Post-stroke lower limb rehabilitation: a comparative study between exoskeleton robots and traditional gait training Authors: Wenzhu Xu, Nai Yeen Gavin Lai, Kok-Hoong Wong, Hamza Azam, Wei Yang Published: 2026 ABSTRACT PurposeThis study systematically compares the efficacy of robotic-assisted gait training (RAGT) with exoskeletons to traditional gait training (TGT) for lower limb rehabilitation after stroke, focusing on motor recovery, neuroplasticity, and muscle function.MethodsA comprehensive systematic review was conducted across PubMed, Scopus, Web of Science, IEEE Xplore and Google Scholar for studies published between 2004 and 2024 following the PRISMA guidelines. A total of 86 studies were selected based on strict inclusion criteria, covering moderate to severe stroke patients undergoing RAGT, TGT, or combined interventions. Outcome metrics included gait symmetry, walking speed, neuroplasticity biomarkers (e.g., BDNF, fMRI), and muscle strength indicators. This systematic review was pre-registered with PROSPERO (ID: CRD420261333541).ResultsRAGT demonstrated superior improvements in gait symmetry index (+ 99.8% trajectory accuracy), walking speed (+ 20%), and muscle strength via high-intensity, repeatable training. TGT excelled in promoting active participation, cortical engagement, and functional independence, particularly in activities of daily living. Neuroplasticity analysis revealed RAGT enhanced spinal-brainstem rhythmic-ity, while TGT reinforced cortical-cerebellar motor planning. Muscle recovery was accelerated by RAGT through anti-gravitational support, though TGT better preserved natural movement patterns and coordination.ConclusionRAGT and TGT exhibit complementary strengths in stroke rehabilitation. Integrating RAGT’s high-intensity, objective training with TGT’s adaptive, patient-centered approach represents a promising direction for personalized, AI-enhanced neurorehabilitation strategies. Introduction Stroke is a significant public health concern and one of the leading causes of disability and mortality worldwide, ranking as the second leading cause of death globally after cardiovascular diseases [1–4]. According to the World Health Organization (WHO), approximately 15 million individuals suffer from strokes annually, with about 5 million resulting in permanent disability [5]. Projections indicate that by 2030, the global prevalence of stroke will rise to 21.9% [6]. The majority of stroke survivors experience gait asymmetry, characterized by stiffness in the affected (hemiparetic) lower limb, impaired unilateral stance, and compromised weight-shifting abilities, which limit their mobility [7]. Although 65% to 85% of individuals can achieve independent walking within six months post-stroke through appropriate surgical, pharmacological, and rehabilitative interventions, suboptimal gait patterns and reduced cardiopulmonary endurance continue to hinder their daily ambulation [5, 8]. Therefore, effective rehabilitation of lower limb function is a crucial component of stroke recovery. Currently, various approaches are employed to improve gait in stroke patients, including Traditional Gait Training (TGT) and Robotic-Assisted Gait Training (RAGT). TGT encompasses physical therapy (PT), occupational therapy (OT), and gait training (GT) [9], and has long been the cornerstone of stroke rehabilitation. PT employs methods such as the Brunnstrom approach [10], Proprioceptive Neuromuscular Facilitation (PNF) techniques [11], Range of Motion (ROM) exercises, Strength Training, Balance Training, and Functional Electrical Stimulation (FES) to enhance lower limb function, strength, and coordination. OT focuses on daily activity-oriented interventions [12] to promote functional independence and improve quality of life [13], making it suitable for comprehensive rehabilitation strategies [14]. GT [15] involves repetitive gait exercises to enhance walking ability [16] and balance coordination [17], facilitating neuroplastic recovery [18], though it demands significant physical effort from patients [19]. Clinical guidelines recommend at least 45 min of therapy daily; however, there remains a substantial gap between the recommended and actual therapy durations [20]. TGT also has certain limitations, such as variability in therapist skills, patient fatigue, and the high demand for therapist-to-patient ratios [21]. RAGT involves the use of wearable rehabilitation robotic devices that assist limb movements, enabling consistent high-intensity training, objective progress assessment, and alleviating the physical burden on therapists [22]. RAGT is widely applied in lower limb rehabilitation for stroke patients [23] and is mainly categorized into mobile exoskeletons (e.g., ReWalk [24], Ekso [25, 26]) and stationary exoskeletons (e.g., Lokomat [27–53], G-EO System [39, 54–56], ALEX [57]) [58–60]. Mobile exoskeletons are fully wearable, providing walking assistance [24], facilitating outdoor and daily activity training [23], and enhancing functional independence [61]; however, they have limitations in supporting multi-planar movements [62] and comprehensive rehabilitation [63]. Stationary exoskeletons [62, 64], typically integrated with treadmills and suspension systems [65], offer high-intensity repetitive training [35, 66], suitable for early rehabilitation stages and patients with severe motor impairments [67], but are confined to indoor training environments, limiting their translation to real-life mobility [68]. Despite growing interest in RAGT and TGT over the past two decades(Fig. 1), direct comparative studies evaluating their combined efficacy remain limited. While early investigations (2015–2019) laid foundational insights, recent literature (2020–2024) has predominantly focused on isolated outcomes (e.g., gait speed or muscle strength) without addressing critical gaps: (1) studies vary in intervention duration (4–12 weeks), exoskeleton types (mobile vs. fixed), and patient cohorts (acute vs. chronic stages), complicating cross-study synthesis; (2) over 80% of trials lack more than 6-month follow-up, obscuring long-term functional maintenance; and (3) few studies integrate neuroplasticity biomarkers (e.g., fMRI connectivity, BDNF levels) with biomechanical outcomes, limiting translational insights. In contrast, this review adopts a multidimensional framework to systematically compare RAGT and TGT across motor recovery, neuroplasticity, and muscle adaptation. By synthesizing evidence from 86 rigorously screened studies, we: (1) decode hierarchical interactions between mechanical assistance and neural control, (2) propose a staged rehabilitation strategy (acute-phase TGT followed by subacute RAGT) to optimize recovery trajectories, and (3) identify design imperatives for next-generation exoskeletons, such as adaptive impedance control and multi-planar actuation. These contributions address the field’s fragmented understanding and provide actionable guidelines for personalized, AI-enhanced rehabilitation paradigms. Fig. 1Trends in the Number of Publications on RAGT and TGT Keywords Over the Past Two Decades. a Annual publication counts related to RAGT, TGT, and combined RAGT–TGT keywords from 2004 to 2024, showing the rapid growth of research in gait rehabilitation technologies. b Annual publication counts for physical therapy, robot-assisted rehabilitation, human-robot interaction, and personalized control, highlighting the expanding scope and interdisciplinary integration within rehabilitation robotics Trends in the Number of Publications on RAGT and TGT Keywords Over the Past Two Decades. a Annual publication counts related to RAGT, TGT, and combined RAGT–TGT keywords from 2004 to 2024, showing the rapid growth of research in gait rehabilitation technologies. b Annual publication counts for physical therapy, robot-assisted rehabilitation, human-robot interaction, and personalized control, highlighting the expanding scope and interdisciplinary integration within rehabilitation robotics Research methods This study was conducted in accordance with the PRISMA guidelines [69], the preferred reporting item for systematic reviews. The primary objective of this study was to evaluate the