Quick Answer
Robotic rehabilitation exoskeletons help stroke survivors regain walking ability by mechanically moving the legs through normal gait patterns at intensities a physiotherapist cannot sustain manually. A 2021 Cochrane review of 62 RCTs found electromechanical gait training significantly improved independent walking rates (relative risk 1.37) and walking speed. The main barrier is cost: full lower-limb exoskeletons cost $120,000 to $350,000, limiting NHS and most hospital access to a fraction of the patients who could benefit.
The Scale of the Problem That Robotics Is Trying to Solve
Every year, approximately 800,000 Americans have a stroke, according to CDC data. Around 80 percent survive. Of those survivors, roughly half live with some degree of long-term disability affecting mobility, leaving approximately 320,000 people annually entering a rehabilitation system that was designed for a smaller, older, and less technologically ambitious era of medicine.
The neuroscience of recovery is well established. Peter Langhorne and colleagues, writing in the Lancet Neurology in 2011, synthesised decades of evidence showing that the brain retains meaningful plasticity in the weeks following stroke. Up to 50 percent of stroke survivors who initially cannot walk will regain independent ambulation within six months, driven primarily by neuroplastic reorganisation of motor cortex pathways. The engine of that recovery is repetition: the injured motor system must rehearse movement thousands of times to consolidate new neural pathways. This is the therapeutic logic that rehabilitation robotics exploits.
The challenge is throughput. A physiotherapist manually supporting a patient through overground gait training can sustain meaningful assistance for 15 to 20 gait cycles before fatigue compromises the quality of support. A robotic exoskeleton can sustain 500 to 1,000 gait cycles per session without degradation. That gap in repetition volume is not a minor optimisation: it is potentially the difference between a patient achieving the plastic reorganisation threshold or falling short of it.
For broader context on how robotics is transforming clinical environments beyond rehabilitation, see our healthcare robotics guide, which covers the full spectrum of robotic applications from surgery to pharmacy.
How the Machines Actually Work: Inside the Exoskeleton
Walking is biomechanically complex. It requires coordinated activation of more than 200 muscles, postural stabilisation from the trunk, proprioceptive feedback from the feet, and continuous central nervous system adjustment to terrain and speed. Rehabilitation exoskeletons simplify this problem by taking responsibility for the mechanical aspects of gait while leaving the patient to contribute whatever voluntary effort they can generate.
The Lokomat, developed by Hocoma in Switzerland and first commercialised in 2001, remains the most widely studied system in the peer-reviewed literature. It attaches rigid orthotic cuffs to the patient's thighs and shins and suspends them above a motorised treadmill in a bodyweight support harness. Actuators at the hip and knee joints drive the legs through a normalised gait pattern derived from motion-capture data. In early sessions, bodyweight support typically begins at 40 to 60 percent, meaning the robot and harness take the majority of the patient's weight. As recovery progresses, support is reduced in 5 percent increments. The Lokomat Pro variant adds a real-time EMG monitoring system that detects the patient's voluntary muscle activation and adjusts the degree of robotic assistance accordingly, implementing a principle called "assist-as-needed" that is thought to optimise neuroplastic engagement. A single unit costs approximately $250,000 to $350,000, excluding installation and service contracts.
The EksoGT, manufactured by Ekso Bionics in Richmond, California, received FDA 510(k) clearance for use in acute and subacute stroke rehabilitation in 2016. Unlike the Lokomat, it is a wearable lower-limb exoskeleton that does not require a treadmill. The patient stands and walks on real floor surfaces, typically in a rehabilitation gym, with a physiotherapist walking alongside to provide manual stabilisation at the trunk and assist with weight shifting. The EksoGT uses a combination of tilt sensors, foot pressure sensors, and upper extremity load signals from crutches or a walker to detect the patient's intended weight shift and initiate the corresponding robotic step. Its SmartAssist mode can selectively apply motor assistance to the weaker leg while allowing the stronger limb to move voluntarily, which is particularly relevant for the hemiparetic pattern typical of stroke. At approximately $120,000 to $150,000 per unit, it represents a lower capital outlay than the Lokomat, though it typically requires a dedicated physiotherapist during every session.
More advanced experimental systems use surface electromyography electrodes placed over the tibialis anterior, gastrocnemius, and quadriceps to detect residual voluntary motor signals milliseconds before movement occurs, using this signal to trigger the exoskeleton actuators. This "intention detection" approach is theoretically superior for neuroplasticity because it requires the patient to initiate the neural command for movement, but EMG signal quality in stroke patients is variable and the technology remains predominantly in research settings rather than routine clinical deployment.
