One typical application of a robotic exoskeleton is to automate rehabilitation, where the robot is worn by a stroke patient and provides assistance to help perform repetitive motions and regain motor functions. The deployment of exoskeletons can alleviate the shortage of experienced therapists and would also play a vital role in countries with aging populations. However, the intelligence level of existing exoskeletons is relatively low, wherein a robot cannot adapt to either the online change of a subject's motion (e.g., the gait pattern) or the variation of his/her body parameters (i.e., a new subject who is going to wear the robot), potentially resulting in conflict between human and robot and possibly even leading to physical damage. As such, the application of a robotic exoskeleton in clinical studies is limited. This article introduces a new bilateral exoskeletal assistive robot for rehabilitation (BEAR-H) in which the main novelty is the integration of multiple intelligent features, such as gait recognition and synchronization, cloud-computing diagnosis, and individualized gait generation. Such an integration helps the robot to better understand the patient's condition and hence provide effective assistance, by the end achieving smart rehabilitation. BEAR-H is a successfully commercialized product, and its performance has been validated in actual clinical studies with 30 patients, producing experimental results from different aspects that are analyzed and presented.
As a typical physical human-robot interaction system, robotic exoskeletons have been developed to provide assistive forces to users through direct and close interactions. For healthy subjects, the role of an exoskeleton is to augment their overall strength, thus reducing their workload. Some of these applications include lowering [1], carrying [2], and lifting [3]. For stroke patients, exoskeletons have been used to assist subjects with performing repetitive motions during rehabilitation.
The demand for robotic rehabilitation can be reflected in many developed countries, where aging populations and stroke patient numbers have been increasing rapidly in recent years. There is also a large population of stroke patients in China (in the tens of millions), and the number is growing at a rate of 10% each year. As a result, there is a lack of approximately 400, 000 therapists in China, which is far more than the total number of the nation’s active therapists [4]. The nature of manual rehabilitation is highly labour intensive as a therapist must hold a patient's limb to assist him/her in carrying out repetitive movements. Moreover, the quality of this rehabilitation is dependent on the therapist’s experience, overall treatment time, and consistency of the procedure. In this regard, deploying rehabilitation robots may help overcome the limitations of manual rehabilitation due to their advantages of high repetition and consistency. This article considers the problem of lower-limb rehabilitation. To automate labor-intensive tasks, a diverse range of robotic exoskeletons have been prototyped, such as lower extremity powered exoskeleton (LOPES) [5], Lokomat [6], WalkTrainer [7], and active leg exoskeleton (ALEX) [8]). The LOPES [5] varies its output impedance to adjust the level of assistance, where the minimal impedance is designed for free walking, and hence achieve task-specific training. In [6], Lokomat works with a bodyweight support module, which is hung on the top of the system and supports the patient when walking on a treadmill. In [7], the WalkTrainer is designed for patients who have already gained a certain level of controlled muscular force, and its main purpose is to enhance the patient's ability of overground walking with leg and pelvic orthoses and closed-loop muscle stimulation. In [8], ALEX applies assistive forces at the foot of the patient with a force-field control scheme such that the undesirable gait motion is rejected while the desirable motion is amplified.
Several exoskeletons have also been commercialized and successfully presented to the market such as Ekso [9], Indego [10], hybrid assistive limb (HAL) [11], and ReWalk [12]. Ekso [9] has four working modes corresponding to the patient's status of sitting, stepping, transfer of weight, and normal walking. The Indego [10] has a slim shape, which allows the patient to wear the robot in a wheelchair, and the robot can assess the patient’s intention by monitoring changes in his/her centre of pressure. In [11], the new version of HAL has a total of six active degrees of freedom of freedom (6 DoF) to actuate the joints of hip, knee, and ankle of both legs, and it refers to the electromyography (EMG) signal to capture the motion intention of patients, thus providing a more convenient control interface. The ReWalk [12] has a maximum walking speed of 2.6 km/hr, and it can detect the events of staircase climbing and falling by using multimodal sensory information.
However, the performance of existing prototypes/products is commonly affected by several open issues. First, the robot joints are commonly driven with rigid actuators, which cannot provide structural safety. Second, the robot joints may not fully cover the entire lower limb in the sense that some of the patient's motions cannot be achieved. Third, many robots are the passive training mode kind, such that the natural functionality after the treatment is not guaranteed. Those issues actually limit the applications of robotic exoskeletons in clinical scenarios.
