Upper-Limb Rehabilitation with a Dual-Mode Individualized Exoskeleton Robot: A Generative-Model-Based Solution

Exoskeleton robots offer several advantages in stroke rehabilitation. In particular, they enable precise control of movement, consistent and repeatable therapy sessions, and collection of real-time data on patient progress (Huang and Krakauer, 2009; Jezernik et al., 2004; Said et al., 2022). Furthermore, they offer high-intensity training while reducing physical burdens on therapists, and they are adaptable to the needs and recovery stages of individual patients (Gull et al., 2020). This paper focuses on rehabilitation using upper-limb exoskeleton robots, a training task that is the focus of several prototypical or commercially available products. Among these products, the Harmony (Kim and Deshpande, 2017) stands out, as it is equipped with an anatomical shoulder mechanism designed to augment and facilitate the arm’s natural movements. Additionally, ANYexo (Zimmermann et al., 2023a) enhances the upper limb mobility of patients and exhibits sufficient flexibility to encompass a wide array of daily activities. Furthermore, the Armeo Power (Hocoma) (Lee et al., 2020) enables early-stage stroke patients to start intensive arm therapy. However, current upper-limb exoskeleton robots may not fully achieve the main goal of upper-limb rehabilitation, i.e., recovery of the manipulability of the human body with high degrees of freedom (DoFs), as they have too few DoFs to match the movement of a healthy subject and a rather low capacity for individualization. Individualization is a key feature of an upper-limb exoskeleton robot used in rehabilitation, as it enables the provision of assistance that is customized to the condition (e.g., stroke duration and medical background) of the patient. Such assistance helps the patient to regain movement ability that is original and natural, i.e., devoid of abnormal patterns. Hence, insufficient individualization results in low-quality rehabilitation training.

Generally, either active mirroring or passive following training modes are applied in the rehabilitation of stroke patients, according to their medical condition and as illustrated in Figure 1. Active mirroring training is used for patients who have mild impairments and an unaffected side of the body and thus retain motor abilities. In this mode, a generative model known as the intention predictor estimates the patient’s motion intentions based on historical motion data. Subsequently, these intentions are mirrored by an upper-limb exoskeleton robot on the affected side of the patient’s body. This process aims to align motion intentions between the unaffected and affected sides of the body at a neural level, thereby increasing patient engagement. Passive following training is used for patients with severely impaired motor abilities. In this mode, the patient follows a pre-defined trajectory facilitated by an upper-limb exoskeleton robot, which provides structured guidance for movement rehabilitation. The aforementioned modes can also be utilized together with manual training for patients with different medical conditions or in different stages of stroke. For example, using the passive mode at the early stage and then the active one, when the healthy side can move naturally. Such dual-mode rehabilitation is important for ensuring that an upper-limb exoskeleton robot can be used by multiple patients with various levels of physical ability.

As patients wear an upper-limb exoskeleton robot to carry out training procedures, ensuring patient safety in the presence of tight and continuous physical interactions is a primary concern. Much progress has been made in this regard, in terms of both hardware and software. In terms of hardware, safety is increased by using compliant actuators such as series elastic actuators (SEAs) (Pratt and Williamson, 1995) to absorb impact forces during interactions. Additionally, bio-inspired cable-driven actuators (Xu et al., 2023) are used to transition between different actuation modes, and safety is further guaranteed by backdrivable actuators (Zhu et al., 2021). In terms of software, safety regulation can be achieved through various control strategies, including impedance control (Li et al., 2018b), velocity field-based control (Martinez et al., 2018), and data-driven ergonomic control (Clark and Amor, 2022), all of which ensure that an upper-limb exoskeleton robot system operates safely. However, most hardware and software approaches are applied after a safety problem has occurred. A better approach is to predict potential safety problems and prevent their occurrence.

Accordingly, this paper describes the development of a new upper-limb exoskeleton robot for rehabilitation training of stroke patients that is superior to previously reported upper-limb exoskeleton robots in terms of the three characteristics described below.

1. 1.

   Safety: The new upper-limb exoskeleton robot contains SEAs in most of its active joints, so that the joints absorb impact forces caused by collisions and unexpected contacts during upper-limb interaction. This design enhances the precision of torque control, thereby providing smoother assistance in rehabilitation training than previously reported upper-limb exoskeleton robots. Moreover, an anomaly detection neural network informed by interactive feedback is implemented to assess the safety and naturalness of the movement of the upper-limb exoskeleton robot in real time. This network enables the early detection of anomalies, allowing for their prevention or mitigation during rehabilitation tasks. Furthermore, an online trajectory refinement module and an impedance controller are incorporated into the upper-limb exoskeleton robot to ensure safety throughout the planning and execution phases of rehabilitation.

