Last reviewed
Apr 26, 2026
Sources
29 citations
Review status
Source-backed
Revision
v1 ยท 5,596 words
Add missing citations, update stale details, or suggest a clearer explanation.
| Exoskeleton | |
|---|---|
| Type | Wearable robotics |
| Also known as | Powered exoskeleton, exosuit, wearable robot, bionic suit |
| Primary domains | Medical rehabilitation, industrial ergonomics, military load assistance, recreation |
| Earliest powered prototype | GE Hardiman (1965) |
| First FDA-cleared personal device | ReWalk (2014) |
| First FDA-cleared self-balancing system | Wandercraft Atalante X (2024) |
| Common sensors | Surface electromyography (sEMG), inertial measurement units (IMUs), foot-pressure insoles, force/torque cells, joint encoders |
| Common AI methods | Deep learning, reinforcement learning, recurrent and convolutional networks, sim-to-real transfer |
| Key research labs | UC Berkeley Robotics and Human Engineering Lab, Harvard Biodesign Lab, Stanford Biomechatronics, University of Tsukuba Cybernics Lab |
| Estimated 2025 market size | USD 2.1 to 3.4 billion (varies by analyst) |
An exoskeleton, sometimes called a powered exoskeleton, exosuit, or wearable robot, is a powered or passive mechanical structure worn on the body to augment, assist, or restore human movement. Modern exoskeletons combine actuated joints, lightweight structural members, biological sensors and increasingly sophisticated artificial intelligence controllers that interpret a wearer's intent and modulate assistance in real time. The field sits at the intersection of robotics, biomechanics, neural engineering, and wearable robotics, and it has produced devices that allow paralyzed patients to walk, factory workers to lift heavy loads with reduced injury risk, and soldiers to carry larger payloads over longer distances.
The modern exoskeleton industry is roughly 60 years old in concept and about 15 years old as a practical commercial reality. After a half-century of research-grade prototypes, lower-limb medical exoskeletons received their first U.S. Food and Drug Administration (FDA) clearances between 2014 and 2016, with ReWalk Robotics, Ekso Bionics, Indego, and Cyberdyne HAL all receiving regulatory clearance for clinical or personal use during that window.[1][2] The industrial segment commercialized in parallel, with passive lifting vests from Comau, German Bionic, Ottobock, Hyundai and several other manufacturers reaching factory floors at Ford, BMW, Toyota, Stellantis, Boeing and Airbus by the late 2010s.[3] Military programs, beginning with General Electric's Hardiman in 1965 and continuing through DARPA's BLEEX program, the Lockheed Martin Human Universal Load Carrier (HULC) and the canceled Tactical Assault Light Operator Suit (TALOS), have repeatedly demonstrated the difficulty of building powered armor that does not exhaust the wearer faster than it helps them.[4]
The most consequential change in the past five years has been the maturation of AI-driven control. Older exoskeletons relied on hand-tuned finite-state controllers that switched between preset assistance profiles for walking, sitting and stair climbing. Current research systems use deep learning for gait phase detection from inertial and electromyographic signals, reinforcement learning policies trained in simulated musculoskeletal environments, and increasingly transformer-based foundation models that generalize across users and activities without per-subject tuning.[5][6] In June 2024 a Nature paper from Stanford's biomechatronics group demonstrated a hip exoskeleton trained entirely in simulation that reduced metabolic cost by 24.3% during walking, 13.1% during running and 15.4% during stair climbing on its first deployment, with no per-subject calibration; the result is widely cited as a milestone for sim-to-real transfer in wearable robotics, although follow-up work in 2026 has reexamined some of the reported energy figures.[6][7]
The term exoskeleton, borrowed from biology, originally referred to the rigid external skeletons of arthropods and crustaceans. In engineering it now refers to any wearable mechanical system that runs in parallel with the wearer's musculoskeletal system, with the intent of either augmenting human strength or endurance, restoring mobility lost to injury or disease, or reducing musculoskeletal strain during repetitive labor. Devices range from rigid, full-body powered armor to soft, fabric-based "exosuits" that look more like athletic apparel than machinery.
Three characteristics distinguish a wearable robotic exoskeleton from a more conventional orthosis or brace:
Passive exoskeletons, such as the Comau MATE-XT and most industrial back-support vests, contain no electric power source. Instead they use carefully tuned springs or elastic elements to redistribute load. Active or powered exoskeletons add motors, batteries, and a software controller, which raises capability but also raises weight, cost, and the risk that the device fights the wearer rather than helping.
