Upper limb motor dysfunction caused by stroke, traumatic brain injury, and other neurological disorders has become a major global public health challenge, imposing heavy burdens on both healthcare systems and patient families. Conventional manual rehabilitation therapy is limited by insufficient therapist resources, inconsistent treatment quality, and a lack of quantitative evaluation, while most existing rehabilitation robots suffer from low terminal accuracy, poor adaptability to individual patient status, and low long-term training compliance. This paper presents a hybrid parallel-serial upper limb rehabilitation robot integrating multi-modal force-tactile perception, intelligent adaptive trajectory control, and immersive augmented reality (AR) interaction. First, a multi-body dynamic model of the hybrid mechanism is established based on the Lagrangian method, and a domestically manufactured hardware platform is developed with 100% localization rate of core components, achieving 0.1 mm terminal repeat positioning accuracy. Second, a multi-modal sensing fusion framework combining surface electromyography (sEMG), 3D force signals, and kinematic data is constructed; an optimized BiLSTM-Attention model with reinforcement learning is proposed to predict patient motion intention 300 ms in advance, and dynamically adjust auxiliary torque and motion trajectory with a control response delay ≤ 30 ms. Third, a lightweight asynchronous AR training system is developed to realize haptic-visual synchronous feedback, with interaction latency controlled within 50 ms. Bench tests and preliminary clinical validation show that the proposed robot outperforms industry average levels in core performance indicators, increases patients’ active motion induction rate by 30%, and shortens the Brunnstrom stage progression cycle by approximately 40% compared with traditional manual training. This system provides a precise, intelligent, and patient-friendly solution for upper limb motor rehabilitation, with broad clinical application prospects.
With the aging of the global population and the rising incidence of cerebrovascular diseases, the number of patients with upper limb motor dysfunction continues to grow rapidly. Statistics show that 70% - 80% of stroke survivors suffer from upper limb motor impairment, which severely reduces their ability to perform activities of daily living (ADL) [
In recent years, upper limb rehabilitation robots have become a research hotspot in the field of rehabilitation engineering. Existing products can be divided into two categories: imported high-end devices represented by MIT-MANUS and InMotion Arm, which have stable performance but extremely high cost and low penetration in primary medical institutions; and domestic devices, which generally have shortcomings such as insufficient terminal control accuracy, weak adaptive adjustment ability, and a single interaction mode. Most existing robots follow fixed preset trajectories, cannot dynamically adjust training strategies based on patients’ real-time muscle status and motor intentions, and lack immersive interactive experiences, resulting in low patient compliance and limited rehabilitation effects [
To address the above problems, this paper develops an AR-enhanced upper limb rehabilitation robot with adaptive trajectory control capability. The main contributions are as follows:
1) A hybrid parallel-serial mechanical structure is designed, which combines the advantages of the large working space of serial mechanisms and high rigidity and precision of parallel mechanisms, and realizes 100% localization of core components.
2) A multi-modal sensing fusion adaptive control strategy is proposed, which integrates sEMG, force, and kinematic signals to realize real-time motion intention recognition and dynamic trajectory adjustment.
3) An immersive AR interactive training system with a lightweight asynchronous architecture is built to realize haptic-visual synchronous feedback and improve patient training compliance.
4) Systematic bench tests and preliminary clinical verification are carried out to prove the advanced performance and clinical application value of the system.
Addressing the contradiction between the working space and terminal accuracy of traditional rehabilitation robots [
The mechanism has 4 degrees of freedom, encompassing the primary motion modes required for upper-limb rehabilitation: shoulder flexion/extension, shoulder abduction/adduction, elbow flexion/extension, and forearm pronation/supination. The modular design of the end effector enables quick replacement of different grips to accommodate patients with varying hand function levels. The overall structure, key components, workspace simulation, and clinical application form of the developed rehabilitation robot are shown in
Based on the Denavit-Hartenberg (D-H) parameter method, the hybrid parallel-serial mechanism is equivalent to a unified serial system, and the inverse kinematics solution model of the robot is established to accurately calculate the drive position of each joint corresponding to the target pose of the end effector. The core inverse kinematics algorithm adopted in this study has been authorized as a national invention patent [
For the nonlinear, complex, closed-chain characteristics of the parallel mechanism [
To break the dependence on imported core components and reduce the overall cost of the machine, all core components, such as servo motors, precision reducers, and multi-dimensional force sensors, are sourced from domestically produced, mature industrial-grade devices, and the localization rate of core functional components reaches 100%. Through mechanism topology optimization and component
adaptation calibration, the terminal repeat positioning accuracy of 0.1 mm is achieved on the premise of ensuring performance, and the weight of the whole machine is much lower than the traditional 50 kg rehabilitation equipment, which is convenient for deployment in multiple scenarios.
