consists of two nonlinear generalized predictive controllers, is proposed to track desired torque. The stimulated muscle dynamics are modeled by Hammerstein cascades, with one representing stimulation to activation, the other representing activation to torque. Ankle dorsiflexion torque and ES-evoked EMG of tibialis anterior were recorded experimentally for model identification. The control scheme is validated by following desired torque trajectories with the identified model. The results show that the stimulation pattern obtained from the dual predictive control can produce good torque tracking according to the current muscle condition as shown in Fig. . The updates of model parameters were switched off after certain instant for both the excitation and contraction model. Consequently, the model prediction in the control was only driven by the model input and the last model parameter estimates. The dual predictive controller can still generate suitable control signal to obtain desired torque trajectory .

Standing up is a common daily activity and a prerequisite to standing or walking. This frequently executed task is one of the most biomechanically demanding activities. The ability to rise from a sitting to a standing position is very important for individuals with paraplegia in order to achieve minimal mobility and has functional and therapeutic benefits related to bone loading, joint extension, cardio-circulatory stimulation, and pressure sore prevention.One method which has been widely investigated is functional electrical stimulation (FES) of the lower extremities. The sit-to-stand method, which is widely used in clinical practice, involves open-loop stimulation of knee extensors activated by hand switches. This technique works adequately in many cases, however, in applying this strategy, stimulation starts without referencespect to the upper body movement. Hence, the whole- body motion is not optimal and requires a high velocity of the joints and large upper limb forces during the rising motion, which may cause both damage of joint tissues and shoulder complications.

We propose a "patient-driven" FES method that would coordinate motion of the trunk, which is under voluntary control of the patient, and motion of the lower limbs, which are under FES control. The proposed approach is based on the observation of trunk movement during rising motion and a detection algorithm, which triggers a pre-programmed stimulation pattern.Trunk acceleration was acquired by a singleone-axis wireless accelerometer positioned on the subject's back. The detection algorithm consists of an online comparison of the movement acceleration of the ongoing motion with the reference pattern (a typical pattern characterizing the sit-to-stand transfer for each subject) using Pearson's correlation coefficients . Experiments on paraplegic subjects are ongoing in refabilitation center PROPARA. The experimental setup and a paraplegic subject of the experiment are presented in Fig. .

We have shown that in the cases where the acceleration and reference signal are similar, our algorithm is able to recognize sit-to-stand motion and to properly trigger leg stimulation at the desired instant. Also, we have showned that there is an influence of stimulation timing on applied hand forces during the motion. The best results were achieved for trials in which motion was similar to the one of the able bodied subjects in terms of trunk motion and the beginning of the leg motion with respect to the trunk acceleration signal.

We also investigated dynamic optimization as a tool to improve FES assisted sit tostand transfers of paraplegic subjects. The objective would be to find optimal strategy for voluntary trunk movement, which would minimize hip, knee and ankle torques and demand minimal upper limb participation during the motion. Our results suggest paraplegic patients should bend their body forward in order to use linear momentum of the trunk in sit off phase. Figure shows optimal coordination of ankle, knee, and hip angles during sit-to-stand motion , .

Last year, we introduced a simple approach to segment in homogeneous phases a long-duration record of locomotion data consisting of body segment acceleration and foot pressure information. We used a system based on a network of wireless nodes embedding various types of sensors . Two cases were considered: walk and run around an indoor running track and outdoor marathon .

We now use this system as part of a study of mobility impairment caused by Parkinson's disease (fig. ). Freezing of gait (FOG) has been identified as one of the main contributors to gait disturbances in this disease. We introduce an ambulatory gait analysis method using body attached gyroscopes and accelerometers to detect the freezing of gait. One hand, we aim at proposing a FOG detection algorithm more robust because the existing algorithms were not able to detect the FOG without tremor. On the other hand, we would like to anticipate the freezing before ir is installed in order to reduce the risk of falling.

Before and during a FOG, a patient tends to walk slowly with short strides and fast rhythm. The detection of an increase of frequency is not enough, because there exists similar variations during the initiation of gait or a voluntary acceleration. The association of an increase of the gait rhythm together with a decrease of stride length allows us to detect a FOG. The computation of correlation coefficient in a moving window allows us to estimate the rhythm of strides. We are also working on different methods to estimate the stride length using one or two IMU (3-axis gyroscope and 3-axis accelerometer).

