康復(fù)機(jī)器人的系統(tǒng)設(shè)計(jì)【基于步態(tài)控制的下肢康復(fù)機(jī)器人】
康復(fù)機(jī)器人的系統(tǒng)設(shè)計(jì)【基于步態(tài)控制的下肢康復(fù)機(jī)器人】,基于步態(tài)控制的下肢康復(fù)機(jī)器人,康復(fù)機(jī)器人的系統(tǒng)設(shè)計(jì)【基于步態(tài)控制的下肢康復(fù)機(jī)器人】,康復(fù),痊愈,機(jī)器人,系統(tǒng),設(shè)計(jì),基于,步態(tài),控制,節(jié)制,下肢
原文：
RUPERT: An Exoskeleton Robot for Assisting Rehabilitation of Arm Functions
Sivakumar Balasubramanian, Student Member, IEEE, Ruihua Wei, Mike Perez, Ben Shepard, Edward Koeneman, James Koeneman, and Jiping He, Senior Member, IEEE
Abstract: The design of a wearable upper extremity therapy robot RUPERT IV? (Robotic Upper Extremity Repetitive Trainer) device is presented. It is designed to assist in repetitive therapy tasks related to activities of daily living which has been advocated for being more effective for functional recovery. RUPERT? has five actuated degrees of freedom driven by compliant and safe pneumatic muscle actuators (PMA) assisting shoulder elevation, humeral external rotation, elbow extension, forearm supination and wrist/hand extension. The device is designed to extend the arm and move in a 3D space with no gravity compensation, which is a natural setting for practicing day-to-day activities. Because the device is wearable and lightweight, the device is very portable; it can be worn standing or sitting for performing therapy tasks that better mimic activities of daily living. A closed-loop controller combining a PID-based feedback controller and a Iterative learning controller (ILC)-based feedforward controller is proposed for RUPERT for passive repetitive task training. This type of control aids in overcoming the highly nonlinear nature of the plant under control, and also helps in adapting easily to different subjects for performing different tasks. The system was tested on two able-bodied subjects to evaluate its performance.
I. INTRODUCTION
RECENT advances in neuroscience indicate that the central nervous system can reorganize after injury [1]. This increased awareness of neuroplasticity coupled with emerging interests in controlling and measuring the results of therapy has had a profound influence on the rehabilitation engineering community. This influence has generated a lot of interest in upper and lower extremity robotic therapy research. To some extent, the emerging interest in therapy robots was also facilitated by the fact that many of the traditional therapeutic approaches were time consuming. The increased therapist required time for stroke therapy coupled with the projected increase in the number of stroke survivors because of the aging “Baby Boomer” population leads to projections for increased demand for physical therapy time. Robots can reduce the physical effort of therapists and provide them with the ability to concentrate more on therapy performance.
Most stroke survivors suffer various degrees of loss in both cognitive and motor functions. Motor control research has shown that sensory motor integration is a key process in motor learning and the subsequent improvement in function. Integrating multi-modal biofeedback with therapy will motivate the stroke survivors to actively and volitionally process sensory information during repetitive therapy for motor function. When the therapy is based on a meaningful purposeful task, cognition is challenged, further enhancing learning. The challenge for therapy robots is to achieve the desired effects of motivation, cognitive challenge, easily customizable tasks specific to therapeutic outcomes, and real-time performance assessment [2-4].