effectiveness of RAGT versus TGT in post-stroke rehabilitation, focusing on three outcome domains: gait recovery, neuroplasticity, and muscle strength. To ensure methodological transparency and minimize bias, the review protocol was prospectively registered on PROSPERO (Registration ID: CRD420261333541). Literature search strategy A comprehensive and systematic literature search was conducted across five major electronic databases: PubMed, Scopus, Web of Science, IEEE Xplore, and Google Scholar. The search was restricted to publications within the period from January 2004 to December 2024, a timeframe selected to capture the full trajectory of development in robotic rehabilitation technology while ensuring contemporary relevance. The decision to include IEEE Xplore reflects the interdisciplinary nature of RAGT research, which spans both clinical rehabilitation and biomedical engineering domains. The search strategy was systematically constructed using a combination of Medical Subject Headings (MeSH) terms and free-text keywords to maximize recall and minimize selection bias. The core search terms included: “Stroke,” “Cerebrovascular Accident,” “Exoskeleton Device,” “Robotic-Assisted Gait Training,” “Robot-Assisted Rehabilitation,” and “Traditional Gait Training.” These terms were combined using Boolean operators (AND/OR) with outcome-related descriptors, including “Neuroplasticity,” and “Muscle Strength.” All database searches were conducted independently and cross-validated to ensure completeness. Search results were exported and managed using reference management software Endnote, with automatic deduplication applied as an initial step. Inclusion and exclusion criteria Study eligibility was determined according to a pre-specified framework based on the PICO [70] model, as detailed below: Inclusion Criteria: Population: Adult patients (≥ 18 years of age) with a confirmed clinical diagnosis of ischemic or hemorrhagic stroke, presenting with lower-limb motor deficits (e.g., hemiplegia or hemiparesis) that impair independent ambulation.Intervention: Studies evaluating RAGT implemented via powered exoskeletal devices or end-effector robotic systems, either as a standalone intervention or as part of a combined rehabilitation protocol.Comparison: Studies including an active comparator condition of TGT or conventional physiotherapy-based gait rehabilitation.Outcomes: Studies reporting at least one objective, quantifiable measure of gait function (e.g., gait speed, step length, cadence, symmetry), balance capacity (e.g., Berg Balance Scale [BBS], Timed Up and Go [TUG] test), muscle activation (e.g., surface electromyography [sEMG] indices), or neuroplasticity (e.g., functional magnetic resonance imaging [fMRI] activation maps, serum BDNF levels). Population: Adult patients (≥ 18 years of age) with a confirmed clinical diagnosis of ischemic or hemorrhagic stroke, presenting with lower-limb motor deficits (e.g., hemiplegia or hemiparesis) that impair independent ambulation. Population: Adult patients (≥ 18 years of age) with a confirmed clinical diagnosis of ischemic or hemorrhagic stroke, presenting with lower-limb motor deficits (e.g., hemiplegia or hemiparesis) that impair independent ambulation. Intervention: Studies evaluating RAGT implemented via powered exoskeletal devices or end-effector robotic systems, either as a standalone intervention or as part of a combined rehabilitation protocol. Intervention: Studies evaluating RAGT implemented via powered exoskeletal devices or end-effector robotic systems, either as a standalone intervention or as part of a combined rehabilitation protocol. Comparison: Studies including an active comparator condition of TGT or conventional physiotherapy-based gait rehabilitation. Comparison: Studies including an active comparator condition of TGT or conventional physiotherapy-based gait rehabilitation. Outcomes: Studies reporting at least one objective, quantifiable measure of gait function (e.g., gait speed, step length, cadence, symmetry), balance capacity (e.g., Berg Balance Scale [BBS], Timed Up and Go [TUG] test), muscle activation (e.g., surface electromyography [sEMG] indices), or neuroplasticity (e.g., functional magnetic resonance imaging [fMRI] activation maps, serum BDNF levels). Outcomes: Studies reporting at least one objective, quantifiable measure of gait function (e.g., gait speed, step length, cadence, symmetry), balance capacity (e.g., Berg Balance Scale [BBS], Timed Up and Go [TUG] test), muscle activation (e.g., surface electromyography [sEMG] indices), or neuroplasticity (e.g., functional magnetic resonance imaging [fMRI] activation maps, serum BDNF levels). Exclusion Criteria: Duplicate records identified across multiple databases, with only the most complete version retained.Studies published outside the designated timeframe of 2004–2024.Non-original research formats, including systematic reviews, narrative reviews, meta-analyses, case reports, editorials, commentaries, conference abstracts, and study protocol papers.Studies conducted on pediatric populations (< 18 years) or on patient populations with neurological conditions other than stroke as the primary diagnosis.Studies published in languages other than English, for which adequate translation could not be verified. Duplicate records identified across multiple databases, with only the most complete version retained. Duplicate records identified across multiple databases, with only the most complete version retained. Studies published outside the designated timeframe of 2004–2024. Studies published outside the designated timeframe of 2004–2024. Non-original research formats, including systematic reviews, narrative reviews, meta-analyses, case reports, editorials, commentaries, conference abstracts, and study protocol papers. Non-original research formats, including systematic reviews, narrative reviews, meta-analyses, case reports, editorials, commentaries, conference abstracts, and study protocol papers. Studies conducted on pediatric populations (< 18 years) or on patient populations with neurological conditions other than stroke as the primary diagnosis. Studies conducted on pediatric populations (< 18 years) or on patient populations with neurological conditions other than stroke as the primary diagnosis. Studies published in languages other than English, for which adequate translation could not be verified. Studies published in languages other than English, for which adequate translation could not be verified. Study selection and data extraction The study selection process was executed in two sequential, independent stages in accordance with PRISMA recommendations. In the first stage, two researchers (W.X. and W.Y.) independently screened the titles and abstracts of all records identified through the database searches against the predefined inclusion and exclusion criteria. Records that were clearly irrelevant, duplicated, or outside the scope of the review were excluded at this stage. In the second stage, the full texts of all potentially eligible studies were independently retrieved and comprehensively reviewed by the same two researchers to confirm final eligibility. Any discrepancy in eligibility decisions between the two reviewers was initially addressed through structured bilateral discussion. In cases where consensus could not be achieved, a third senior arbitrator (L.N.Y.G.) was consulted to provide a definitive adjudication. Data extraction was performed independently by the same two reviewers (W.X. and W.Y.) using a standardized, pre-piloted data collection form developed specifically for this review. The extracted information encompassed the following domains: (1) study identification details (authors, year of publication, country, and journal); (2) participant characteristics, including sample size, mean age, sex distribution, stroke type (ischemic vs. hemorrhagic), stroke chronicity (acute, subacute, or chronic phase), and baseline functional status; (3) intervention parameters, including the type and model of robotic device, training intensity (session frequency, session duration, and total training weeks), and the nature of any combined interventions; (4) comparator