2. 2.

   Effectiveness: The upper-limb exoskeleton robot is configured with one passive joint and five active joints, endowing it with a high number of DoFs and thus enabling the workspace to accommodate a broad spectrum of daily activities. Given the inherent uncertainty and randomness associated with human motion intentions and upper-limb movements, the system efficiently generates trajectories by exploiting a probabilistic model of motion for explicit consideration during the sampling process so as to match the nature of human intention. The effectiveness of the upper-limb exoskeleton robot was validated through clinical trials, which demonstrated that the experimental group, which underwent passive following training, recovered motor abilities more quickly than the control group.

3. 3.

   Friendliness: The active joints of the upper-limb exoskeleton robot are capable of sensing force and thus can detect interaction torque. Therefore, when the upper-limb exoskeleton robot is operated in transparent mode for demonstration purposes, it can be maneuvered effortlessly by the patient. Moreover, given the substantial load sustained by the shoulder joint during upper-limb movements, a direct-drive motor is incorporated into this joint in the upper-limb exoskeleton robot. This motor delivers enhanced driving torque and increases backdrivability by eliminating nonlinear friction forces. Furthermore, to reduce the effects of inertia during movement and thus increase comfort, a cable-driven mechanism is utilized in joints that must have a broader range of motion in space than other joints.

The key novelty of this study is its design and implementation of a generative-model-based refinement framework that is capable of generating highly individualized trajectories and ensuring safety. Specifically, the framework operates in either of the two aforementioned distinct modes, namely active mirroring mode and passive following mode.

The active mirroring mode exhibits the newly developed key features described below.

• -

  A novel diffusion model-based motion intention predictor, which uses historical motion data of the patient’s unaffected side to estimate the patient’s upper-limb motion intentions. This allows for the prediction of future trajectories and establishes the mean and variance of the patient’s motion intentions.

• -

  A preemptive tuning algorithm that ensures that a predicted trajectory remains within a safe region by exploiting the predicted distribution to mitigate potential risks.

The passive following exhibits the newly developed key features described below.

• -

  A diffusion model-based anomaly detection network capable of evaluating safety and naturalness based on an anomaly score. This score is utilized to guide the online trajectory refinement and evaluation for each rehabilitation task.

• -

  A probabilistic movement primitives (ProMPs)-based approach to trajectory generation that captures the distribution of the patient’s motion intention. In addition, based on each sampling result’s performance, this approach iteratively optimizes the assistive trajectory.

Both modes customize the assistance and generate an individualized trajectory for the patient, enabling natural and original motion patterns to be recovered by the patient. In particular, the intention predictor and anomaly detector utilize generative models to capture patterns in the latent space of large datasets. Moreover, the intention predictor and anomaly detector have distinct functions. The intention predictor accurately predicts the patient’s original motion intentions. This prediction serves as a self-mirrored trajectory that is finely tailored to the needs of the patient’s affected side and thus aligns closely with the patient’s motion intentions. Simultaneously, the anomaly detector quantitatively identifies the differences between the patient’s rehabilitation movements and those of healthy individuals. These differences are used to guide the adjustment of rehabilitation exercises, thereby enhancing the individualization of the treatment.

In a previous study (Chen et al., 2023b), we developed an individualization framework for passive following training using a variational autoencoder (VAE)-based anomaly detector. In the current study, we expand this approach to a dual-mode upper-limb exoskeleton robot by incorporating generative model technology, i.e., by utilizing a diffusion model. We present results from a series of experiments and comparisons that confirm the safety and performance of the new upper-limb exoskeleton robot. In addition, we present results from a clinical trial that validate its effectiveness and friendliness.

This section reviews related work on upper-limb exoskeleton robots, trajectory generation, and interaction control. Unlike those that have been previously reported, the upper-limb exoskeleton robot reported in this paper has an efficient generative capability, which enables it to effectively provide individualized assistance.

This study examined a prototype cable-driven upper-limb exoskeleton robot that was co-developed by Tsinghua University and Shenzhen MileBot Robotics Co., Ltd. A computer-aided design model of the upper-limb exoskeleton robot is illustrated in Figure 2. The upper-limb exoskeleton robot is equipped with one passive and five active joints that allow refinement of the upper-arm and forearm lengths within ranges of 20 and 40 cm, respectively. The passive DoF, labeled as Joint 0, accommodates the eccentric movements of the shoulder joint and thereby eliminates the limitations of the upper-limb motion range and increases the space for patient motion training. The active joints are numbered 1 to 5 and serve specific functions, as described below.