A prerequisite for the elaborate movements of the upper extremity, and dexterous abilities of humans in general is the ability to coordinate multiple joints and regulate forces produced by limb segments. While the disturbance of voluntary upper extremity movement in subjects with stroke is typically apparent upon visual examination, little is known about the mechanisms responsible for these disturbances [5]. This is due in part to the dearth of quantitative studies of multi-joint movements in such subjects [3]. During a simple target-directed pointing task, subjects with stroke could reach into all parts of the workspace with their affected limb, [6] suggesting that movement planning was intact for these subjects. However, when inter-joint coordination was assessed by expressing elbow angle as a function of shoulder angle, subjects with stroke exhibited an irregular and variable relationship. This disruption in inter-joint coordination resulted in movement paths that were more segmented and variable. Roby-Brami and colleagues showed that prehension (reaching and grasping) movements of subjects with stroke were characterized by a spatiotemporal dyscoordination between the arm and trunk [7]. As a result subjects’ developed a new pattern of coordination represented by more trunk recruitment during prehensile actions. More recently, a kinematic analysis of reaching movements of subjects with stroke indicated that subjects with stroke, unlike healthy controls, recruited the trunk to assist in transporting the hand to the object [8]. Thus, these subjects were recruiting a new degree of freedom (e.g. the trunk) to perform this task. Kinematic studies have also shown that subjects with stroke have a more variable reaching path, orientation of the hand relative to the object, final hand position on the object, and a disruption in inter-joint coordination. These data suggest that subjects with stroke have difficulty with motor execution. Therefore rehabilitation of the affected upper extremity should be oriented toward restoring the normal sensorimotor relationships between the joints [6].
II. METHODS
A. Design Basis
Motor control research has shown that both subcortical and cortical networks participate in the control and modulation of movement. The principles of neuroplasticity propose that these networks can be “rewired” through repetitive training. It is our hypothesis that in the process of training neurologically injured subjects, therapy tasks should be based on tasks related to activities of daily living in a natural environment. This will encourage the retraining of the subtle coordination in cortical and subcortical networks to perform real-life activities in a natural environment, rather than training in a gravity compensated or any other unnatural environments. Therefore, our device and therapy protocol consists of a wearable device so that skills needed for numerous activities of daily living can be trained in a natural gravity environment.
Traditional robots are usually driven by electric motors attached to gear boxes which can be very stiff and can supply very large torques which could result in injury to a stroke survivor with spasticity. Electric motor actuation in wearable robotic applications presents a mismatch in the compliance of the actuator and the limb being assisted. A compliant interface for human interaction requires a system that can easily control resisting forces. The traditional approach to designing a compliant interface for human interaction is to actively control the compliance of motors using appropriate control algorithms such as those developed by Salisbury [9] or Paul [10]. The main disadvantage in this method is that in such systems, the mechanical stiffness is typically very large and it is necessary to rely on high-performance actuators (expensive motors) and high bandwidth control systems to provide compliance using robots that are particularly designed for position control tasks. Such schemes have inherent limitations during interaction with stiff environments (i.e., such as repeated impacts with a walking surface). However, a better approach is to build into the system some mechanical compliance and then use active control to vary this compliance. The main advantage of this approach is that there is always some compliance in the system regardless of the stiffness of the environment. As a result, the requirements on the actuator and the control bandwidth are more modest. Therefore, the pneumatic McKibben artificial muscle was chosen as the actuator for our therapy robot.
The RUPERT therapy protocols that are being developed to provide the ability to assist with therapies in the same manner that therapists approach stroke patient treatment, i.e., RUPERT should provide proximal stabilization with the capability of providing therapy on isolated joints. After working on the component parts (individual joint motion), therapy progresses to combined motions. Combined motion tasks are developed so the temporal-spatial difficulty of the task can be increased as the motor ability of the subject improves.
The tasks are tied to the perception of functional importance to the patient. Eventually RUPERT will have a catalog of tasks from which the therapist/patient can select the therapy tasks of interest. Three tasks are initially being tested: reaching, washing the contralateral arm and drinking. The difficulty of these tasks depends on the target location and desired timing.
B. Structural Design
If restricting continuous tone exists in stroke survivors, it usually occurs in the “anti-gravity” muscles, i.e., the flexors of the fingers, hand, wrist, and elbow, the pronators of the forearm, the extensors of the shoulder, and internal rotator muscles of the humerus. The design of an earlier version (RUPERT III) has been described [11]. The current design (RUPERT IV) has added shoulder external rotation to expand the space available for performing assisted tasks. Assistance and measurement of the following joint motions is provided in RUPERT IV; hand/wrist extension, forearm supination, elbow extension, humeral external rotation and shoulder elevation. To accommodate a large range of subject sizes, adjustments in the structure are provided (three degrees of adjustment at the shoulder, and humeral length, forearm length, hand length adjustment). To keep the weight of the structure low, many of the structural components are made of graphite composite materials. A picture of RUPERT IV is shown in Figure 1.