intervention details; and (5) primary and secondary outcome measures, including the specific instruments used and the time points of assessment. The complete study selection workflow, including the number of records at each decision point and the reasons for full-text exclusion, is illustrated in the PRISMA flow diagram presented in Fig. 2. Fig. 2PRISMA flowchart of the search and inclusion process PRISMA flowchart of the search and inclusion process Comparative analysis of rehabilitation methods Although stroke recovery encompasses a wide range of care, including in-hospital, outpatient, and community-based rehabilitation, this section focuses on the comparative analysis of RAGT versus TGT for motor function recovery, neuroplasticity, and muscle function recovery. To facilitate multidimensional comparison between RAGT and TGT across heterogeneous outcome domains that cannot be statistically pooled, a semi-quantitative evidence synthesis approach was developed. Each rehabilitation outcome indicator (e.g., gait symmetry, muscle activation, cortical engagement) was assigned a composite score from 0 to 10 based on a structured, evidence-based scoring algorithm integrating three dimensions: (A) Clinical Significance and Magnitude of Evidence (0–4 points): Evaluates whether studies consistently demonstrate benefit and whether the effect magnitude exceeds the Minimal Clinically Important Difference (MCID). Higher scores reflect stronger evidence of clinically meaningful improvement.(B) Methodological Quality (0–3 points): Integrates PEDro-based quality assessment of supporting studies. Higher scores are assigned when evidence derives from studies with PEDro scores ≥ 6 (low risk of bias).(C) Consistency and Replicability (0–3 points): Assesses the degree of agreement across independent studies. Higher scores are assigned when ≥ 4 independent studies report convergent findings. (A) Clinical Significance and Magnitude of Evidence (0–4 points): Evaluates whether studies consistently demonstrate benefit and whether the effect magnitude exceeds the Minimal Clinically Important Difference (MCID). Higher scores reflect stronger evidence of clinically meaningful improvement. (A) Clinical Significance and Magnitude of Evidence (0–4 points): Evaluates whether studies consistently demonstrate benefit and whether the effect magnitude exceeds the Minimal Clinically Important Difference (MCID). Higher scores reflect stronger evidence of clinically meaningful improvement. (B) Methodological Quality (0–3 points): Integrates PEDro-based quality assessment of supporting studies. Higher scores are assigned when evidence derives from studies with PEDro scores ≥ 6 (low risk of bias). (B) Methodological Quality (0–3 points): Integrates PEDro-based quality assessment of supporting studies. Higher scores are assigned when evidence derives from studies with PEDro scores ≥ 6 (low risk of bias). (C) Consistency and Replicability (0–3 points): Assesses the degree of agreement across independent studies. Higher scores are assigned when ≥ 4 independent studies report convergent findings. (C) Consistency and Replicability (0–3 points): Assesses the degree of agreement across independent studies. Higher scores are assigned when ≥ 4 independent studies report convergent findings. The composite score for each indicator was calculated as: Score = A + B + C (range: 0–10). For negative indicators (e.g., mechanical dependency), directionality was inverted so higher scores consistently represent more favorable rehabilitation outcomes. Scoring was performed independently by two researchers (WX and WY). Discrepancies exceeding 2 points were resolved through consensus discussion, with a third senior researcher (NYGL) consulted when necessary. Results are visualized using radar charts (Figs. 3, 4 and 5) to provide a “top-down” perspective of the rehabilitation profile across multiple domains simultaneously, allowing clinicians to visualize trade-offs between mechanical assistance (RAGT) and active neural engagement (TGT). Motor function recovery: RAGT mechanized assistance versus active neural engagement of TGT The differences between RAGT and TGT in the recovery of motor functions reflect complementary paradigms. RAGT emphasizes a programmable adaptive training protocol that can modulate intensity (from mild to high) based on real-time physiological feedback (e.g., EMG signals or heart rate variability), while TGT prioritizes therapist-guided, context-sensitive adjustments. While early RAGT systems focused on high-intensity repetitive training for severe motor deficits, modern exoskeletons integrate sensor-driven algorithms to dynamically adjust assistance levels, enabling graded challenge progression for patients across recovery stages. For instance, subacute stroke patients may benefit from low-intensity, error-amplification modes to enhance cortical engagement, while chronic patients tolerate high-intensity patterns to optimize neuromuscular remodeling. This adaptability addresses a critical limitation of traditional TGT, where therapist-dependent manual adjustments may struggle to maintain precision in intensity titration. In-depth analysis reveals that these differences extend beyond apparent efficacy to fundamental distinctions in motor learning mechanisms.The 13 angle comparison of motor-function metric changes between RAGT and TGT is given in Fig. 3.The scoring criteria utilize a 0–10 scale, where higher values indicate better performance. Additionally, a directionality adjustment is applied for negative indicators, such that higher scores represent lower dependency. RAGT’s closed-loop systems leverage mechanical assistance to reduce compensatory strategies in early rehabilitation, while TGT’s emphasis on active participation strengthens volitional control in later stages. Thus, the optimal integration of both approaches lies in staged intensity modulation, ensuring patients receive task-specific challenges aligned with their neurophysiological readiness. Fig. 3Changes in Various Metrics of Motor Function Between RAGT and TGT(1) Mechanical Dependency Resistance [71–73] (2) Improvement in Motor Variability [74] (3) Autonomous Control [75] (4) Active Muscle Load [71, 73, 76, 77] (5) Personalization [71–73] (6) Range of Motion (ROM) [78, 79] (7) Gait Activation [71, 73, 76] (8) Feedback [71–73, 80] (9) Increase in Walking Speed [23, 26, 28–33, 35–44, 46, 48, 52, 55, 56, 73, 77, 81–116] (10) Improvement in Walking Endurance [73, 77] (11) Reproducibility [71–73, 117](12) GSI [72, 73, 118, 119] (13) Precision of Motor Control [120] Changes in Various Metrics of Motor Function Between RAGT and TGT(1) Mechanical Dependency Resistance [71–73] (2) Improvement in Motor Variability [74] (3) Autonomous Control [75] (4) Active Muscle Load [71, 73, 76, 77] (5) Personalization [71–73] (6) Range of Motion (ROM) [78, 79] (7) Gait Activation [71, 73, 76] (8) Feedback [71–73, 80] (9) Increase in Walking Speed [23, 26, 28–33, 35–44, 46, 48, 52, 55, 56, 73, 77, 81–116] (10) Improvement in Walking Endurance [73, 77] (11) Reproducibility [71–73, 117](12) GSI [72, 73, 118, 119] (13) Precision of Motor Control [120] Table 1Comparison between RAGT and TGT in motor function recoveryEIRAGT effectsTGT effectsComparative conclusionGSI [72, 73, 118, 119]Significant improvement in GSI with high trajectory accuracy (99.8%)GSI improvement relies on therapist feedback, with considerable fluctuations in enhancementRAGT is superior in standardized training, but may suppress autonomous error correctionWS [23, 26, 28–33, 35–44, 46, 48, 52, 55, 56, 73, 77, 81–116]Significant short-term improvement (+ 20%), with mechanical assistance exceeding physiological limitsStable long-term improvement, with synergistic enhancement of cardiopulmonary function when combined with HIITRAGT is suitable for rapid recovery, while TGT enhances exercise motivation through active participationROM [50, 78, 79, 97, 109, 114, 121, 122]Rapid short-term improvement in ROM (13–42%)Superior dynamic adjustmentRAGT is suitable for early recovery, while TGT offers stronger support for neural adaptationMV [44, 74, 88]Suppresses movement variability, which may hinder autonomous controlEnhances