The control system adopts a closed-loop data flow architecture of “end sensor acquisition → data preprocessing → main control unit operation → power execution → attitude feedback calibration”. The end 3D force sensor, surface electromyography electrode, and inertial measurement unit synchronously collect multi-source data, and the control board calculates the control quantity in real time and drives the actuator to move, forming a complete closed-loop control.
Aiming at the problems of nonlinear characteristics, multi-source heterogeneity, and noise interference of multi-modal sensor data, a data fusion method combining an extended Kalman filter (EKF) and a convolutional neural network (CNN) is proposed. The EKF algorithm performs optimal state estimation on multi-source dynamic signals through a closed prediction-update iteration cycle, effectively suppressing random noise in sEMG, 3D force, and kinematic signals, and outputs smoothed state quantities with minimum variance [
the latent nonlinear features of different modal data, and a nonlinear fusion strategy is adopted to realize the deep interaction of cross-modal features, which provides high signal-to-noise ratio data input for subsequent motion intention recognition and trajectory prediction.
To realize an accurate prediction of the robot end-effector motion trajectory in nonlinear and high-dynamic rehabilitation scenarios, a bidirectional long short-term memory (BiLSTM) network with an attention mechanism is constructed in this study. As shown in
Aiming at the time-series characteristics of rehabilitation motion data, a bidirectional long short-term memory (BiLSTM) model with an attention mechanism is constructed to predict the robot’s end motion trajectory over the next 300 ms. The trajectory prediction method based on multi-modal rehabilitation data has been protected by an authorized invention patent [
On this basis, a reinforcement learning framework is introduced to build an adaptive control strategy. The system can adjust the auxiliary torque and motion trajectory in real time according to the patient’s posture change, external force interference, and sudden spasm, so as to avoid secondary injury caused by “one-size-fits-all” training.
An impedance control model based on force-position hybrid control is established, and the stiffness coefficient K of the robot is adjusted according to the patient’s muscle strength level:
F a s s i s t = K ( θ t a r g e t − θ a c t u a l ) + B θ ˙ (1)
where F a s s i s t is the equivalent auxiliary force at the end effector, θ t a r g e t and θ a c t u a l are the target joint angle and actual joint angle, respectively, θ ˙ denotes the angular velocity of the joint, and B is the damping coefficient. The stiffness coefficient K is adaptively adjusted in real time based on the patient’s sEMG-derived muscle strength level: it is set to a large value in the early rehabilitation stage to provide sufficient trajectory guidance and support force, and decreases gradually as motor function recovers [
Aiming at the problem of boring training process and low patient compliance of traditional rehabilitation robots [
The system supports multi-terminal adaptation, including desktop computers, tablets, and AR head-mounted displays. Through motion synchronous rendering and a haptic-visual linkage mechanism, it realizes real-time mapping between the robot end-effector motion and virtual scene interactions, providing patients with an immersive integrated rehabilitation experience. Representative AR training scenarios of the system are shown in
A multi-person collaborative scene editing system is developed, which allows doctors and therapists to design or modify AR training tasks collaboratively through the cloud, and customize trajectory constraints, resistance parameters, and task objectives. The system is pre-installed with more than 10 types of standardized virtual training scenarios covering upper limb fine function training, such as transfer, guidance, positioning, and grasping.
The system can automatically switch the scene complexity according to the patient’s Brunnstrom stage: low-difficulty tasks are adopted for early passive training, and multi-joint coordination challenges are upgraded for later active training, so as to realize the precise matching of rehabilitation schemes in different stages.