Fig.?1.?Picture?of?RUPERT?IV
C. Control Design
The immediate design goal for the RUPERT aims to help the stroke survivors to practice basic arm movement relevant to activities of daily living repeatedly to achieve therapeutic benefit for improved functionality and independence. The challenge for the control system is to provide a consistent performance for stroke survivors who all have different functional impairments and neurological conditions. Safety and comfort impose significant constraint on the design of control algorithm.
The hardware of control system for the robot consists two major components, (i) a control box housing a PC104 single board computer, and five pressure regulators that actuate each of the DOF in RUPERT IV, (ii) a main PC (mPC) which acts the terminal for the physician/therapist/operator to interact with the robot.
The overall structure of the RUPERT control system has two layers, namely, the Inner Loop (IL) and the Outer Loop (OL) as shown in Figure 2. The IL controller works at individual joint level, while the OL controller works at the level of functional tasks.
The current closed-loop controller for RUPERT IV is developed for a passive therapy mode. The controller was designed in a Matlab/Simulink environment on the mPC and loaded into the control box.
Fig.?2.?Overall?Closed?Loop?Control?Structure.
1. Outer Loop Controller: The OL controller is currently a simple (open loop) trajectory planning module that generates a command trajectory in the joint space for the controllers in the IL. It should be noted that the current version of the OL controller does not form a closed-loop as shown in Fig. 2. Given an initial and target position in joint space, the OL generates a smooth trajectory in the joint space using the following equations,
Where, is the command angle, is the initial angle, is the target angle, T is the task duration, and t is time.
The above equation is the minimum jerk equation in the Cartesian space [12], but we use it in the joint space in our reaching tasks just to provide a smooth command signal to the IL controllers.
2. Inner Loop Controller: Each of the five joint controllers in the IL consists of a PID feedback controller. Additionally, three of the joint controllers (Shoulder flexion/extension, Elbow /flexion/extension, and Humoral rotation) have an Iterative Learning Controller (ILC) in parallel to the PID feedback controller (Fig. 3). Because of the highly nonlinear nature of the plant (robot + Stroke subject’s arm) being controlled, the use of a traditional linear PID controller can produce highly varying responses for different command signals and for different subjects. Apart from this, the response becomes slightly oscillatory (or feels jerky) for slower movements; thus, even for a smooth command signal the actual movement appears as if it is composed of many sub-movements. This jerky movement is prominent especially in the joints requiring large PMA (Internal tube diameter >in1) to provide a strong force to assist movement. A large PMA has a slower and nonlinear dynamic response, which might be a major reason for the jerky movement.
One way to compensate for the slow nature of the actuator is to use a feedforward controller. Since RUPERT will be used for performing different types of tasks at different speeds, it is unlikely that a single feedforward controller will perform equally well for all of these situations. Designing different feedforward controllers for different tasks on different subjects will be an impossible task and make the IL controller too complicated.
Fig. 3. Individual Joint Controller with PID and ILC Controller (Shoulder Flexion/Extension, Elbow Flexion/Extension and Humoral Rotation have PID+ILC controller; Elbow Supination/Pronation and Wrist Flexion/Extension have only PID controllers)..
The repetitive nature of the therapy and training provides a good opportunity for machine learning techniques to generate a logical control approach: the Iterative Learning Control (ILC). The basic idea is that the controller learns from errors measured from the previous trials and updates the control command to optimize the performance. ILC is a type of learned open loop control strategy that is used in applications with repetitive tasks. ILC improves the performance of the system by learning from the previous executions [13]. Thus, instead of designing a large rule base with control rules that can provide satisfactory performance for difficult tasks on different subjects, we take advantage of repetitive nature of the therapy. The controller can learn for each individual on the first few trials of each session to update and shape a satisfactory feedforward control command for a given training task.