This article discusses the development of BEAR-H, a new bilateral exoskeletal assistive robot developed by Shenzhen MileBot Robotics Co., Ltd., in China [14]. A comparison among BEAR-H and other robotic exoskeletons for rehabilitation is listed in Table 1. Compared with other products, BEAR-H integrates multiple intelligent features in the training mode: 1) it can automatically recognize the patient's gait event to synchronize the robot's angular trajectory with respect to the patient's actual gait; 2) it can inspect the body parameters of a patient and then customize an individualized gait, corresponding to the gait profile of a healthy subject with similar body parameters; and 3) it carries a cloud-based artificial intelligence diagnosis platform, which allows doctors and family members to monitor the patient's progress when he/ she is under home rehabilitation. The integration of the aforementioned features helps BEAR-H better understand the patient's condition and then generate a more "intelligent" gait for rehabilitation. Additionally, an adaptive impedance control scheme is proposed to drive BEAR-H to track such a gait, which guarantees a safe interaction and also encourages the patient to execute voluntary efforts and hence speed up the recovery. Moreover, compliant actuators are fully deployed in the hip, knee, and ankle joints of BEAR-H to further ensure safety at the structural level. BEAR-H has been successfully implemented in clinical applications, and comparative studies have been carried out for 30 hemiplegic patients. The experimental results from different aspects are presented to validate the effectiveness of BEAR-H.
BEAR-H is a lower-limb exoskeleton robot designed to facilitate smart rehabilitation. Specifically, the robot fully assesses the body parameters, online status, and history data of the patient and then customizes the rehabilitation strategy to better suit his/her condition, thereby improving the rehabilitation effect. The overall system consists of a backpack (housing the control module), segments of waist and both legs, and a touchpad as the interface.
The leg length can be adjusted to suit users with different heights, from 150 cm (i.e., children) to 190 cm (i.e., adults). The dimension of the waist joint can also be adjusted to suit users who weigh up to 85 kg. Each leg has three active and two passive DoF. Hence, the whole robot has six active DoF in total. The active DoF are designed as the flexion/extension around the hip, knee, and ankle joints, while the passive DoF are designed as the inversion/eversion of the ankle joint and the adduction and abduction of the hip joint. Therefore, the whole robot has 10 DoF in total, and such a high-DoF structure supports full rehabilitation for lower limbs. All of the active joints are driven with compliant actuators, guaranteeing safety from the structural aspect. A module in the shape of a shoe is designed at the end of both legs to support the patient's feet. The module is displayed in Figure 1.
The central control system is embedded into the backpack to realize multiple interaction modes and the algorithm of intelligent rehabilitation. The backpack also integrates the modules of emergency stop, input/outputs, and indicators to show the level of assistance and the ground contact status. The battery inside the backpack can support BEAR-H's continuous operation for roughly 8 hr. There are also two handles on the backpack, which enables therapists to help patients put on the robot [see Figure 1(b)].
The touchscreen conveys information during rehabilitation, where therapists specify the training mode and parameters according to the progress of rehabilitation. During the patient's development, such an interface is useful for debugging as well.
A range of sensors is installed in BEAR-H to capture the change of human limbs and monitor the robot's posture. The robot refers to the multimodal sensory data and then Intelligently recognizes the gait of the patient to output proper assistance at different joints. Specifically, the encoders are used to measure the flexion/extension of the robot's joints, the inertial measurement units are used to measure the angle of the lower limb and the acceleration information, and the pressure sensors embedded in both shoes are used to inspect the contact status.
BEAR-H provides three rehabilitation modes, i.e., weight support, training, and intelligent interaction. The differences among multiple modes, summarized in Table 2, are as follows:
The weight support mode is designed for early stages of rehabilitation. Under this mode, the motion at the patient's healthy side is converted to the trajectory for the disabled side according to the walking pattern between two legs. The robot is controlled to track the trajectory to provide assistance for the disabled side while compensating for the weight of the healthy side.
In the training mode, the robot is controlled to follow a predefined trajectory to provide assistance for the patient. The therapist monitors the progress of the training and adjusts the frequency and level of assistance if necessary. This mode is designed for patients whose motor functionabilities are partially restored.
In the intelligent interaction mode, the level of assistance and frequency of the robot's trajectory is automatically adjusted online by monitoring the patient's walking pattern. This mode is designed for rehabilitation at the recovery stage, where the patient is able to walk independently with several abnormal patterns.
These multiple modes can meet the demand of patients with different backgrounds, and a criterion is also proposed to identify the patient's degree of severity (see Table 2). Both the of training and intelligent interaction belong to active training modes, which is important for pushing the patient to master the normal and regular gait pattern. For patients with dyskinesias, active training is even more significant in the sense that a closed loop is established in the central neural system to repair the disordered functionality and lead to a positive long-term training effect.
The following development of this article mainly focuses on the intelligent interaction mode to specify generation of the intelligent gait pattern. Note that the initial frequency of the robot's trajectory in the intelligent interaction mode is set as 0.3 Hz, and it is progressively scaled up or down to suit the actual walking frequency of the patient. When the rate of synchronization is higher than 75%, the corresponding frequency and level of assistance is displayed in the touch panel. In addition to the aforementioned modes, BEAR-H also has the ability of customizing a highly individualized gait according to specific body parameters, which can better suit patients with different conditions.