A single-threaded asynchronous rendering architecture is adopted to realize parallel processing of multi-sensor data on low-computing terminals such as tablets and AR glasses, which reduces memory occupation by 70% and controls the interaction feedback delay within 50 ms. The lightweight asynchronous rendering method for rehabilitation scenarios involved in this system has been authorized as an invention patent [
A haptic-visual synchronization mechanism is established: combined with sEMG signals and end force control data [
To verify the performance of the developed rehabilitation robot, systematic bench tests were carried out. All bench tests were conducted in a standard temperature and humidity laboratory environment (23˚C, 50% RH), and performed in strict accordance with the YY 9706 (aligned with IEC 60601) series standards for medical electrical equipment [
The preliminary clinical pilot trial was reviewed and approved by the Ethics Committee of Hangzhou First People’s Hospital (Approval No. [
Inclusion criteria were defined as: first-onset stroke confirmed by cranial CT/MRI, age 40 - 75 years, Brunnstrom stage II - V of the affected upper limb, stable vital signs, and ability to cooperate with rehabilitation training. Exclusion criteria included severe cognitive impairment, severe limb spasm (Modified Ashworth Scale ≥ grade 4), combined upper limb joint deformity, and other serious systemic diseases.
A self-controlled study design was adopted. A total of 20 eligible post-stroke patients with upper limb motor dysfunction were recruited, with baseline Brunnstrom stages covering II to V. All subjects had received conventional manual rehabilitation for 2 weeks before enrollment, and their baseline motor function and training efficiency data were collected as the control period. After enrollment, all patients received 4 weeks of robot-assisted rehabilitation training (30 min per session, 5 sessions per week) on the basis of conventional rehabilitation therapy. The primary evaluation indicators included the Fugl-Meyer Assessment for Upper Extremity (FMA-UE) score, modified Barthel Index (MBI), active motion induction rate, daily effective training duration, and incidence of adverse events [
Bench test results show that the terminal repeat positioning accuracy of the robot reaches 0.08 mm ± 0.02 mm, and the steady-state trajectory tracking error is controlled at 0.1 mm, which is significantly better than the ±0.5 mm accuracy level of mainstream imported rehabilitation equipment [
The closed-loop control system achieves a dynamic response delay of 22 ms ± 4 ms, which is far better than the industry average level of >100 ms, and fully meets the real-time control requirements of rehabilitation training. The system supports four training modes: passive, assistive, active, and resistive, with a weight assistance accuracy error ≤ 0.15 kg, which can cover the rehabilitation needs of patients at different Brunnstrom stages [
Benefiting from the proposed lightweight asynchronous rendering architecture, the end-to-end delay of AR interactive feedback is controlled at 42 ms ± 6 ms, which is significantly lower than the >200 ms latency level of similar AR rehabilitation systems. The system is equipped with more than 10 built-in standardized virtual training scenarios covering upper limb fine motor functions such as transfer, positioning, and grasping, and supports multi-terminal adaptation of tablets and AR head-mounted devices.
The detailed comparison of core performance indicators between the proposed system and the industry average level is summarized in
| Index Category | Specific Parameters of This System | Industry Average Level |
|---|---|---|
| Hardware performance | 100% localization rate of core components; 0.1 mm terminal repeat positioning accuracy; 0.1 mm motion trajectory accuracy | 60% localization rate for domestic devices; ±0.5 mm terminal accuracy for imported devices |
| Control performance | System response delay ≤ 30 ms; supports 4 training modes (passive/assistive/active/resistive); weight assistance accuracy error ≤ 0.2 kg | Response delay > 100 ms; most devices only support 2 - 3 training modes |
| Interaction performance | AR interaction delay ≤ 50 ms; built-in 10+ virtual training scenarios; supports multi-terminal adaptation | Most devices have no AR interaction, or a delay > 200 ms, with few scenarios |
| Safety performance | Equipment noise ≤ 55 dB(A); integrated multi-safety protection (anti-pinch, emergency stop, overload) | Meets general medical electrical standards, with lower safety redundancy |
The prototype completed a 100,000-cycle reciprocating motion fatigue test under full-load operating conditions, with no mechanical failure, component wear, or control accuracy degradation observed during the test. The estimated mean time between failures (MTBF) is greater than 10,000 hours, which supports 7 × 24 hours of long-term continuous operation and meets the high-frequency use requirements of clinical rehabilitation scenarios.
In terms of electrical safety and electromagnetic compatibility, the product has passed a full-item inspection by a CMA-accredited third-party testing institution and fully complies with the YY 9706.1 series standards for electric shock protection, mechanical safety, electromagnetic compatibility, and environmental adaptability. The device integrates multiple graded safety protection mechanisms, including anti-pinch detection, a physical emergency stop button, and overload current limiting, which can trigger protection actions within 10 ms in abnormal conditions such as excessive traction force or sudden limb spasm, effectively preventing secondary injury to patients.