cortical error processing through error amplification strategiesTGT has advantages in promoting autonomous motor learningTI [35, 66]High-intensity repetitive training, executable in a standardized mannerIntensity fluctuates due to limitations of therapist stamina and patient enduranceRAGT is more advantageous when resources are limitedEI = Evaluation Index, GSI = Gait Symmetry, WS = Walking Speed, ROM = Range of Motion, MV = Movement Variability, TI = Training Intensity Table 1Comparison between RAGT and TGT in motor function recoveryEIRAGT effectsTGT effectsComparative conclusionGSI [72, 73, 118, 119]Significant improvement in GSI with high trajectory accuracy (99.8%)GSI improvement relies on therapist feedback, with considerable fluctuations in enhancementRAGT is superior in standardized training, but may suppress autonomous error correctionWS [23, 26, 28–33, 35–44, 46, 48, 52, 55, 56, 73, 77, 81–116]Significant short-term improvement (+ 20%), with mechanical assistance exceeding physiological limitsStable long-term improvement, with synergistic enhancement of cardiopulmonary function when combined with HIITRAGT is suitable for rapid recovery, while TGT enhances exercise motivation through active participationROM [50, 78, 79, 97, 109, 114, 121, 122]Rapid short-term improvement in ROM (13–42%)Superior dynamic adjustmentRAGT is suitable for early recovery, while TGT offers stronger support for neural adaptationMV [44, 74, 88]Suppresses movement variability, which may hinder autonomous controlEnhances cortical error processing through error amplification strategiesTGT has advantages in promoting autonomous motor learningTI [35, 66]High-intensity repetitive training, executable in a standardized mannerIntensity fluctuates due to limitations of therapist stamina and patient enduranceRAGT is more advantageous when resources are limitedEI = Evaluation Index, GSI = Gait Symmetry, WS = Walking Speed, ROM = Range of Motion, MV = Movement Variability, TI = Training Intensity Comparison between RAGT and TGT in motor function recovery EI = Evaluation Index, GSI = Gait Symmetry, WS = Walking Speed, ROM = Range of Motion, MV = Movement Variability, TI = Training Intensity RAGT demonstrates superior performance over TGT in metrics such as gait symmetry, walking speed, and joint range of motion. However, its efficacy in enhancing motor variability is limited, suggesting that mechanical assistance may suppress the ability for autonomous motor adaptation.RAGT and TGT motor function recovery outcomes are compared in Table 1. In the recovery of joint ROM, RAGT employs torque compensation and fixed trajectory designs to effectively reduce patients’ active muscle load and expand movement range. Nevertheless, its mechanized assistance mode might inhibit patients’ active perception and correction of movement errors. For instance, a study by Major et al. reported similar ROM improvements (13%-42%) in both groups [79], implying that the ”compensatory advantage” of mechanical assistance could mask potential deficits in neural adaptability. Conversely, ROM improvements achieved through TGT interventions (e.g., shockwave therapy) [78] are fundamentally based on dynamic adjustments of peripheral neuromuscular interactions. Such interventions are more likely to activate patients’ proprioceptive feedback loops, thereby promoting autonomous optimization of motor control. Therefore, while standardized RAGT training can rapidly enhance joint mobility, its support for the adaptive mechanisms of motor learning requires further investigation. Improvements in walking endurance and speed reveal the dual role of rehabilitation intensity. High-intensity training with RAGT breaks through patients’ physical limits via mechanical assistance; for example, Louie’s study reported a 20% improvement in walking speed [23]. However, this ”passive high-intensity” approach may diminish patients’ cognitive engagement in motor tasks. In contrast, TGT combined with high-intensity interval training (HIIT) necessitates that patients actively regulate movement pace and energy distribution. This ”active high-intensity” approach not only enhances cardiopulmonary function but also sustains motor motivation through activation of the prefrontal-basal ganglia pathway [75]. A study by Mustafaoglu et al. demonstrated that the combined therapy group significantly outperformed single therapy in walking distance (78.2 m vs. 25 m) [123], suggesting that the synergistic effect of mechanical assistance and active training may stem from dual stimulation of motor control hierarchies. RAGT breaks through patients’ physiological limits (first threshold) via mechanical assistance, while TGT, through active participation, surpasses neural control limits (second threshold). The synergy of both can activate the dual ”spinal-cortical” motor learning pathways. The disparity in the precision of motor control reflects the core contradiction in human-machine interaction. RAGT’s closed-loop control system boasts a trajectory accuracy of 99.8% [120], achieving an ideal GSI. However, its rigid control strategy may suppress motor variability [74]. In traditional rehabilitation, therapists guide patients to adjust gait through real-time tactile feedback, essentially employing an ”error ampli-fication” strategy to enhance the motor cortex’s error signal processing capability [124]. Therefore, if RAGT’s precision advantage excludes tolerance for motor variability, it may hinder the transition from ”mechanical dependence” to ”autonomous control.” Future exoskeleton systems need to integrate adaptive impedance control algorithms to introduce controllable movement perturbations while ensuring safety, thereby simulating the trial-and-error process inherent in natural gait learning. Neuroplasticity: deep mechanisms from synaptic remodeling to whole-brain network reconstruction Research on neuroplasticity has expanded from merely examining changes in synaptic efficacy to analyzing the reorganization of whole-brain functional networks. The distinct induction pathways of RAGT and TGT in promoting neuroplasticity reflect inherent differences in their neural regulatory hierarchies, as illustrated in Fig. 4. At the microscopic level, the high-intensity repetitive training delivered by RAGT enhances long-term potentiation (LTP) through stable input patterns. Cataldo et al. demonstrated that the RAGT group exhibited a significant increase in brain-derived neurotrophic factor (BDNF) levels (+ 5.96 ng/ml) [125],as shown in Table 2. This upregulation of BDNF is likely attributable to the regular mechanical rhythmic stimulation of spinal central pattern generators, which indirectly enhances cortical excitability via the brainstem-thalamic pathway. However, this “bottom-up” stimulation mode may overly rely on exogenous rhythmic inputs, potentially leading to a diminished cortical representation of voluntary movement intention. In contrast, the active training provided by TGT, through the synergistic activation of the premotor cortex (PMC) and supplementary motor area (SMA), directly reinforces the synaptic connections within the motor intention-execution circuitry [126]. Moreover, Kim’s fNIRS study revealed that PMC activation increased by 173% during voluntary walking tasks in the TGT group [127], suggesting that active training more specifically strengthens the cortical-cerebellar motor planning network. This “top-down” regulatory approach is more consistent with the neural mechanisms underlying natural motor learning. Thus, elucidating the hierarchical differences between RAGT and TGT in modulating neuroplasticity requires an integrated strategy that combines both “bottom-up” (spinal rhythmicity) and “top-down” (cortical control) stimulation. Fig. 4Changes in various neuroplasticity indicators between RAGT and TGT. (1) SMA [126] (2) PMC [127] (3) Self-organization ability [127, 128] (4) Voluntary movement intention [129, 130] (5) Long-term neural function improvement [65, 131] (6) Cognitive load [132] (7) Cerebellum-Basal Ganglia-Thalamus pathway [133] (8) Complex neural circuit activation [134] (9) Environmental adaptability [135] (10) SMN [136, 137] (11) DMN [136] (12) BDNF [125] (13) Spinal Cord-Brainstem pathway [138] Changes in various neuroplasticity indicators between RAGT and TGT. (1) SMA [126] (2) PMC [127] (3) Self-organization ability [127, 128] (4) Voluntary movement intention [129, 130] (5) Long-term neural function improvement [65, 131] (6) Cognitive load [132] (7) Cerebellum-Basal Ganglia-Thalamus pathway [133] (8) Complex neural circuit activation [134] (9) Environmental adaptability [135] (10) SMN [136, 137] (11) DMN [136] (12) BDNF [125] (13) Spinal Cord-Brainstem pathway [138] Table 2Comparison of the effects of RAGT and TGT on neuroplasticityEIRAGT effectsTGT effectsComparative conclusionMSR[92, 97, 103, 121, 139–143]Regular stimulation enhances LTP, with elevated BDNF levels (+ 5.96 ng/ml)Active training activates the PMC and SMA, reinforcing the motor intention–execution loopRAGT indirectly enhances plasticity through spinal rhythmic input, while TGT directly reinforces cortical controlMBN [92, 97, 103, 121, 139–144]Enhanced connectivity between the DMN and SMN, reducing cognitive loadActivation of the cerebellum–basal ganglia–thalamus circuit, optimizing motor coordinationintegration disorders, while TGT is more effective for ataxiaAMI [90, 129]Cortical representation may weaken, relying on exogenous rhythmsActivation of the prefrontal–basal ganglia pathway, maintaining motor motivationTGT is superior in training autonomous controlLTA [32, 77, 83, 114, 143, 145]May induce a “neural dependency trap” (reduced SMN clustering coefficient)Promotes the self-organization of neural networksTGT holds greater potential in functional maintenanceEI = Evaluation Index, MSR = Microscopic Synaptic Remodeling, MBN = Macroscopic Brain Network Connectivity,AMI = Autonomous Motor Intention, LTA = Long-term Adaptability Table 2Comparison of the effects of RAGT and TGT on neuroplasticityEIRAGT effectsTGT effectsComparative conclusionMSR[92, 97, 103, 121, 139–143]Regular stimulation enhances LTP, with elevated BDNF levels (+ 5.96 ng/ml)Active training activates the PMC and SMA, reinforcing the motor intention–execution loopRAGT indirectly enhances plasticity through spinal rhythmic input, while TGT directly reinforces cortical controlMBN [92, 97, 103, 121, 139–144]Enhanced connectivity between the DMN and SMN, reducing cognitive loadActivation of the cerebellum–basal ganglia–thalamus circuit, optimizing motor coordinationintegration disorders, while TGT is more effective for ataxiaAMI [90, 129]Cortical representation may weaken, relying on exogenous rhythmsActivation of the prefrontal–basal ganglia pathway, maintaining motor motivationTGT is superior in training autonomous controlLTA [32, 77, 83, 114, 143, 145]May induce a “neural dependency trap” (reduced SMN clustering coefficient)Promotes the self-organization of neural networksTGT holds greater potential in functional maintenanceEI = Evaluation Index, MSR = Microscopic Synaptic Remodeling, MBN = Macroscopic Brain Network Connectivity,AMI = Autonomous Motor Intention, LTA = Long-term Adaptability Comparison of the effects of RAGT and TGT on neuroplasticity EI = Evaluation Index, MSR = Microscopic Synaptic Remodeling, MBN = Macroscopic Brain Network Connectivity, AMI = Autonomous Motor Intention, LTA = Long-term Adaptability At the macro-network level, fMRI and graph theory analyses have unveiled differences in whole-brain functional connectivity reorganization between the two therapies. Tang et al. found that the RAGT group showed a significant increase in the functional connectivity between the default mode network (DMN) and the sensorimotor network (SMN) [136], suggesting that mechanical assistance may alleviate cognitive load (e.g., balance anxiety) and thereby free up regulatory resources in higher-order brain regions. In contrast, the TGT group exhibited enhanced connectivity in the cerebellum-basal ganglia-thalamus circuit [133], indicative of a specific activation of pathways related to motor coordination and error correction. This network differentiation implies that RAGT may be more suitable for patients with sensorimotor integration deficits, whereas TGT might offer advantages in neuroplastic remodeling for ataxic gait. Specifically, RAGT primarily enhances the rhythmic generation capacity of the spinal cord–brainstem pathway, making it well-suited for early-stage rehabilitation, while TGT directly reinforces the motor planning abilities of the cortical-cerebellar pathway, rendering it appropriate for mid- to late-stage functional remodeling. It is important to note that the fixed motion patterns associated with RAGT may lead to a “neural dependency trap.” Molteni’s brain connectivity research demonstrated that patients who underwent long-term use of Exoskeleton Robots exhibited a significant reduction in the clustering coefficient within nodes of the SMN after discontinuing device usage [137]. This finding suggests that mechanical assistance might suppress the intrinsic self-organizing capability of neural networks. While modern RAGT systems employ sensor-driven adaptive algorithms [146, 147], further integration of neuroplasticity biomarkers, such as fNIRS-derived cortical activation patterns, could refine personalized training by dynamically aligning robotic assistance with patients’ evolving neural reorganization. Future studies should prioritize decoding the causal links between task difficulty, neural resource allocation, and long-term functional recovery to optimize closed-loop neurorehabilitation paradigms. Muscle function recovery: systemic remodeling from mechanical compensation to metabolic adaptation The recovery of muscle function should not be confined merely to the superficial enhancement of muscle strength; rather, it must address the systematic impact of various stimuli on the muscle-nerve-metabolic axis. A comparative analysis between RAGT and TGT in this regard reveals an evolutionary trajectory in rehabilitation interventions, from localized compensation to overall adap-tation, as illustrated in Fig. 5. Muscle function recovery necessitates a dual focus on both strength enhancement and the optimization of synergistic movement patterns. Future designs of Exoskeleton Robots should therefore move beyond the constraints of the sagittal plane by integrating support for multi-degree-of-freedom movements.Muscle function recovery differences are listed in Table 3. Fig. 5Changes in various muscle function indicators under RAGT and TGT. These include: (1) activation of rotational muscle groups [58] (2) activation of adductor muscle groups [58] (3) activation of deep stabilizing muscles [76] (4) balance of muscle strength [71] (5) activation of the HIF-1α pathway [148] (6) similarity to natural gait [149] (7) muscle cross-sectional area [150] (8) coordination [151] (9) activation of the PGC-1α pathway [152] (10) muscle strength [153] (11) anti-gravitational load [154] (12) activation of sagittal plane muscle groups [71] (13) training consistency [155] Changes in various muscle function indicators under RAGT and TGT. These include: (1) activation of rotational muscle groups [58] (2) activation of adductor muscle groups [58] (3) activation of deep stabilizing muscles [76] (4) balance of muscle strength [71] (5) activation of the HIF-1α pathway [148] (6) similarity to natural gait [149] (7) muscle cross-sectional area [150] (8) coordination [151] (9) activation of the PGC-1α pathway [152] (10) muscle strength [153] (11) anti-gravitational load [154] (12) activation of sagittal plane muscle groups [71] (13) training consistency [155] Table 3Differences between RAGT and TGT in muscle function recoveryEIRAGT effectsTGT EffectsComparative conclusionMCSA [155]Significant short-term muscle hypertrophy (+ 14.5%), activating the mTOR pathwayWeaker hypertrophic effect, but with a more natural synergy patternRAGT is suitable for rapid strength recovery, while TGT optimizes muscle synergyMA [41, 88, 95, 156–158]Lactate accumulation activates PGC-1α, improving oxidative metabolism(VO2max increased)Intermittent load fluctuations activate HIF-1α, enhancing anaerobic capacityTGT is superior in comprehensive muscle activationMPA [77]Primarily sagittal plane flexion–extension, with insufficient activation of