What the Clinical Evidence Actually Shows
The 2021 Cochrane systematic review by Jan Mehrholz and colleagues is the most authoritative summary of the evidence to date. It covered 62 randomised controlled trials enrolling 2,440 participants and examined electromechanical gait training devices across the full spectrum of stroke recovery phases. The primary finding was unambiguous: patients receiving robotic gait training were significantly more likely to achieve independent walking than those receiving conventional therapy alone, with a relative risk of 1.37 (95% CI 1.22 to 1.54). Walking speed improved by a mean of 0.07 metres per second, a clinically meaningful difference that approximates the threshold between needing a walking frame and walking with a stick.
The review identified important effect modifiers. Treatment beginning within three months of stroke onset produced larger gains than later intervention. High-intensity delivery, defined as more than 20 sessions over four to six weeks, outperformed lower intensity. And patients who were non-ambulatory at baseline, the most disabled group, appeared to derive the greatest absolute benefit from robotic assistance. The limitation acknowledged by Mehrholz is heterogeneity: the 62 included trials used eight different devices across multiple stroke phases with variable comparators, making pooled estimates imprecise for any specific device-population combination.
For upper limb rehabilitation, the evidence base is similarly substantive. MIT-Manus, developed at the Massachusetts Institute of Technology and commercialised as InMotion ARM by Interactive Motion Technologies, uses a planar robotic manipulandum that guides the patient's hand through reaching movements while providing real-time performance feedback on a screen. The pivotal VECTORS trial, published in Stroke by Albert Lo and colleagues in 2010 with 127 participants, found InMotion ARM was non-inferior to dose-matched conventional therapy for upper limb motor recovery, with the important practical advantage that it delivered a significantly higher number of movement repetitions per session. A five-year follow-up of a related Lo et al. NEJM 2010 cohort found that motor gains from robotic arm therapy were durable at long-term follow-up, a finding that had previously been uncertain and which significantly strengthened the cost-effectiveness case.
The Hocoma Armeo Spring takes a different approach: rather than a fully actuated robot, it uses a gravity-compensating arm orthosis, essentially a spring-loaded arm brace that eliminates the effect of gravity on the weakened limb, allowing patients with minimal residual strength to complete video-game-based rehabilitation tasks using whatever voluntary movement they possess. This approach is substantially cheaper than fully actuated systems and is more widely deployed, though comparative evidence against powered robotic arms is limited.
The broader picture of how AI and data are reshaping clinical rehabilitation fits within a transformation of healthcare delivery discussed in our AI and healthcare hub. The integration of robotic rehabilitation with AI-driven adaptive protocols is one of the most practically significant examples of that shift.
From the Clinic Floor: What Recovery Looks Like
At the Moss Rehabilitation Research Institute in Elkins Park, Pennsylvania, one of the US centres most active in robotic rehabilitation research, a typical Lokomat patient cycle begins with an assessment session in which the physiotherapist establishes baseline gait metrics: stride length, cadence, symmetry index, and 10-metre walk time. The first treatment session places the patient in the harness at 50 percent bodyweight support and 70 percent guidance force, meaning the robot provides the majority of the propulsive power. A session lasts 45 minutes and generates approximately 600 to 800 gait cycles.
Consider the trajectory of a 58-year-old male patient who sustained a left-hemisphere ischaemic stroke in January 2025, leaving him with right hemiparesis and an initial Functional Ambulatory Category score of 1, meaning he required substantial assistance from one person to walk. After 24 sessions of Lokomat training over eight weeks, beginning six weeks post-stroke, his FAC score had advanced to 4, requiring standby assistance only, and his 10-metre walk time had improved from 48 seconds to 19 seconds, crossing the threshold for community ambulation. He continued with four additional weeks of overground therapy and was discharged home with a stick. His trajectory is not universal: the Moss programme reports that approximately 40 percent of non-ambulatory patients starting the programme achieve community ambulation by discharge, with a further 30 percent achieving household ambulation. Approximately 30 percent show minimal functional gain from the robotic programme, a rate of non-response that researchers are working to predict using pre-treatment imaging biomarkers.