A time-frequency representation permits to show an increase of fundamental frequency and a duplication before a FOG, Fig. . The variations of the fundamental frequency are already detected with the correlation. In the future works, we aim at characterizing these duplications and to propose an algorithm of automatic detection.

It is now possible to perform resections of slowgrowing tumors in awake patients. Using direct electrical stimulation (DES), real-time functional mapping of the brain can be used to prevent the resection of essential areas near the tumor. For now, simple clinical tests are performed on conscious patients and combined with DES in order to discriminate functional and non-functional areas invaded by the tumors. In this work we try to develop a simple device based on a simple technology to better quantify the performances of the patients during the surgery itself and give a real-time feedback to the neurosurgeon that will help to further guide the surgery by improving the sensibility of the functional mapping. This procedure should also allow building a strong database that should serve retrospectively to improve the surgical procedure and reinforce the neurosurgeons' experience as well as to monitor the patients' performances all along their life.

Center of Mass control has been used in humanoid robotics to create stable standing postures and movements. By controlling the CoM position and acceleration, joint trajectories which respect to the ZMP stability criterion can be generated. FES may be used to drive joint torques in order to maintain a standing posture within a closed loop controller. Current work is focused on locating a human's CoM by creating a statically equivalent serial chain (SESC) model using widely accessible equipment, such as the Kinect camera and the Wii balance board.

In the context of the timeproject, one cafe12-based stimulator will be used for chronic experiment in human. During the validation of the stimulator, it appeared that we needed to improve both the linearity and the current matching of the 12 channels. We thus define a calibration process consisting in:

The measurement setup makes use of a characteristic analyzer (HP4156A) to maintain the voltage load to $7.5V$for all the daccurrent values.

The improvement in the matching of the 12 channels can be seen in figure  . The results in terms of linearity are an integral non-linearity of $\pm 2.5\mathrm{LSB}$and a differential non-linearity of $\pm 0.3\mathrm{LSB}$.

In the context of fes, neural recording is one of the main issues, as the control requires information carried on afferent peripheral nerves. Because specific information are carried in different fascicles, we propose to realize a non-invasive and spatial-selective electrode. Last year, based on investigation on the topic of extracellular Action Potentials (ap), we proposed a new tripole design, where the tripolar output signal is the image of the activity in the close vicinity of this tripole, providing high spatial selectivity.

We showed however, that this high spatial selectivity is achieved at the expense of signal amplitude. This first result jeopardizes the feasibility of this kind of electrode since the signal amplitude appears to be on the same range of the expected noise. First, we propose to estimate the performance of the proposed electrode with a quantitative study of the electrode selectivity. Then, to conclude on the feasibility of this electrode, the snrhas to be determined. So with a more accurate model, we studied the sensitivity of the proposed tripole, allowing to determine precisely the amplitude level of the expected signal. Thus, the snrcan be estimated knowing the expected noise.

In short, the work of this year aims at characterizing the performances and evaluating the feasibility this new multi-contact cuff electrode.

This year we proposed a method for enhancing the noise-power tradeoff of front-end amplifiers in parallel recording applications of analog signals with respect to a common reference. One example of application is shown in the Fig.  for spatial-selectivity engrecordings.

The circuit architecture is based on a Shared-Input Amplifier ( sia), composed by shared-input transconductance amplifiers and a differential stage. Averaging null signals and subtracting the result from every signal reduce the noise because the correlated noise between parallel outputs is attenuated. It results significant supply current savings without noise penalty. One example is shown in the Fig.  for the specific case of two average of two null signals.

Also, a method for reducing the remaining noise with little power penalty is possible. And it is possible to combine both methods, either noise level, total supply current, or both can be significantly reduced. The benefits of combining both methods and the related trade-offs was validated by simulations using models of a $0.35\mu \mathrm{m}$BiCMOS process. So we have shown that the total supply current can be reduced by more than $50\%$that of a comparable system using conventional differential amplifiers with equivalent output noise. Alternatively, the noise can be reduced by approximately $35\%$with comparable power consumption.

This should enable low-noise recording of signals with significantly better efficiency than even the theoretical limit of any conventional differential amplifier. Future work will include circuit optimization and investigation of the impact of the architecture over other performances of the system, such as crosstalk, linearity, distortion, and channel mismatch. Implementation of the circuit for full characterization is expected to be completed in the near future.