The PID gains for PID+ILC controllers are set to fairly low values to ensure an overdamped response from the controller when there is zero output from the ILC. The low values for the PID gains were chosen primarily to prevent a overshoot in the closed-loop response. It should be noted that overshoots cannot be actively corrected by RUPERT due to the unidirectional nature of the actuation in each DOF. The ILC helps the IL adapt to different subjects by learning from the previous performance of the controller. The learning procedure for the ILC consisting of two separate steps,
Step1: Learning from error signal from the previous iteration.
Where, is the control signal for the current iteration (j) learned from the previous iteration’s control signal (uj?1), L (q) is the filter applied on the previous iteration’s error signal (ej?1), is the learning rate for the ILC, q is the forward shift operator, and n is the sample number.
The value of determines the extent to which an error signal from the current iteration is learned for the next iteration. The learning rate was made a nonlinear function of three performance metrics, to ensure that we learn from the previous performance only when it is necessary. The performance metrics used were, (i) Absolute Mean Error for the previous iteration, (ii) Standard deviation of the absolute mean error for the previous iteration, and (iii) Correlation coefficient of the last two performance error signals. The nonlinear mapping between the learning rate and the three performance metrics is obtained through a fuzzy rule base. The rules for the fuzzy rule base were selected to ensure that the ILC learned only underlying nonlinearities of the plant and not the unwanted disturbances. There were a total of 13 rules used in the rule base. Some of the decisive rules and the rationale behind these rules are as follows,
1. If Absolute Mean Error is VERY LOW then
=ZERO
Rationale: Do not learn when the performance is good.
2. If Absolute Mean Error is MEDIUM (or HIGH) and Error Correlation is LOW (MEDIUM or HIGH) then
= LOW (MEDIUM or HIGH)
Rationale: Learn to correct only the errors that are consistent. This is to prevent the controller from learning unwanted disturbances in the system.
3. If Absolute Mean Error is LOW and Variance is LOW (MEDIUM or HIGH) and Error Correlation is HIGH then
= LOW (MEDIUM or HIGH)
Rationale: Learn any consistent disturbance that causes an increase in the variance of the error signal.
Step2: The next step in the ILC learning procedure is to smooth. A smooth ILC control signal is very essential for producing a smooth movement. A smooth is achieved by representing as a high order polynomial function of time. The coefficients of this polynomial function are obtained by performing a least square fit between and.
Where, bk ’s are the coefficients of the polynomial, T is the sampling time, NT is the total number of samples from an iteration, and N is the order of the polynomial. The order of the polynomial was chosen as 20 for our learning algorithm.
The final control signal (uj) from the IL is a weighted combination of the PID feedback control signal () and the ILC ()
Where, WFB is the weight of the PID controller and WILC is the weight on the ILC.
III. RESULTS
The designed closed-loop controller was tested on two able-bodied performing reaching tasks. The main objective of the testing procedure was to see if the controller was able to adapt to different subjects for different reaching tasks. During the reaching tasks, the subjects (especially the able-bodied subjects) were instructed to remain passive at the shoulder, and simulate some tone at the remaining degrees-of-freedom. This is necessary to prevent overshoots in the controller response, which cannot be actively corrected by the controller. Figure 4 shows the response of the closed-loop controller for the first four reaching tasks involving the shoulder, elbow flexion/extension and humoral rotation. The command signals for these three degrees-of-freedom and the corresponding response of the closed-loop controller are shown in Figure 4 for four consecutive reaching trials. The increased convergence between the command angle (blue trace) and the actual angle (red trace) can be seen with each trial.
The convergence between the command and actual angle was quantified using the absolute mean error between these two signals. The absolute mean error for the four trials shown in Fig 4 is plotted in Fig 5 as a function of the trial number. This figure clearly demonstrates the improvement in the performance of the PID+ILC controller.
IV. DISCUSSION & CONCLUSIONS
The design of a five degree-of-freedom exoskeleton wearable robot, along with a closed-loop controller has been presented. An adaptive controller combining a PID-based feedback controller and a ILC-based learning controller is proposed for RUPERT. The ability of the PID+ILC controller to adapt to different subject and different tasks was tested on two able-bodied subjects and the performance of the controller indicated that controller was able to adapt to the two different subjects.
Our current research is focused towards improving the functionality of RUPERT. The following are proposed as the future work:
1. Implementation of pro
收藏