Wearing a robotic exoskeleton involves substantial physical interactions, hence the safety of patients is always the main concern. Many robotic exoskeletons in the market are driven with rigid actuators (see Table 1). Although the structure of rigid actuators is relatively simple, it is unable to provide additional flexibility for the movement of the human body. More importantly, as the output force of the rigid actuator is directly transferred to the robot joint, forces that are too large might cause physical damage to the human body.
BEAR-H is designed with the deployment of compliant actuators, which makes it safe from a structural point of view. The compliant actuator is developed by referring to the concept of series elastic actuator [15], where an elastic element is installed between the driving motor and the robot joint. This design has the following attractive features:
The elastic property is explored to store the excessive energy and thus counterbalance too large an impact.
The output torque/force can be obtained by simply measuring the deflection of the elastic element (usually with the encoders' reading on both ends of the element).
The actuator is backdrivable such that a certain level of free motion is allowed for the patient.
Although various designs of compliant actuators have been reported in the literature, BEAR-H is the first commercialized exoskeleton robot (i.e., a medical product with an official registration) that converts the prototype of compliant actuators into an actual implementation. The structure of compliant actuators used in BEAR-H is illustrated in Figure 2. The motor drives the motion of the rope coiled around a spool, and the curved spring is then pulled by the rope via a rotary table, and then the motion is transmitted to the output rod so as to realize the rotation of the robot joint by the end. Such a design achieves a balance between the small size, lightweight, high-density output, and relatively low cost, thus guaranteeing the actual implementation. Table 3 summarizes the specifications of the developed compliant actuators.
BEAR-H considers seven phases of the human gait, as listed in Table 4, and each phase is initiated with a gait event. The seven phases of healthy subjects nearly follow the same sequence, and the duration of each phase is also similar. For patients, abnormal patterns might be detected in the gait, e.g., one or several missed phases, too-long or too-short durations, or asymmetry between both legs. Consequently, phases of the gait can be treated as a measurement to assess the quality of rehabilitation. The gait events are depicted in Figure 3.
In BEAR-H, a hidden Markov model (HMM) with seven states is initialized to describe seven gait phases, that is, loading response, midstance, terminal stance, preswing, initial swing, midswing, and terminal swing, which represent different gait events (see Table 4). The HMM can be described with the tuple of parameters as λ=(π, A, B), where π, A, B denote the prior probability, state-transition probability, and observation probability, respectively. Then, the observation vector is defined as
zt=[ωR, αR, ωL, αL, θRknee, θRshank], (1)
where ωR and ωL denote the angular speed of both feet, αR and αL are their first-order derivatives (i.e., acceleration), and iRknee and iRshank are the angles of the right knee and right shank, respectively. Note that the value of those variables can be obtained with sensors mounted on the human limb.
Given the observation vector (1), the Baum–Welch learning algorithm [16] is used to find the optimal value of model parameters by solving λ*= argmaxλ P(zt; λ), where P(·) is the probability. This is basically achieved by following Expectation-Maximization steps, that is, (Expectation step): to calculate the expectation of the states and (Maximization step): to estimate the model parameters by maximizing the expectation. Then, with the well-trained HMM, the most likely gait phase during overground walking can be recognized.
To quickly detect specific gait events, the adaptive thresholding scheme was proposed in our previous works [17], where a linear regression model was introduced to approximate the relationship between three signal inputs (i.e., thigh angle, thigh angular velocity, and forward axial acceleration) and the output value of threshold. The proposed scheme has been validated in the event detection of heel rise and toe-off [17].
The detected events can be used to adjust the robot's reference trajectory online for synchronization between robot and the patient. To achieve this, an adaptive oscillator was developed for BEAR-H in [16]. The adaptive oscillator refers to the gait events to update its frequency as
ω= ε·(Ω-φ)+∑∞k=0 δ(t-kT)·Q·G(ω, Δφ), (2)
where ω denotes the frequency; ε denotes the adaptation rate; ∑∞k=0 δ(·) is a periodic Dirac delta function describing the occurrence of gait event; Ω is the corresponding phase of actual gait; T is the period of gait cycle; Q is a positive value related to the amount of the frequency change; Δφ=((φ-Ω)/2)+2π representing the phase difference between the adaptive oscillator and the actual gait; and G(·) is a function curving the phase response defined as
G(ω, Δφ)=max(sin(Δφ)/2π,0)(ωmin-ω)-min(sin(Δφ)/2π,0)(ωmax-ω), (3)
where ωmax and ωmin represent the upper and lower bounds of frequency, respectively. According to (3), when 0<φ-Ω≤2π, the oscillator leads the gait event, sin(Δφ)≥0, G(·)＜0, its frequency will be lowered; when -2π≤φ-Ω＜0, the oscillator lags behind the gait event, sin(Δφ)≤0, G(·)>0, its frequency will be scaled up.