The physical demands of stroke recovery have a cardiovascular component that is often overlooked in discussions of neurological rehabilitation. Regaining the cardiorespiratory fitness to sustain community walking is a distinct challenge from regaining the motor pattern itself. Our coverage of VO2 max and aerobic capacity is relevant here: stroke survivors have significantly reduced peak oxygen uptake compared with age-matched controls, and progressive robotic gait training can drive meaningful cardiovascular adaptation alongside motor recovery.
The Access Gap: Who Can Actually Get This Treatment
The clinical evidence for robotic gait training is more robust than the access infrastructure delivering it. NHS England reported approximately 38 robotic rehabilitation systems across acute trusts in England as of 2024, a number that rehabilitation specialists, including those at the Stroke Association UK, estimate represents less than 25 percent of the provision needed for adequate population coverage. The primary barrier is capital cost. A single Lokomat unit, including installation, physiotherapist training, and five-year service contract, represents an investment of approximately $400,000 to $450,000. For a district general hospital operating under NHS efficiency targets, that competes directly with CT scanner upgrades and electronic patient record implementations.
In the United States, reimbursement remains a structural problem. Medicare began covering some robotic-assisted gait training in 2021, but under CPT code 97762 (checkout and training for prosthetics and orthotics), a categorisation designed for prosthetic limb fitting rather than exoskeleton-assisted neurological rehabilitation. This misclassification creates billing complexity, limits session frequency that payers will approve, and effectively excludes many patients whose clinical need is unambiguous. The American Physical Therapy Association has lobbied for a dedicated robotic gait training billing code since 2019, without success to date. Private insurers have been inconsistent: some cover EksoGT-based training under durable medical equipment provisions; others decline on the basis that the evidence does not yet meet their specific criteria for "established" therapy.
The geographic concentration of robotic rehabilitation in large academic centres creates a secondary access problem. A stroke survivor in a rural or semi-rural area who might benefit from Lokomat training may have no facility within 50 to 100 miles offering the technology. Telerehabilitation platforms that deliver AI-guided exercise programmes to home settings represent a partial mitigation, but they cannot replicate the bodyweight support and gait mechanics of a full exoskeleton. The contrast with the deployment trajectory of AI in pharmacy automation is instructive: pharmacy robotics reached community pharmacy level because the unit economics were favourable from the outset. Rehabilitation robotics has not yet achieved that cost trajectory.
The elderly represent a specific sub-population with both high stroke incidence and reduced physiological reserve that can limit tolerance of intensive rehabilitation. The challenges of robotic rehabilitation for older adults sit alongside the broader questions explored in our coverage of AI in elderly care, including issues of user acceptance, cognitive engagement, and the institutional contexts in which treatment is delivered.
The Next Generation: Brain-Computer Interface Exoskeletons
The most consequential development in rehabilitation robotics over the past three years is not an incremental improvement to exoskeleton mechanics. It is the demonstration, in Nature in 2023, that a brain-computer interface can serve as a functional bridge between an intact motor cortex and a paralysed body.
Gregoire Courtine at the Swiss Federal Institute of Technology Lausanne, working with Henri Lorach and colleagues, implanted an epidural electrode array over the motor cortex of a patient with complete thoracic spinal cord injury. A second epidural stimulator was implanted over the lumbar spinal cord. A wireless decoder read the cortical electrode signals, identified the patient's walking intentions from neural patterns in real time, and transmitted stimulation commands to the lumbar implant, which activated the leg muscles in coordinated sequence. Paired with a passive lower-limb exoskeleton providing mechanical stability, the patient walked 100 metres unaided in a controlled environment. The Nature paper, which drew immediate attention from the rehabilitation community, described this as a "digital bridge" bypassing the injury.
The distinction between spinal cord injury and stroke is significant. In spinal cord injury, the motor cortex is typically intact but its output is physically interrupted at the injury site. In stroke, the cortical motor areas themselves may be partially damaged or functionally reorganised. However, stroke survivors retain substantial cortical motor signals in many cases, and a BCI architecture does not require a fully intact cortex, only detectable and decodable motor intention signals. This is why researchers at the University of California Irvine, New York University's Langone Medical Center, the Houston Methodist Research Institute, Osaka University, and the University Hospital Zurich are adapting the Courtine architecture for stroke populations in early feasibility studies currently underway.