After 4 weeks of robot-assisted rehabilitation intervention, the Fugl-Meyer Assessment for Upper Extremity (FMA-UE) score of the enrolled patients increased from 28.4 points ± 6.3 points at baseline to 39.7 points ± 7.1 points, with a mean improvement rate of 39.8% (P < 0.01, paired t-test). The modified Barthel Index (MBI) increased from 52.5 points ± 8.6 points to 68.3 points ± 9.4 points (P < 0.05), indicating a significant improvement in patients’ activities of daily living.
Among the 20 subjects, 14 patients achieved at least one-stage improvement in the Brunnstrom stage, with an improvement rate of 70%. Compared with the functional progression rate observed during the pre-enrollment conventional manual rehabilitation phase, the average Brunnstrom stage progression cycle was shortened by approximately 40% under robot-assisted training.
Quantitative statistics show that under robot-assisted training, the active motion induction rate of patients reaches 72%, and the average daily effective training duration reaches 45 min. Both indicators are notably higher than the 42% induction rate and 25 min effective duration widely reported in conventional manual rehabilitation literature [
During the entire pilot trial, no adverse events such as muscle strain, joint pain, skin pressure injury, or secondary motor injury occurred in all subjects. The equipment operated stably throughout the trial period, with no safety failures such as out-of-control motion, abnormal force output, or system crash, which preliminarily verified the clinical safety and operational stability of the system.
This study proposes an AR-enhanced upper limb rehabilitation robot with adaptive trajectory control, which has made breakthroughs in mechanism design, intelligent control, and interactive experience. Compared with existing research, the innovation of this system is mainly reflected in three aspects:
1) The hybrid parallel-serial mechanism takes into account both working space and control accuracy, and realizes full localization of core components, which significantly reduces the cost of the whole machine and improves the accessibility of primary medical scenarios.
2) The multi-modal fusion adaptive control strategy realizes real-time perception and dynamic adjustment of the patient’s motor state, which solves the problem of poor adaptability of traditional fixed trajectory training.
3) The lightweight asynchronous AR interaction system realizes haptic-visual synchronous feedback, which effectively improves patient compliance and neural plasticity activation efficiency.
This study still has some limitations. First, the sample sizes of preliminary clinical trials are small, and large-sample randomized controlled trials are needed to further verify clinical efficacy. Second, the current evaluation system focuses on motor function, and the evaluation of cognitive function and psychological state needs further enrichment. In future work, we will continue to promote clinical trials, optimize the multi-modal rehabilitation evaluation model, and explore the integration of brain-computer interface technology to further improve the intelligence level of the system.
This paper presents the design, implementation, and experimental validation of an AR-enhanced upper limb rehabilitation robot with adaptive trajectory control. The system adopts a hybrid parallel-serial mechanical structure, builds a multimodal sensing adaptive control system, and develops an immersive AR interactive training module, effectively addressing the pain points of traditional rehabilitation robots, such as low accuracy, poor adaptability, and low compliance. Bench tests and preliminary clinical verification prove that the system has advanced core performance and good rehabilitation effect, and has broad application prospects in clinical rehabilitation, community rehabilitation, and home rehabilitation.
In the next step, we will carry out large-scale multicenter clinical trials to further verify the clinical efficacy and safety of the system; promote the registration of Class II medical devices to accelerate the industrial transformation of the achievements; and continue to iterate the intelligent algorithm to build a more complete closed-loop rehabilitation system of “evaluation-training-feedback-optimization”.
Weiwei Wen: Writing―original draft; formal analysis (clinical data statistics and analysis). Hanyang Xu: Methodology; development and implementation of the adaptive trajectory control system. Fanghui Qiu: Validation; design of the clinical validation protocol. Tianxiao Chen: Investigation; clinical data collection and curation. Yu Wang: Conceptualization; funding acquisition; Writing―review & editing; overall study supervision.
The authors sincerely thank all volunteers who participated in the prototype performance testing and the preliminary clinical pilot trial of the rehabilitation robot. We highly appreciate their dedicated time, active cooperation, and valuable feedback, which have made essential contributions to the system performance verification and functional optimization of this study.
This research was financially supported by the Key Research and Development Program of Jiangxi Province, China (Grant No. 20252BCE310013).
The authors declare no conflicts of interest regarding the publication of this paper.