Next, the adaptive oscillator phase is dependent on the adjusted frequency as φ=ω, where phase φ varies between zero and 2π and grows uniformly within one cycle. The adaptive oscillator phase is referred to so as to compute a stride percentage as stride %= (φ/2π)×100. Then, the trajectory of a robot can be adjusted by normalizing a reference trajectory (e.g., the individualized gait trajectory in the subsequent development) with the stride percentage, based on a lookup table. The normalized trajectory can now match the actual gait of the patient, i.e., starting from the same gait event under the same frequency. Therefore, the synchronization between the robot and the patient is realized. The aforementioned workflow is illustrated in Figure 3.
The robot tracks the reference trajectory to assist the patient under the desired impedance model. Compared to the position control, the impedance control allows a certain level of deviation from the reference trajectory and hence provides the freely moving space, which can encourage the patient’s voluntary efforts, thus benefiting the rehabilitation process. The control objective is specified as the desired impedance model in joint space as
Md=(q-qd)+Cd(q-qd)+Kd(q-qd)=τe, (4)
where q∈ℜn denotes the vector of the joint angles of the robot; n is the number of DoF; qd∈Rn is the reference trajectory generated with the aforementioned algorithm (i.e., the gait recognition, event detection, and adaptive oscillator); Md, Cd, Kd∈ℜn×n represent the parameters of the desired inertia, damping, and stiffness in the desired impedance model, respectively; and τe∈ℜn is the torque due to the physical interaction between human and robot, which is measured with force gauges mounted on human joints during rehabilitation. Equation (4) describes a dynamic relationship between the interaction torque and the joint angles of robot.
Note that the desired impedance model can be rewritten as
(q-qd)+M-1d Cd(q-qd)+M-1d Kd(q-qd)-M-1d τe=ż+Γz, (5)
where z=q-qz=q-qd+Λ(q-qd)-τl qz=qd-Λ(q-qd)+τl is a reference vector, and Λ+Γ=M-1d Cd, Λ·Γ=M-1d Kd, and τ1+Λτl=M-1d τe. That is, Λ and Γ are two matrices that are derived from the impedance parameters, and τ1 is the low-pass filtered signal of τe. From (5), the objective for impedance control can now be reformulated as z→0 as t →∞, guaranteeing the realization of the desired impedance model in the low-frequency range.
Next, the dynamic model of a compliantly actuated exoskeleton robot is described as [18]
M(q)q+C(q,q)q+g(q)=K(θ-q)+τe, (6)
Bθ+K(θ-q)=u, (7)
Where M(q)∈ℜn×n is the inertia matrix, C(q-q)∈ℜn×n is a matrix related to the Centrifugal and Coriolis forces, and g (q)∈ℜn is a vector of torques due to gravity, θ∈ℜn denotes the vector of the rotation of the driving motor, K∈ℜn×n is the matrix of spring stiffness, B ∈ℜn×n is the inertia matrix of the motor, and u∈ℜn is the control input. The control input u corresponds to the driving motor's output torque and is generated according to the proposed control algorithm.
Note that (6) and (7) describe a high-order nonlinear system, where both subsystems are coupled with the output torque of the actuator [i.e., K(θ-q)]. The controller development for such a system is not trivial because 1) both subsystems should be stabilized and 2) the actual control input is not exerted on the robot joint but the driving motor instead. To fulfill the requirement, an adaptive impedance control scheme is developed by following the approach of backstepping [18]. Specifically, the dynamic model of the subsystem (6) can be rewritten as
M(q)q+C(q,q)q+g(q)+Kq=Kθd+KΔθ+τe, (8)
where Δθ=θ-θd, and θd denotes a virtual control input. The term Δθ can be treated as a perturbation to the subsystem. When the actual control input u is designed to ensure that state θ converges to the virtual control input (i.e., θ→θd), that is, the perturbation term Δθ→0, then the virtual control input θd works to ensure that z→0 and hence the realization of the desired impedance model.
Then the virtual control input is proposed as
θd=q-K-1[τe+Kzz-Y(qz, qz, q, q)ψ], (9)
where θd∈ℜn is the virtual control input and Kz∈ℜn×n is a diagonal and positive definite matrix, which represents the control parameter; Y(·)∈ℜn×m symbolizes a dynamic regressor, which is dependent on the joint angle, joint velocity, reference trajectory, and the interaction torque; ψ∈ℜm is the vector of uncertain dynamic parameters; and m denotes the number of parameters. The uncertain parameters are updated with the online adaption law as ψ∈－LYT(·)z, where L∈ℜm×m governs the rate of parameter convergence. Note that the update of parameters is driven by impedance vector z, and therefore, it automatically stops after the desired impedance mode is realized, that is, z→0.