adductor/rotator muscle groupsPNF diagonal patterns activate deep stabilizing musclesTGT is superior in comprehensive muscle activationNMP [90, 116, 151]The synergy pattern is less similar to natural gaitMuscle synergy optimization is closer to physiological gaitTGT has greater advantages in naturalizing movement patternsEI = Evaluation Index, MCSA = Muscle Cross-sectional Area, MA = Metabolic Adaptation, MPA = Multi-planar Activation, NMP = Naturalness of Movement Pattern Table 3Differences between RAGT and TGT in muscle function recoveryEIRAGT effectsTGT EffectsComparative conclusionMCSA [155]Significant short-term muscle hypertrophy (+ 14.5%), activating the mTOR pathwayWeaker hypertrophic effect, but with a more natural synergy patternRAGT is suitable for rapid strength recovery, while TGT optimizes muscle synergyMA [41, 88, 95, 156–158]Lactate accumulation activates PGC-1α, improving oxidative metabolism(VO2max increased)Intermittent load fluctuations activate HIF-1α, enhancing anaerobic capacityTGT is superior in comprehensive muscle activationMPA [77]Primarily sagittal plane flexion–extension, with insufficient activation of adductor/rotator muscle groupsPNF diagonal patterns activate deep stabilizing musclesTGT is superior in comprehensive muscle activationNMP [90, 116, 151]The synergy pattern is less similar to natural gaitMuscle synergy optimization is closer to physiological gaitTGT has greater advantages in naturalizing movement patternsEI = Evaluation Index, MCSA = Muscle Cross-sectional Area, MA = Metabolic Adaptation, MPA = Multi-planar Activation, NMP = Naturalness of Movement Pattern Differences between RAGT and TGT in muscle function recovery Lactate accumulation activates PGC-1α, improving oxidative metabolism (VO2max increased) EI = Evaluation Index, MCSA = Muscle Cross-sectional Area, MA = Metabolic Adaptation, MPA = Multi-planar Activation, NMP = Naturalness of Movement Pattern During stages of pathological muscle change (e.g., post-stroke muscle atrophy), Karelis et al. study reported a 14.5% increase in leg muscle cross-sectional area [150]. This effect may be attributed to the anti-gravitational load provided by the exoskeleton, which stimulates mTOR-mediated protein synthesis. However, a training regimen predominantly based on sagittal plane flexion and extension leads to insufficient activation of the adductor and rotational muscle groups [58], potentially exacerbating imbalances in muscle strength. In contrast, TGT employs multi-planar resistance training (such as the PNF diagonal pattern) to activate deep stabilizing muscles. Although its hypertrophic effects are less pronounced than those of RAGT, TGT can optimize muscle synergy through enhanced coordination training. Tanaka’s EMG study demonstrated that the TGT group exhibited increased activation of the rectus femoris alongside a reduction in antagonist muscle activation [76]; this synergistic optimization may stem from adaptive reorganization of the cerebellar–cortical pathways rather than from mere mechanical stimulation. On the metabolic adaptation front, the high-intensity training characteristic of RAGT induces lactate accumulation that activates the PGC-1α pathway, thereby promoting mitochondrial biogenesis and oxidative metabolism [152]. Do et al. revealed a significant increase in VO2max within the RAGT group [75], suggesting that enhanced muscle energy metabolism efficiency may delay post-stroke fatigue. However, the “metabolic laziness” induced by mechanical assistance, resulting in a reduction of patients’ active metabolic demand, might counteract this benefit. Conversely, TGT utilizes intermittent load fluctuations (e.g., stair climbing training) to induce HIF-1α pathway activation, which enhances anaerobic metabolic capacity. This improvement is crucial for boosting the speed of muscular responses during sudden postural adjustments [148]. Regarding long-term adaptations, a limitation of RAGT is its interference with the formation of muscle memory. Infarinato found that despite improvements in gait function, the muscle synergy patterns in the RAGT group remained less similar to those of natural gait compared to the TGT group [159]. This suggests that mechanical assistance may inadvertently reinforce non-physiological movement patterns. To resolve this issue, the control strategy of Exoskeleton Robots should transition from “fixed trajectory tracking” to “dynamic muscle activation-based torque optimization.” For instance, Ma et al. have developed a human–machine cooperative control strategy that dynamically adjusts the assistance torque profile based on electromyographic synergy similarity indices [160]. Such approaches hold promise for achieving a balance between enhancing muscle strength and naturalizing movement patterns. Discussion This systematic review adopts a three-perspective analytical framework, examining motor function, neuroplasticity, and muscle recovery separately, to elucidate the distinct therapeutic mechanisms of RAGT and TGT in post-stroke rehabilitation. Rather than pursuing quantitative meta-analysis, we employed this mechanistic approach because the substantial clinical, methodological, and statistical heterogeneity across the 86 included studies would render any pooled effect size misleading. The rationale for this decision is grounded in three critical observations. First, in the subacute phase of stroke recovery, the critical window for neuroplasticity, data synthesis revealed that gait endurance (6-minute walk test, 6MWT), gait speed (10-meter walk test, 10MWT), and dynamic balance (Berg Balance Scale, BBS) improved significantly in the majority of trials for both RAGT and TGT groups, regardless of intervention modality [30, 34, 35, 40, 56, 97, 105, 106, 112, 115, 161]. However, a minority of trials showed no improvement in either group. Some trials reported that both RAGT and TGT yielded significant improvements, others found RAGT superior to TGT [29, 33, 43, 53, 66, 90, 91, 93, 122, 162, 163], while a few concluded that RAGT was equivalent to or even inferior to TGT. These contradictory findings cannot be reconciled through simple numerical averaging; instead, they demand a deeper mechanistic investigation, which we provide through our three-perspective framework. Through critical appraisal of the included literature, these discrepancies can largely be attributed to three underlying methodological and physiological factors that constitute the primary sources of heterogeneity: The Baseline Severity Paradox The most significant source of heterogeneity lies in the baseline functional status of the cohorts. RAGT demonstrates clear, statistically significant superiority primarily in severely impaired, non-ambulatory patients (e.g., Functional Ambulation Category [FAC] 0–1) [31, 40, 47, 55, 111]. In these patients, RAGT safely delivers the high-repetition, standardized kinematic stepping cycles that physical therapists cannot manually sustain. Conversely, for moderately impaired patients capable of initiating stepping, TGT yields equivalent outcomes, suggesting that the “ceiling effect” of robotic assistance is reached earlier in patients with residual motor control [26, 28, 161]. This baseline-dependent efficacy pattern fundamentally precludes pooling studies with heterogeneous severity profiles, as the treatment effect is not constant across the severity spectrum. Assistance strategy and the “slacking” effect Contradictory outcomes are strongly correlated with the control algorithms of the robotic devices utilized. Studies employing 100% fixed-trajectory assistance often report inferior or equivalent results to TGT, as excessive guidance diminishes active muscular engagement and induces the “slacking” effect [44, 164]. In contrast, trials utilizing Assist-As-Needed (AAN) paradigms or progressive force reduction show significant advantages in lower limb muscle strength and balance [81, 165], highlighting that active central nervous system (CNS) engagement is a prerequisite for functional gains. Pooling fixed-trajectory and AAN studies would artificially dilute the true efficacy of optimized RAGT protocols. Kinematic quality vs. absolute speed A recurrent discrepancy