1. -

   Joint 1: shoulder abduction and adduction

2. -

   Joint 2: shoulder flexion and extension

3. -

   Joint 3: upper-arm internal and external rotation

4. -

   Joint 4: elbow flexion and extension

5. -

   Joint 5: forearm internal and external rotation

A block diagram of the upper-limb exoskeleton robot is shown in Figure 3. Joints 1 and 2 consist of direct-drive joint modules (RJSIIT-17-RevB2 and RJSIIT-17-RevB5, respectively), each of which is outfitted with force sensors and supported by the frame. Joints 3 to 5 utilize harmonic deceleration servo motors (AK80-64) that are configured as cable-driven SEAs and equipped with torque sensors (TK17-191151) and encoders (3590S-2-104L for the internal and external rotation joints, and QY2204-SSI for the elbow joint). Additionally, these joints incorporate two potentiometers for measuring spring compression during motion. The detailed technical specifications of the upper-limb exoskeleton robot are provided in Table 2. Motor drivers are implemented to control Joints 3–5, and together with two joint modules, facilitate communication with the main board through a controller area network. The main board receives commands from a personal computer (PC) via a universal asynchronous receiver/transmitter.

The upper-limb exoskeleton robot employs a hierarchical control system. At the low level, the main board executes motion control and environmental perception, enabling precise force control and real-time acquisition of sensory data. At the high level, a Linux system on the PC operates a motion planning module and performs computationally intensive tasks, such as dynamic calculation and network inference. All programs are integrated into Robot Operating System (ROS) (Quigley et al., 2009) and run concurrently.

The joints that undergo significant spatial movement, namely Joints 3, 4, and 5, utilize our self-developed SEAs equipped with cable-driven mechanisms. A working principle diagram of a SEA is provided in Figure 4. Each joint motor is outfitted with two sets of pulleys, with each set corresponding to a different rotational direction. Each end of the motor output shaft is affixed with a steel cable that is connected to a potentiometer to constitute the SEA. The potentiometer assembly includes two fixed blocks, namely fixed block 1 and fixed block 2, and a spring component. The cable from the motor end (depicted in red) is connected via a spring to fixed block 1, while another cable (depicted in blue) connects the moving joint components to fixed block 2. As the motor applies torque, fixed block 1 moves along a sleeve attached to fixed block 2, thereby compressing the spring and enabling the measurement of tension within the cable. Upon spring compression, the blocks form a unified structure that moves in synchrony with and in the same direction as the joint, thereby facilitating overall movement.

For safety, the upper-limb exoskeleton robot is equipped with an emergency stop button that physically disconnects power during urgent situations. Additional safety measures are incorporated at the software level, including measures that limit the range of motion, prevent self-collision, restrict joint velocity, and automatically cease operation in cases of excessive interaction force or torque.

A dynamic model of a SEA-driven upper-limb exoskeleton robot can be defined as follows (Albu-Schäffer et al., 2007):

where $\bm{q}\in\Re^{5}$ is the vector of joint positions; $\bm{\theta}\in\Re^{3}$ is the vector of motor-rotor-shaft positions; and $\bm{M}(\bm{q})\in\Re^{5\times 5}$, $\bm{C}(\dot{\bm{q}},\bm{q})\dot{\bm{q}}\in\Re^{5\times 5}$, and $\bm{g}(\bm{q})\in\Re^{5}$ are the inertia matrix, centripetal and Coriolis torques, and gravitational torques of the robot, respectively. The selection matrices $\bm{S}_{1}=\text{diag}(1,1,0,0,0)\in\Re^{5\times 5}$ and $\bm{S}_{2}=[\bm{0},\bm{I}_{3}]\in\Re^{3\times 5}$ are used to separate the cable-driven joints. In addition, $\bm{K}\in\Re^{3\times 3}$ is the stiffness matrix, $\bm{B}\in\Re^{3\times 3}$ is the inertia matrix of the motor, $\bm{\tau}_{e}\in\Re^{5}$ is the physical interaction torque vector, $\bm{\tau}_{f}\in\Re^{3}$ is the disturbance due to the cable transmission and the friction of the joint system, and $\bm{u}\in\Re^{5}$ is the control torque applied to the actuator.