Next, the actual control input u is specified as
u=K(θ-q)-Kss+Bθr, (10)
where Ks is a positive definite and diagonal matrix, s=θ-θr=θ-θd+α(θ-θd) is a sliding vector, θr=θ-θd-α(θ-θd) is another reference vector, and α is a positive constant. The purpose of control input is to guarantee that the state of the actuator subsystem converges to the virtual control input, i.e., θ→θd such that the functionality of the virtual control input is guaranteed. It can be proved that the proposed control scheme described by (9) and (10) ensures the realization of the desired impedance model (4) in the presence of uncertain dynamic parameters.
A cloud-based management platform is established for BEAR-H (see Figure 4), and the platform structure is based on the Internet of Things. It can receive and store the real-time training data of patients and then allow doctors to master the situation and progress of rehabilitation. With some algorithms of intelligent diagnosis, doctors can quickly provide feedback to the patient via remote access, laying the foundation of home rehabilitation. Additionally, the established cloud-based platform provides the following multiple interfaces.
The doctor can use the mobile application to check the training record and training report of patients or control the robot to adjust the strategy of rehabilitation before the next course of treatment.
As the platform visualizes all of the real-time data, the doctor can closely monitor the robot's motions (e.g., trajectory of the robot’s joints and level of assistance) and then provide immediate feedback.
The platform allows the user to revisit the history data (which are stored as personalized archives) and thus have an overview of the whole procedure of rehabilitation; the platform can also be used as the interface for remote maintenance and fault diagnosis.
Additionally, the platform stores healthy subjects' data, that is, the body parameters (e.g., height, age, wrist width, and so on) and gait trajectories under different walking speeds. By referring to the large data set, a neural network (NN) is constructed in BEAR-H for the generation of individualized gaits, under the structure shown in Figure 5. The overall structure consists of four modules: input, feature extraction, mapping, and output, and the workflow can be summarized as follows:
The NN is constructed to realize a probabilistic mapping from the body parameters of patients to the features of gait trajectory. The features are designed as a series of coefficients of Fourier transform of the gait trajectory, which can then be reconstructed by computing the Fourier inversion with the reference to the predicted features (i.e., outputs of NNs).
The training data set is collected from healthy subjects. Specifically, the gait trajectory of the subject is recorded using anthropometry methods and set as the ground truth. The multiscale body parameters (e.g., height, weight, leg length, and so forth) are measured and then fed into the NN. The NN outputs the predicted gait features, which are compared with the ground truth to update the weights and train the NN.
The well-trained NN receives body parameters of the patient and then generates an individualized gait trajectory (i.e., the representation with gait features) for a specific subject corresponding to the trajectory of a healthy subject with similar body parameters. As the body parameters vary for different subjects, the individualized gait is also varied and automatically adjusted by the trained NN.
Note that the backbone network is based on a log-linearized Gaussian mixture network, and the training and implementation phases are illustrated in Figure 6. Therefore, the individualized gait is actually generated by estimating the posterior probability of the gait in a data-driven way, given the specific body parameters of a patient. Such a formulation explores the underlying relationship between the body parameters and gait; hence, it can better suit the patient to guarantee his/her natural motion after rehabilitation.
Note that the subject-specific gait is related with EMG signals and the strength of a human limb. However, a stroke patient has significantly different EMG signals and strength data compared with the situation when he/she was in healthy condition. Therefore, that information is not used as the input to retrieve the original healthy gait pattern.
Clinical trials are carried out to study the performance of BEAR-H, where the patient wears the robot and walks in a straight line in a hospital environment (see Figure 7). Before clinical trials, the patients are informed about the purposes of the study and sign informed consent. Then, some preliminary training is performed to help patients get used to wearing the robot, which ensures that the patient has the ability to participate in the subsequent formal trials; the preliminary training lasts for three to five days. Next, the formal trial is performed for five days each week, and the patient walks by wearing the robot twice per day, with each walking trial lasting for approximately 30 min. An experimental protocol was approved on 23 May 2019 by the Ethics Committee of The Second Affiliated Hospital, Shenzhen University. A clinical process was also supported by The Second Affiliated Hospital, Shenzhen University.
A total of 30 patients were involved in the clinical trials, including 23 males and seven females. The patients are selected according to the following criteria:
The subject can walk with some assistance.
The subject's age is between 18 and 75 years old, including both male and female subjects.
The subject has not been previously diagnosed with hemiplegia, and the duration is between two weeks and six months.
The upper limb can function properly and is able to hold the stand (which is used as the walking assistant device).
The walking is not stable, and the speed is significantly slower than healthy subjects.
He/she can understand commands from the doctor and clearly acknowledges all of the training details.