arises between clinical scales and biomechanical assessments. While TGT and RAGT may produce identical absolute improvements in the 10MWT, the underlying biomechanical strategies often differ profoundly. TGT patients frequently increase speed through compensatory mechanisms (e.g., increased step width, hip hiking). In contrast, RAGT significantly improves underlying gait quality, evidenced by enhanced spatiotemporal symmetry, increased joint excursion, and reduced abnormal co-contraction indices (CCI) [166]. Thus, relying solely on absolute speed metrics may mask the specific biomechanical and neuroplastic advantages conferred by RAGT, further invalidating the assumption of a common treatment effect necessary for meta-analysis. Given these fundamental sources of heterogeneity, we concluded that quantitative pooling would obscure rather than illuminate the true therapeutic value of RAGT and TGT. Instead, our three-perspective analytical framework, motor function, neuroplasticity, and muscle recovery, provides mechanistically grounded insights that acknowledge and leverage this heterogeneity to identify clinically actionable rehabilitation strategies. The sections below elaborate on each perspective. Motor function TGT and RAGT exhibit complementary characteristics in motor function recovery. TGT is more flexible, allowing for personalized adjustments based on the patient’s individual rehabilitation progress, thereby enhancing active participation. Through task-oriented training (e.g., the Brunnstrom approach and PNF techniques), it improves lower limb strength, coordination, and balance. Traditional methods are limited by treatment intensity, especially when patients are physically exhausted or therapists become fatigued, making it difficult to achieve high-intensity repetitive training. In contrast, RAGT provides patients with high-intensity, repetitive gait training in a short period via standardized assistance modes (e.g., fixed trajectory control and torque compensation), significantly improving gait speed, symmetry, and stability [13, 123]. This effect is particularly pronounced in patients with moderate to severe motor impairments. The synergistic effect of mechanical assistance and active training can overcome the limitations of a single mode. Future efforts should further optimize the adaptive control algorithms of RAGT, integrating personalized adjustment strategies (such as assistance torque profile optimization based on real-time EMG signals [160]) while maintaining training intensity to balance standardization with flexibility. Neuroplasticity The period from the first week to the first month post-stroke (acute and early subacute phases) is critical for neuroplasticity [167–169]. The nervous system restores motor ability by reshaping damaged neural pathways. TGT stimulates cortical neural activity and enhances synaptic connections. This active training effectively engages neural networks associated with specific motor tasks, enabling patients to gradually recover motor function through repeated practice. It markedly facilitates neural reorganization by reinforcing activity in the cerebral cortex and motor-related regions. Exoskeleton Robots have demonstrated superior outcomes over traditional training in gait control and neural circuit reconstruction. Nonetheless, because the movement mode of RAGT is primarily confined to the sagittal plane, it may not fully activate all neural networks related to complex gait. Therefore, future rehabilitation should combine the high-intensity repetitive training of RAGT with the comprehensive neural stimulation provided by traditional methods to better promote neuroplasticity and overall motor function recovery. It is noteworthy that post-stroke neurological deficits are not only manifested as motor impairments but are also accompanied by significant fatigue. Recent meta-analyses show that 47% (95% CI = 43–50%) of stroke survivors suffer from post-stroke fatigue [170]. This pathological state, often misdiagnosed as depression or apathy, is in fact closely related to neural network dysfunction [171]. Evidence from neuromodulation studies supports this; non-invasive brain stimulation techniques such as transcranial direct current stimulation can significantly alleviate fatigue symptoms by modulating neural network excitability [172, 173]. Additionally, exercise can improve aerobic regulation, reduce mild to moderate depression and anxiety, enhance sleep quality, increase dopamine levels, boost central nervous system growth factors, and reduce inflammation [174, 175], thereby mitigating fatigue in stroke patients and, to some extent, improving neural network dysfunction. This highlights the need to pay special attention to the interaction between neuroplasticity mechanisms and fatigue symptoms when formulating rehabilitation protocols. Muscle recovery The recovery of lower limb muscle function is key to restoring motor function in stroke patients. TGT enhances muscle strength and coordination through resistance training, balance exercises, and joint mobility practices, with core muscle training being critical for overall stability and functional recovery. However, the limitation of TGT is that therapists cannot continuously provide high-intensity repetitive training, resulting in suboptimal recovery of muscle strength in some cases. RAGT, by offering continuous assistance, can significantly improve muscle strength, especially during the early rehabilitation phase, as it accelerates the recovery of major lower limb muscles such as the quadriceps and gastrocnemius. Nonetheless, because RAGT primarily focuses on the sagittal plane, it is challenging to effectively train muscle groups involved in rotational and lateral movements. Thus, while RAGT shows significant advantages in muscle recovery speed, it still has limitations in comprehensively restoring lower limb function. Research findings and clinical significance RAGT has demonstrated significantly superior outcomes compared to TGT in stroke patients with moderate to severe motor impairments, particularly in terms of gait stability and muscle strength enhancement [73]. For patients with residual motor ability or mild stroke and with high demands for daily living activities, TGT can maximize their recovery potential. Compared with conventional methods, robotic-assisted gait training, especially when using demand-based assistance algorithms, shows better performance in helping severely disabled patients recover symmetrical gait patterns. However, despite the significant improvements in gait function provided by robotic-assisted rehabilitation, its effects do not clearly surpass those of therapist-provided manual assistance. TGT and occupational therapy have unique advantages in enhancing patient engagement, building confidence, and improving quality of life. Although TGT performs well in restoring physiological movement patterns and improving coordination, its outcomes are highly dependent on the therapist’s skill and training intensity, and research findings remain inconclusive. Rehabilitation is not merely about the recovery of physical function but also about rediscovering the nervous system’s capacity for self-repair. Although traditional rehabilitation methods are effective, they are constrained by the physical and time demands on therapists, making it difficult for all patients to achieve deep neural activation. RAGT, on the other hand, has shown deeper stimulation of neuroplasticity by emphasizing neuronal remodeling and the strengthening of synaptic connections for functional recovery. This training mode breaks the limitations of therapist-dominated traditional rehabilitation by closely integrating rehabilitation with technology, thereby demonstrating the potential of “human-machine collabora-tion” in future healthcare. While traditional gait rehabilitation relies on the therapist’s technique and experience, Exoskeleton Robots provide standardized, repeatable high-intensity training marking a transition from labor-intensive to technology-intensive rehabilitation. Moreover, despite the irreplaceable advantages of exoskeleton devices in rehabilitation, their widespread application faces