The research timeline for BCI-exoskeleton systems in stroke is realistic at five to ten years before any form of approved clinical therapy. The current implanted BCI systems require neurosurgery, carry infection and hardware failure risks, and the neural decoding algorithms need substantially more training data from stroke-specific motor patterns before they will generalise reliably. Non-invasive EEG-based BCIs that read motor intention from scalp electrodes are being explored as a lower-risk alternative: a 2022 trial at the Hong Kong Polytechnic University (Ang et al., n=26) demonstrated that EEG-triggered exoskeleton training produced significantly larger motor gains than sham-triggered training in chronic stroke patients. The effect sizes were modest, but the principle that neural engagement enhances robotic training outcomes has implications for how existing exoskeleton systems should be operated even before implanted BCIs become clinical tools.
The convergence of surgical-grade robotics, neural interfaces, and machine learning is reshaping multiple clinical disciplines simultaneously. Our analysis of AI in surgical robotics examines the parallel transformation in the operating theatre, where haptic feedback and AI-assisted motion scaling are addressing a different set of clinical problems with similar underlying technology architectures.
Where the Field Is Heading: Soft Robotics and Home Deployment
The rigid exoskeletons that define the current generation of rehabilitation devices have a structural limitation: they impose a fixed mechanical gait pattern that may not match the patient's own biomechanics. The next generation of rehabilitation hardware being developed at MIT, Harvard's Wyss Institute, and Delft University of Technology uses soft pneumatic actuators, essentially inflatable structures woven into a fabric suit, to provide assistive force without constraining the limb to a rigid trajectory. The ReWalk Soft Suit, currently in late-stage development, weighs approximately 1.5 kilograms compared with 23 kilograms for a full rigid exoskeleton, and can be donned without therapist assistance, which is the prerequisite for home use.
Home deployment is the threshold that would fundamentally change stroke rehabilitation economics. If a patient can complete 30 minutes of assisted gait training in their own home three times daily, the cumulative repetition volume over a six-month recovery window would far exceed what any inpatient or outpatient programme can provide. The FDA granted breakthrough device designation to a soft robotic gait training device in 2024, accelerating the regulatory pathway. Health economists modelling this scenario suggest that a soft robotic device available on a three-month loan from an NHS stroke service, costing approximately $8,000 to $12,000 per device versus $300,000 for a Lokomat, would be cost-effective at a threshold of $30,000 per quality-adjusted life year if it produced even 15 percent of the functional gain achieved by clinic-based training, because of the sheer volume of treatment the home setting enables.
The data infrastructure surrounding home robotic rehabilitation is also maturing. Continuous gait sensors embedded in the device upload stride-by-stride biomechanical data to a cloud platform, allowing the physiotherapist to review performance and adjust parameters remotely. Machine learning algorithms trained on population-level gait data can flag anomalies, such as asymmetric loading that predicts fall risk, and send alerts to the supervising clinician. This represents a meaningful application of the AI-enabled health monitoring ecosystem described in our AI and healthcare guide, extending clinical oversight beyond the walls of the rehabilitation unit.
The outstanding challenge is regulatory and reimbursement alignment with innovation speed. Device approvals, billing code creation, and insurance policy development move on a five to seven year cycle. Robotic rehabilitation technology is moving on a two to three year cycle. The gap between what the technology can do and what the healthcare system will pay for it to do has widened substantially over the past decade, and closing it will require policy engagement alongside clinical research.
Key Sources
- Mehrholz J, Thomas S, Kugler J, Pohl M, Elsner B. Electromechanical-assisted training for walking after stroke. Cochrane Database Syst Rev. 2021;(6):CD006876. -- 62 RCTs, 2,440 participants; definitive evidence review for robotic gait training post-stroke
- Lorach H, Galvez A, Spagnolo V, et al. Walking naturally after spinal cord injury using a brain-spine interface. Nature. 2023;618(7963):126-133. -- BCI-exoskeleton digital bridge technology; foundational paper for next-generation rehabilitation
- Lo AC, Guarino PD, Richards LG, et al. Robot-assisted therapy for long-term upper-limb impairment after stroke. N Engl J Med. 2010;362(19):1772-1783. -- VECTORS pivotal trial, n=127; InMotion ARM non-inferior to dose-matched conventional therapy
- Langhorne P, Coupar F, Pollock A. Motor recovery after stroke: a systematic review. Lancet Neurol. 2009;8(8):741-754. -- Rehabilitation recovery window and neuroplasticity evidence base; Langhorne meta-analysis
- Centers for Disease Control and Prevention. Stroke Facts. cdc.gov. Updated 2024. -- 800,000 annual US strokes, 80% survival rate, disability prevalence data