At the same time, the patients with the following conditions will be excluded:
His/her joints can only move within a significantly limited range.
The fracture has not fully recovered, or the osteoporosis is significant.
There are other serious symptoms involved.
The statistical summary of their body parameters and medical condition is given in Table 5. In Table 5, the duration represents the interval since the subject was diagnosed (for example, a duration of 15 means that the subject was diagnosed 15 days ago), and there are two types of hemiplegia causes: cerebral hemorrhage (CH) and cerebral infarctions (CIs). From Table 5, it can be seen that the age is between 22 (young) and 72 (old), the height is between 150 cm (short) and 176 cm (tall), and the weight is between 45 kg (thin) and 80 kg (fat). Hence, the selected patients are representative, in the sense that patients with different body parameters have been considered. The selection also covers different medical conditions, ranging from new cases (e.g., 15 days ago) to long-standing ones (e.g., 162 days ago). In addition, the number of CIs is double that of CH, which is reasonable and matches the occurrence probability in clinical trials.
The formal trial is carried out as follows: 1) BEAR-H is tested to ensure that the system functions properly; 2) the conditions of the subject are evaluated by measuring blood pressure and heart rate, and the trial proceeds if all readings are normal; 3) the clinical staff help the patient to wear the robot, fasten the safety belt of the stand, and then input the body parameters of the patient; 4) the patient starts to walk once the robot is activated; 5) during the walking phase, the doctor and clinical staff continue to monitor the procedure and take action to stop the trial if any safety issues arise; and 6) after the walking phase is complete, the patient takes off the robot, and the staff measures the blood pressure and heart rate of the patient again.
Comparative studies were performed in the clinical trials, where all 30 patients are equally divided in two groups: Group 1 includes patients from number 1 to 15 who were subject to the traditional manual rehabilitation, that is, an experienced therapist holds and assists the patient during the walking test. Group 2 includes patients from number 16 to 30 who were subjected to robotic rehabilitation by wearing BEAR-H with minimal manual assistance (the doctor intervenes only when there might be a safety issue). Both groups have the similar distribution of body parameters and medical conditions given in Table 5.
The performance of BEAR-H was studied in the six-minute walk test (6MWT) [17], where the patient wears the robot and then walks continuously for six minutes. During the walking phase, the doctor accompanies the patient with minimal communication (i.e., the doctor talks to the patient to encourage him/her), and the patient can rest between the two tests. The trials were performed in a closed corridor such that other pedestrians could not interrupt. The walking path is straight and solid, and the total length is more than 30 m. There is a returning sign at the end of the path, and there are also a series of reminders every 5 m.
In the first experiment, the heart rate of the patient before and after 6MWT are recorded and compared. For example, a variation of 10% of heart rate implies that the Patient's heart rate after walking increases for 10% with respect to that before walking. Note that smaller increases in heart rate implies better effect, where the assistance from doctor/robot benefits the patient's motion such that the patient is relaxed while walking. Significant increases in heart rate is a sign of ineffective training, indicating that the patient is not comfortable with the assistance from the doctor/robot and needs to exert additional effort while walking. The experimental results are presented in Figure 8, where the blue bar denotes the heart rate before walking, and the red bar denotes the heart rate after walking. The variation of heart rate for group 1 (i.e., manual rehabilitation) is summarized in Table 6, which is between 5.71 and 14.63%, and the average change of heart rate is 8.89%. The variation of heart rate for group 2 (i.e., robotic rehabilitation) is listed in Table 7, which is between 1.22 and 8.05%, and the average change of heart rate is 4.10%, less than group 1, thus proving the effectiveness of robot.
In the second experiment, the distance that the patient can walk within 6 min is compared before and after training, and a farther distance implies a faster walking speed and hence a better effect of rehabilitation and vice versa. The experimental results for both groups are shown in Figure 9, represented as the improvement of distance (i.e., the difference between the distance of how far the patient can walk after training and that before it). Note that the training does not function for patient number 18, who does not gain progress. This is common in clinical trials, mainly due to the specific condition of patients (e.g., other diseases or extreme discomfort). From Figure 9, it is obtained that the average improvement with manual rehabilitation is 61.07 m and that with robotic rehabilitation is 65.51 m (after the removal of data for number 18), proving the effectiveness of robotic rehabilitation.
In the third experiment, the effect of training is assessed by referring to the Fugl–Meyer assessment for lower extremity (FMA-LE) [19], which has become a standard tool for the evaluation of poststroke recovery. To carry out FMA-LE, the doctor asks the patient to perform a series of tasks, then assesses his/her difficulty in performing such tasks. The performance before and after training is compared, and a higher assessment mark implies a better effect of rehabilitation and vice versa. Experimental results of FMA-LE are shown in Figure 10, represented as the improvement of FMA-LE marks after the training. Note that FMA-LE marks each task with three points, that is,
Score 0: the patient cannot perform the task
Score 1: the patient can perform the task partially
Score 2: the patient can perform the task fully.