The dynamic model of the overall system is described by (1) and (2) and exhibits different time-scales. Specifically, the SEA subsystem operates on a fast time-scale, while the upper-limb exoskeleton robot subsystem functions on a slow time-scale. To control this system effectively, the control input is formulated according to singular perturbation theory (Spong, 1987), as follows:

where $\bm{u}_{f}$ is the fast time-scale control term for stabilizing the model defined by (2), and $\bm{u}_{s}$ is the slow time-scale control term for stabilizing the model defined by (1). A representative form of $\bm{u}_{f}$ is expressed as follows:

where $\bm{K}_{v}\in\Re^{3\times 3}$ is a diagonal and positive-definite matrix.

Substituting (3) and (4) into (2) yields

which can be rewritten as follows by defining $\bm{\tau}_{o}=\bm{K}(\bm{\theta}-\bm{S}_{2}\bm{q})$:

By introducing $\bm{K}=\bm{K}_{1}/\varepsilon^{2}$ and $\bm{K}_{v}=\bm{K}_{2}/\varepsilon$, with $\varepsilon$ being a small positive parameter, (6) can be written as follows:

When $\varepsilon=0$, the solution of (7) is $\bar{\bm{\tau}}_{o}=\bm{S}_{2}\bm{u}_{s}-\bm{B}\bm{S}_{2}\ddot{\bm{q}}$.

If the fast time-scale is set as $\gamma=\frac{t}{\varepsilon}$, $\bar{\bm{\tau}}_{o}$ is achieved at $\gamma\rightarrow\infty$. $\bar{\bm{\tau}}_{o}$ remains constant at $\varepsilon=0$. Next, we introduce a new variable $\bm{\eta}=\bm{\tau}_{o}-\bar{\bm{\tau}}_{o}$ to rewrite (7) on the fast time-scale, as follows:

which defines the boundary-layer system.

Substituting the solution of (8) into (1), affords a quasi-steady-state system that captures the slow dynamics of the overall system. These dynamics can be expressed as follows:

where $\bar{\bm{B}}=\bm{S}_{2}^{\mathsf{T}}\bm{B}\bm{S}_{2}$ is the projected motor inertia matrix.

According to singular perturbation theory, the stability of the overall system is guaranteed if the boundary-layer system and the quasi-steady-state system are both exponentially stable. A stability analysis of this system is provided in the Appendix.

The overall dynamic model described by (1) and (2) is a high-order system, where (1) represents the rigid-joint side, and (2) represents the compliant actuator side. It is non-trivial to stabilize and control such a system.

An upper-limb exoskeleton robot needs to guide the patient to perform repetitive motions via close interaction to help the patient to regain motor function. Thus, the upper-limb exoskeleton robot is controlled to track a time-varying trajectory in accordance with the following impedance model:

where $\bm{q}_{d}\in\Re^{n}$ is the vector of the desired trajectory, and $\bm{M}_{d},\bm{C}_{d},\bm{K}_{d}\in\Re^{n\times n}$ are the desired inertia, the desired damping, and the desired stiffness matrices, respectively, which are diagonal and positive-definite. Tracking the desired trajectory in accordance with the impedance model allows the patient to deviate from the trajectory and hence provides a certain level of compliance to improve safety.

The proposed dual-mode training framework is illustrated in Figure 5. Both the anomaly detector and intention predictor are designed using trained generative models based on a dataset of motions of healthy individuals. These models are designed to capture the motion intentions of the unaffected side of the patient and evaluate the comfort and naturalness of the current patient–robot interaction. Individualization is facilitated by employing two methods, i.e., employing the anomaly score to guide the refinement of a desired trajectory that complies with dynamic constraints, and employing ProMPs to integrate various assistive trajectories into a personalized assistance distribution. Additionally, the safety of the motion process is enhanced by adjusting the impedance parameters according to the anomaly score, thereby ensuring a secure and effective rehabilitation environment.

This section introduces the new trajectory generation method for the upper-limb exoskeleton robot. This method is based on the generative models and can be refined online to suit the patient while maintaining safety and individualized features, and hence maintaining effectiveness. First, in both training modes (i.e., active mirroring and passive following modes), we introduce a general integrated vector $\bm{x}_{d}=[\bm{q}_{d}^{\mathsf{T}},\dot{\bm{q}}_{d}^{\mathsf{T}}]^{\mathsf{T}}$, which encapsulates a sequence of discretized trajectory points within the joint space, thereby parameterizing the trajectory. The following quadratic programming problem is formulated to describe the planning process:

where $N_{p}$ is the predictive horizon, $\bm{Q}$ and $\bm{R}$ are symmetric positive-definite weighting matrices, $\bm{u}_{d}$ is the acceleration of the desired trajectory, $\bm{C}_{a}$ and $\bm{K}_{a}$ are the varying impedance parameters to be defined later, and $f(\cdot)$ is a functional representation of the anomaly detection network. In addition, the superscript indicates the time-step, and $\mathcal{X}$ and $\mathcal{U}$ are the sets of trajectory space constraints and acceleration constraints, respectively. Thus, (11) enables various constraints and safety measures to be included in both training modes.