a series of challenges. The high cost of exoskeleton devices limits their adoption in many healthcare institutions, leading to an imbalance in rehabilitation resources. This issue concerns not only technological accessibility but also social equity, raising the question of how to ensure that every stroke patient has access to the best rehabilitation program. The recommended intervention timings for RAGT and TGT during the recovery process are illustrated in Fig. 6. The window for limb motor recovery post-stroke ranges from several hours to several months [176]. Endogenous plasticity enhancement and improvements in both structural damage and function peak in the acute phase (1–7 days post-onset) [177]. Most behavioral recovery and rapid changes occur within the first week to several months after stroke [177]. Few rehabilitation trials initiate restorative treatment in the later stages post-stroke [178, 179]. However, recent studies indicate that beginning training within 7 days of stroke onset achieves a significantly higher percentage of gait recovery compared to control groups starting training between 7 and 90 days post-stroke [180]. Nevertheless, for patients with hemorrhagic stroke, it is still recommended to commence vigorous exercise or high-intensity exoskeleton-assisted gait training only after the patient’s condition stabilizes, to prevent secondary hemorrhage. A staged rehabilitation strategy is advised: during the first 1–7 days post-stroke (acute phase), TGT should be the primary method to enhance basic motor function through personalized training; then, during 1–12 weeks (early subacute phase), RAGT should be gradually introduced for high-intensity gait training in patients with moderate to severe impairments to maximize functional recovery. Fig. 6Intervention timing of RAGT and TGT in stroke rehabilitation Intervention timing of RAGT and TGT in stroke rehabilitation Quality of evidence The methodological quality of the 86 included studies was evaluated using the PEDro scale [181]. The mean PEDro score was 7.04 (range: 4–9), with 82 studies (95.3%) classified as high quality (≥ 6 points), 4 studies (4.7%) as moderate quality (4–5 points), and none as low quality (≤ 3 points). Risk of bias in the included RCTs was further assessed using the revised Cochrane Risk of Bias tool (RoB 2; Fig. 7). Domain-level results indicated consistently high proportions of low-risk ratings: randomization process (85%), deviations from intended interventions (90%), missing outcome data (98%), measurement of the outcome (80%), and selection of the reported result (80%). At the overall bias level, 67% of RCTs were rated as low risk, 10% as some concerns, and 23% as high risk; the elevated high-risk proportion in the overall judgment reflects convergent concerns primarily in the outcome measurement and selective reporting domains. Collectively, these quality indicators confirm that the body of evidence is predominantly of moderate-to-high methodological quality. Results derived from the minority of studies with high risk of bias or moderate PEDro scores are interpreted with appropriate caution in the sections below. Fig. 7Risk of bias summary presented as percentages across all included Risk of bias summary presented as percentages across all included Limitations and future directions The limitations of this review include high heterogeneity among the included studies, such as variations in the patients’ stages (from acute to chronic), intervention durations (4–12 weeks), and exoskeleton types (mobile vs. fixed), which may affect the generalizability of the conclusions. Future research should involve multi-center, large-sample randomized controlled trials to validate these findings. Moreover, of the 86 studies, 66 (76.7%) reported outcomes only at the immediate post-intervention time point, 5 (5.8%) included follow-up at 3 months, and only 3 (3.5%) extended follow-up beyond 6 months.Over 96% of the trials lacked follow-up beyond 6 months, which masked long-term maintenance of function.Exoskeleton and robotic-assisted rehabilitation have demonstrated significant short-term functional improvements, especially in resource-limited settings where they can alleviate therapists’ burdens, but their long-term effects and the maintenance of function after discontinuing device use require further study. Additionally, due to the limited degrees of freedom of current exoskeleton structures, there is a need to develop multi-degree-of-freedom self-balancing exoskeleton systems that can facilitate movements in the sagittal, frontal, and horizontal planes, thereby more comprehensively activating patients’ neural networks. Future research should aim to optimize the integration of exoskeleton devices with traditional rehabilitation strategies to meet the diverse needs of different patient groups, particularly considering individual differences such as age, weight, and condition, and explore ways to improve long-term rehabilitation outcomes and multi-dimensional motor function. With advancements in artificial intelligence, machine learning, and neuroimaging techniques, exoskeleton devices can be further optimized to dynamically adjust training parameters based on patients’ neural feedback and a broader range of inputs, thereby more accurately assessing current status and needs. This approach will not only optimize the process of neuroplasticity stimulation but also enable more efficient restoration of specific impaired regions through personalized training protocols. Future rehabilitation models can thus be explored from the perspectives of multimodal training and long-term efficacy evaluation. Additionally, while the semi-quantitative evidence profiles presented in radar charts (Figs. 3, 4 and 5) provide a useful multidimensional overview of comparative rehabilitation outcomes, this approach inherently carries a degree of interpretive judgment. Although we have mitigated this through a structured scoring rubric, independent dual-reviewer assessment, and integration with validated quality metrics (PEDro scores), readers should interpret these visualizations as evidence synthesis tools rather than statistical effect sizes. Future studies employing standardized outcome measures across comparable interventions may permit formal quantitative meta-analysis for individual domains, which would complement the mechanistic insights provided by our three-perspective analytical framework. Conclusion RAGT significantly improves gait parameters and promotes neuroplasticity through high-intensity, standardized training, whereas TGT enhances patients’ active participation and functional independence through individualized interventions. An ideal rehabilitation strategy should combine the flexibility of traditional rehabilitation with the high-intensity training provided by exoskeleton technology. By employing traditional methods in the early stages to enhance basic functions and gradually introducing exoskeleton-assisted training as patients adapt, using exoskeleton devices to reinforce fundamental gait training while traditional methods refine motor skills and bolster daily self-care capabilities, the overall comprehensiveness and effectiveness of treatment can be improved. This integrated rehabilitation strategy will better enhance overall recovery outcomes in stroke patients and also guide future intelligent decision-making research for the transition of Exoskeleton Robots from “mechanical compensation” to “autonomous control.”The recommendations advanced in this review are grounded in a body of evidence of predominantly moderate-to-high methodological quality, as confirmed by PEDro scale assessment (mean score 7.04/10; 95% of studies rated ≥ 6 points) and RoB 2 evaluation of the included RCTs (67% rated low risk of overall bias). These quality indicators strengthen confidence in the clinical applicability of our conclusions. The elevated high-risk overall bias rating observed in 23% of RCTs, attributable largely to outcome measurement and selective reporting concerns, and the four studies of moderate PEDro quality represent residual limitations that future well-designed randomized controlled trials should seek to address.