The full mark of FMA-LE is 34. Similarly, the rehabilitation does not function for a few patients (i.e., numbers 3, 14, and 28). After the removal of invalid data, the average improvement with manual and robotic rehabilitation is 4.38 and 5.21 marks, respectively, which also proves the effectiveness of robotic rehabilitation. The field performance of BEAR-H can be found in the multimedia attachment.
The current weight of BEAR-H is 24 kg, which is more or less at the same level as EksoNR (27 kg) and ReWalk (23.3 kg). The weight of BEAR-H can be reduced by replacing the battery with a lower-volume and replaceable one. Additionally, a titanium alloy or carbon-fiber material can be used in the mechanical structure of BEAR-H in future versions to further reduce its weight but with a higher cost as the tradeoff. Furthermore, other advanced actuators (e.g., shape-memory alloys [20]) may be implemented in the robot; however, these have a relatively lighter weight but are less stable and reliable due to their nonlinearity, and thus may not be suitable for formal deployment in commercial products at their current stage.
This article reported the development of BEAR-H, a new, intelligent lowerlimb exoskeleton designed for smart rehabilitation. Compared to other rehabilitation robots, BEAR-H has the following intelligent features: 1) The robot can recognize the gait events of the patient and then automatically synchronize the reference trajectory with the actual gait of the patient; 2) a cloud-based management platform is established to allow doctors and family members to monitor the progress of the rehabilitation, and such a platform also provides access to personalized files with history data to quickly understand the overview of the whole procedure; and 3) the individualized gait-generation scheme is proposed to retrieve the original healthy pattern of a specific patient by exploring the big data of different healthy subjects such that the generated gait can well suit the patient and guarantee the effect of rehabilitation. The safety of BEAR-H is guaranteed in both the hardware and software. In terms of hardware, all the joints are driven by compliant actuators (which is first among all commercialized products thus far). For software, an adaptive impedance controller is proposed to regulate the relationship between interaction torque and joint angle in the presence of uncertain dynamics, and the realization of the impedance model also provides a certain level of free moving space, therefore suiting the rehabilitation paradigm of assistance as needed. The performance of BEAR-H has been validated in clinical trials.
This work was supported in part by the National Natural Science Foundation of China under grant 61803321; Institute for Guo Qiang, Tsinghua University; Beijing National Research Center for Information Science and Technology; Guangdong Medical Research Fund under grant A202169, and Shenzhen Science and Technology Program under grant KQTD20200909114235003. The corresponding authors are Gong Chen (gong118911@126.com) and Jing Ye (yejing@milebot.cn).
[1] T. Zhang and H. Huang, "A lower-back robotic exoskeleton: Industrial handling augmentation used to provide spinal support," IEEE Robot. Autom. Mag., vol. 25, no. 2, pp. 95–106, 2018, doi: 10.1109/MRA.2018.2815083.
[2] I. Pacifico et al., "An experimental evaluation of the Proto-MATE: A novel ergonomic upper-limb exoskeleton to reduce workers' physical strain," IEEE Robot. Autom. Mag., vol. 27, no. 1, pp. 54–65, 2020, doi: 10.1109/MRA.2019.2954105.
[3] S. J. Baltrusch, J. H. van Dieen, C. A. M. van Bennekom, and H. Houdijk, "Testing an exoskeleton that helps workers with low-back pain: Less discomfort with the passive SPEXOR trunk device," IEEE Robot. Autom. Mag., vol. 27, no. 1, pp. 66–76, 2020, doi: 10.1109/MRA.2019.2954160.
[4] L. Liu, D. Wang, K. S. Lawrence Wong, and Y. Wang, "Stroke and stroke care in China," Stroke, vol. 42, no. 12, pp. 3651–3654, 2011, doi: 10.1161/STROKEAHA.111.635755.
[5] J. Meuleman, E. van Asseldonk, G. van Oort, H. Rietman, and H. van der Kooij, "LOPES II—Design and evaluation of an admittance controlled gait training robot with shadow-leg approach," IEEE Trans. Neural Syst. Rehabil. Eng., vol. 24, no. 3, pp. 352–363, 2016, doi: 10.1109/TNSRE.2015.2511448.
[6] R. Riener, L. Lunenburger, I. Maier, G. Colombo, and V. Dietz, "Locomotor training in subjects with sensori-motor deficits: An overview of the robotic gait orthosis Lokomat," J. Healthcare Eng., vol. 1, no. 2, pp. 197–216, 2010, doi: 10.1260/2040-2295.1.2.197.
[7] Y. Stauffer et al., "The WalkTrainer - A new generation of walking reeducation device combining orthoses and muscle stimulation," IEEE Trans. Neural Syst. Rehabil. Eng., vol. 17, no. 1, pp. 38–45, 2009, doi: 10.1109/TNSRE.2008.2008288.