The joint below the shoulder joints is cable-driven, which decreases the inertia associated with movement and thus increases comfort. Given the presence of these joints, there is substantial friction, which can hinder the movement of joints and therefore should be considerably compensated for.

To achieve this, the upper-limb exoskeleton robot first performs movements in the absence of the patient, where $\bm{\tau}_{e}=\bm{0}$. In addition, the disturbance torque can be parameterized as follows (De Wit et al., 1995):

where $\bm{a}_{f},\bm{b}_{f},\bm{c}_{f},\bm{d}_{f}$ are the unknown parameters, $\bar{\bm{a}}_{f},\bar{\bm{b}}_{f},\bar{\bm{c}}_{f}$ are derived from the Taylor expansion as simplifications for the model, $\odot$ denotes the Kronecker product, $\bm{sgn}(\cdot)$ represents a sign function, $\bm{Y}(\cdot)$ represents a regressor matrix, and $\bm{\zeta}$ represents the vector of model parameters. The approximation presented in (40) is reasonable because the velocities of the joints of the upper-limb exoskeleton robot remain rather low during a given rehabilitation process.

As the upper-limb exoskeleton robot is equipped with force sensors on all of its active joints, friction can be directly measured and recorded in the absence of a patient. Together with the recorded joints’ velocities, the parameters of friction model are learned via polynomial fitting. The estimated friction is represented as follows

Once the friction is estimated, a variable impedance model is proposed. This model must be capable of identifying the human–robot interaction condition and addressing the conflict during rehabilitation. These capabilities enable the model to regulate the action of the upper-limb exoskeleton robot.

To regulate the impedance model, a weighting function (Zhang et al., 2023) is introduced to consider the anomaly score. This function is mathematically defined as follows:

where $\lambda_{1}$ and $\lambda_{2}$ are positive constants that determine the range and median of the weighting function, respectively, $\chi_{1}$ is a constant that normalizes the anomaly score into a specified small range, and $\chi_{2}$ is the offset of the weighting function from the origin of the coordinates along the positive horizontal axis. Based on this function, the desired impedance model is redefined as follows:

Multiplying both sides of (43) by $w(s)$ yields

where $\bm{C}_{a}(t)\stackrel{{\scriptstyle\triangle}}{{=}}w(s)\bm{C}_{d}$ and $\bm{K}_{a}(t)\stackrel{{\scriptstyle\triangle}}{{=}}w(s)\bm{K}_{d}$, are the time-varying apparent impedance parameters. These parameters are utilized to deduce the relationship between the desired trajectory and the interaction torque, as outlined in (11). The function of the mechanism is such that an increase in the anomaly score leads to a decrease in $w(s)$, thereby reducing the magnitude of the impedance parameters. This reduction ensures that there is an increase in the passiveness with which the upper-limb exoskeleton robot responds to any detected conflicts. Conversely, when the anomaly score is low, the impedance parameters revert to their original values, thereby maintaining the level of assistance provided by the upper-limb exoskeleton robot.

Subsequently, an impedance vector is introduced, as follows

where

is a reference vector. According to (45), the convergence of $\bm{z}\rightarrow\bm{0}$ implies the realization of the desired impedance model (43).

The overall control input is designed as defined in (3), with the fast time-scale control term defined as in (4). Next, the slow time-scale control term is established by using the estimated friction $\hat{\bm{\tau}}_{f}$ to stabilize the dynamics expressed in (9) and achieve the desired impedance model, as follows:

where ${\bm{sgn}}(\cdot)$ is the sign function and is defined as follows:

where $k_{g}$ is a positive constant, and $\bm{K}_{z}\in\Re^{n\times n}$ is a diagonal and positive-definite matrix.

The proposed variable impedance controller, as delineated in (3), (4), and (47), can be demonstrated to be exponentially stable, as shown by the stability analysis provided in the Appendix.

The proposed dual-mode trajectory refinement method was implemented within the upper-limb exoskeleton robot to assess the effectiveness of both rehabilitation modes. Figure 8 depicts the experimental configuration, in which the main board controlled the impedance of the upper-limb exoskeleton robot. This board was interfaced via a serial port connection with a PC, which was outfitted with an Intel i5-13490F CPU and an RTX 4060Ti graphics card. The PC executed the anomaly detection module, calculated the reference trajectory $\bm{q}_{r}$, and dynamically updated the desired trajectory $\bm{q}_{d}$ online. Subjects performed rehabilitation training in passive following mode or active mirroring mode. In the latter mode, the subject wore a brace equipped with markers on the unaffected side of the body to facilitate engagement in rehabilitation exercises. Optical motion capture equipment (Nokov) was used to accurately capture motion intentions.

Implementation of the variable impedance controller required knowledge of the dynamic model. Thus, the dynamic parameters were computed analytically in real time using the open-source Orocos Kinematics and Dynamics Library¹¹1https://www.orocos.org/wiki/orocos/kdl-wiki.html. In this computation, the upper-limb exoskeleton robot is deconstructed into a sequence of links and joints to formulate a model defining its physical configuration, encompassing characteristics such as the length, mass, and inertia of each link. Subsequently, the forward dynamics are derived using Newton–Euler equations based on this model.

The intention predictor and the anomaly detection module required motion data, which was collected from the upper-limb exoskeleton robot during operation by healthy subjects. To facilitate free and natural movement, the upper-limb exoskeleton robot was set to operate in a transparent mode during the data collection phase (Chen et al., 2023a). Specifically, the controller was designed as follows (Zimmermann et al., 2020):

where $\gamma_{0}$ represents a parameter controlling the magnitude of virtual mass, and $\ddot{\bm{q}}_{0}$ denotes the desired acceleration. In this transparent mode, healthy subjects or patients were able to maneuver the upper-limb exoskeleton robot effortlessly, without experiencing significant discomfort.

The next section presents the results of five experiments, which are described below.

1. -

   Intention Predictor: This experiment aimed to validate the efficacy and accuracy of the proposed intention predictor. A diverse set of intention prediction methods were trained using the collected data, and then a comparative analysis was performed to demonstrate the superiority of the proposed intention predictor.

2. -

   Anomaly Detection: This experiment aimed to evaluate the performance of the anomaly detection network. Its detection accuracy was demonstrated in various simulated anomaly scenarios, including movements outside the normal range, stroke-induced convulsions, and human–robot interaction conflicts. Furthermore, comparative studies were conducted to illustrate that the proposed anomaly detector, which is based on a diffusion model, exhibits detection accuracy that is significantly better than that exhibited by other anomaly detection methods.

3. -

   Interaction Control: This experiment aimed to assess the dynamic capabilities of the system and its ability to reject disturbances. Thus, a trajectory tracking task was conducted. In addition, the efficacy of the proposed variable impedance controller was validated in a scenario involving anomalies.

4. -

   Active Mirroring Training: This experiment aimed to assess the effectiveness of the online trajectory refinement. Thus, ablation studies were conducted. In addition, how the anomaly score influences the trajectories generated by this refinement process was examined. Furthermore, an active mirroring mode was implemented, and the motion capture system was used to verify that the proposed method effectively constrains assistive trajectories and maintains safety throughout a rehabilitation process.

5. -

   Passive Following Training: This experiment aimed to validate the improvements in rehabilitation facilitated by passive following training. An ablation study was conducted to demonstrate that the online trajectory refinement significantly enhances movement naturalness and task performance throughout the training process. Moreover, a clinical trial was performed with stroke patients to obtain evidence that this rehabilitation framework significantly aids in the recovery of motor functions.

Two able-bodied participants with no prior experience with upper-limb exoskeleton robots were recruited for motion data collection. Table 3 presents the motion range of the upper-limb exoskeleton robot during data collection. This motion range covers the main spatial areas of upper-limb daily activities.

Table 4 presents an overview of the participants’ statistical attributes. The participants signed a written informed consent form prior to the experimental sessions. Next, the participants underwent a 6-minute pre-training phase to familiarize themselves with moving while wearing the upper-limb exoskeleton robot. During the data collection phase, no specific movement guidelines were imposed. Therefore, the participants were permitted to move the upper-limb exoskeleton robot at their preferred speed, i.e., rapidly or slowly, or to maintain it in a stationary position. The collected data were compiled into a dataset that was utilized for training the intention predictor and anomaly detection modules.