[8] S. K. Banala, S. H. Kim, S. K. Agrawal, and J. P. Scholz, "Robot assisted gait training with active leg exoskeleton (ALEX)," in Proc. IEEE/RAS-EMBS Int. Conf. Biomed. Robot. Biomechatron., pp. 653–658, doi: 10.1109/BIOROB.2008.4762885.
[9] A. Kozlowski, T. Bryce, and M. Dijkers, "Time and effort required by persons with spinal cord injury to learn to use a powered exoskeleton for assisted walking," Topics Spinal Cord Injury Rehabil., vol. 21, no. 2, pp. 110–121, 2015, doi: 10.1310/sci2102-110.
[10] C. Hartigan et al., "Mobility outcomes following five training sessions with a powered exoskeleton," Topics Spinal Cord Injury Rehabil., vol. 21, no. 2, pp. 93–99, 2015, doi: 10.1310/sci2102-93.
[11] A. Tsukahara, Y. Hasegawa, K. Eguchi, and Y. Sankai, 'Restoration of gait for spinal cordinjury patients using HAL with intention estimator for preferable swing speed,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 23, no. 2, pp. 308–318, 2015, doi: 10.1109/TNSRE.2014.2364618.
[12] A. Esquenazi, M. Talaty, A. Packel, and M. Saulino, “The ReWalk powered exoskeleton to restore ambulatory function to individuals with thoracic-level motor-complete spinal cord injury," Amer. J. Phys. Med. Rehabil., vol. 91, no. 11, pp. 911–921, 2012, doi: 10.1097/PHM.0b013e318269d9a3.
[13] J. Zhang, T. Wang, Q. Zhao, and S. Liu, "Impacts of a lower limb exoskeleton robot on the muscle strength of tibialis anterior muscle in stroke patients," in Proc. Int. Conf. Energy, Environ. Bioeng., 2020, pp. 1–4. doi: 10.1051/e3sconf/202018503036.
[14] MileBot. http://milebot.com.cn (accessed 2021).
[15] G. A. Pratt and M. M. Williamson, "Series elastic actuators," in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 1995, pp. 399–406, doi: 10.1109/IROS.1995.525827.
[16] G. Chen, P. Qi, Z. Guo, and H. Yu, "Gait-event-based synchronization method for gait rehabilitation robots via a bioinspired adaptive oscillator," IEEE Trans. Biomed. Eng., vol. 64, no. 6, pp. 1345–1356, 2016, doi: 10.1109/TBME.2016.2604340.
[17] J. Ye et al., "An adaptive method for gait event detection of gait rehabilitation robots," Frontiers Neurorobot., vol. 14, p. 38, Jul. 2020, doi: 10.3389/fnbot.2020.00038.
[18] X. Li, Y. Pan, G. Chen, and H. Yu, "Adaptive human-robot interaction control for robots driven by series elastic actuators," IEEE Trans. Robot., vol. 33, no. 1, pp. 169–182, 2016, doi: 10.1109/TRO.2016.2626479.
[19] A. Fugl-Meyer, "The post-stroke hemiplegic patient. 1. a method for evaluation of physical performance," Scand. J. Rehabil. Med., vol. 7, no. 1, pp. 13–31, 1975.
[20] X. Liu, H. Liu, and J. Tan, "Actuation frequency modeling and prediction for shape memory alloy actuators," IEEE/ASME Trans. Mechatronics, vol. 26, no. 3, pp. 1536–1546, 2021, doi: 10.1109/TMECH.2020.3023097.
Xiang Li, Department of Automation, Tsinghua University, Beijing, 100084, China. Email: xiangli@tsinghua.edu.cn.
Xuan Zhang, Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, 518055, China. Email: 1729980156@qq.com.
Xiu Li, Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, 518055, China. Email: li.xiu@sz.tsinghua.edu.cn.
Jianjun Long, Rehabilitation Center, The First Affiliated Hospital, Shenzhen University, Shenzhen, 518025, China. Email: longjianjun@szu.edu.cn.
Jian'an Li, Department of Rehabilitation Medicine, The First Affiliated Hospital (Jiangsu Province Hospital), Nanjing Medical University, Nanjing, 210049, China. Email: lijianan@carm.org.cn.
Lanshuai Xu, Shenzhen MileBot Robotics Co., Ltd., Shenzhen, 518055, China. Email: xulanshuai0123@163.com.
Gong Chen, Shenzhen MileBot Robotics Co., Ltd., Shenzhen, 518055, China. Email: gong118911@126.com.
Jing Ye, Shenzhen MileBot Robotics Co., Ltd., Shenzhen, 518055, China. Email: yejing@milebot.cn.
English