| Category | Primary Goal | Typical Users | Representative Products |
|---|---|---|---|
| Medical and rehabilitation | Restore or retrain gait after spinal cord injury, stroke, multiple sclerosis, brain injury, cerebral palsy | Inpatient and outpatient rehab patients, personal users with paraplegia | ReWalk, Ekso GT, EksoNR, Indego, Cyberdyne HAL, Wandercraft Atalante X |
| Industrial and occupational | Reduce musculoskeletal strain during lifting, overhead work, kneeling and prolonged standing | Automotive assembly, warehouse, construction, shipbuilding, agriculture | Comau MATE-XT, German Bionic Cray X, Hyundai H-WEX and CEX, Levitate Airframe, Ottobock Paexo |
| Military and load-bearing | Carry heavy loads over long distances without fatiguing the wearer | Infantry, special operations, logistics personnel | GE Hardiman (research), DARPA BLEEX, Lockheed Martin HULC, Lockheed ONYX, Sarcos Guardian XO |
| Augmentation and recreation | Increase performance or endurance for able-bodied users in sport or daily life | Skiers, hikers, mobility-limited consumers | Roam Robotics Elevate, Roam Ascend, Arc'teryx MO/GO (with Skip Innovations), Hypershell |
| Pediatric and home-care | Restore mobility in children with neuromuscular disease or rehabilitate at home | Pediatric patients with cerebral palsy, spinal muscular atrophy | Trexo Robotics Plus, MyoSwiss MyoSuit, ABLE Human Motion |
A single device may straddle two categories. The Cyberdyne HAL platform, for example, exists in both a clinical version (HAL for Medical Use) and an industrial-labor version (HAL for Care Support and HAL Lumbar Type for caregivers).[8]
The earliest patented design for a body-worn assistance device dates to the 1890 Russian inventor Nicholas Yagn, who proposed a passive bag of compressed-gas springs to assist walking, running and jumping. Yagn's apparatus was never built and would not have functioned as drawn. Several other speculative patents appeared in the early 20th century, but no powered prototype was constructed until General Electric's Hardiman in the 1960s.
General Electric's Hardiman was the first serious attempt to build a powered exoskeleton. Funded by the U.S. Office of Naval Research and the U.S. Army, it was developed at GE's Schenectady research center between 1965 and 1971. Hardiman was a hydraulic and electric full-body suit intended to multiply the wearer's strength by a factor of 25, allowing a 110 kilogram lift to feel like 4.5 kilograms.[4] In practice the suit weighed roughly 680 kilograms, moved at about 0.76 meters per second, and was so unstable that engineers could only safely operate one arm at a time. When both legs were powered, GE's own engineers described the resulting motion as "violent and uncontrollable," and the project was shelved without a single human ever wearing the full suit.[4] Hardiman established that powered exoskeletons would require dramatic advances in sensors, actuators, batteries and control theory before they could ever leave the laboratory.
In 2000 the U.S. Defense Advanced Research Projects Agency (DARPA) launched the Exoskeletons for Human Performance Augmentation (EHPA) program, which funded a new generation of academic and corporate research. The most influential output was the Berkeley Lower Extremity Exoskeleton (BLEEX), built at the University of California, Berkeley by Homayoon Kazerooni's Robotics and Human Engineering Lab. Unveiled in 2004, BLEEX had seven degrees of freedom per leg, a hydraulic actuation system, and an internal-combustion engine that made it the first untethered, energy-autonomous lower-limb exoskeleton. It could carry a 75 kilogram payload at 0.9 meters per second.[9] BLEEX directly seeded several commercial efforts: Berkeley Bionics (later renamed Ekso Bionics) was founded in 2005, and Lockheed Martin licensed a derivative platform that became HULC.
In parallel with BLEEX, the Japanese roboticist Yoshiyuki Sankai at the University of Tsukuba had been developing the Hybrid Assistive Limb (HAL) since the late 1990s. Sankai began mapping motor neurons in human legs in 1990, designed early prototypes between 1997 and 2003, and founded Cyberdyne in June 2004 to commercialize HAL. HAL is distinguished by a unique control approach. Skin-mounted electrodes detect bioelectric signals leaking through the skin a few hundred milliseconds before the underlying muscles contract, and the suit provides matching mechanical assistance during the same movement, creating a tight closed loop between intent and motion that Sankai calls cybernic voluntary control.[10]
HAL received CE marking in Europe in 2013 and U.S. FDA 510(k) clearance in late 2017, with commercial U.S. deployment beginning at Brooks Rehabilitation in Jacksonville, Florida in March 2018. Cleared indications include rehabilitation for spinal cord injury, stroke, and certain progressive neuromuscular diseases.[10][11]
Between 2010 and 2018 a cohort of medical exoskeletons reached the market, mostly aimed at rehabilitation centers and a smaller number aimed at personal use by people with spinal cord injuries.
A notable structural feature of these devices was their reliance on crutches or other balance aids: none of them was self-balancing, which meant the wearer still had to manage trunk balance manually.
The French startup Wandercraft, founded in Paris in 2012, developed Atalante, the first self-balancing lower-limb exoskeleton that allows hands-free walking by patients with paralysis. Atalante uses model-based dynamic balance control inspired by humanoid bipedal robotics, eliminating the need for crutches.[15] Atalante X received its initial FDA clearance in early 2024 for spinal cord injury at levels T5 to L5, was cleared for hemiplegia from cerebrovascular accidents soon after, and in November 2025 the FDA expanded the cleared population to spinal cord injury from C4 to L5 and to multiple sclerosis.[16] In June 2025 Wandercraft closed a $75 million Series D round to fund global commercialization of Atalante and the rollout of Eve, a personal at-home version of the self-balancing exoskeleton planned for U.S. launch in 2026.[17] During the 2024 Paris Olympics torch relay, Wandercraft's personal exoskeleton allowed Kevin Piette, a wheelchair user, to walk a leg of the torch route, the first time a self-balancing exoskeleton had been used in such a public ceremonial setting.[15]
| Product | Manufacturer | First Cleared | Indication | Distinguishing Feature |
|---|---|---|---|---|
| ReWalk Personal | ReWalk Robotics | FDA 2014 | Spinal cord injury, T7 to L5 | First FDA personal-use clearance for a powered exoskeleton |
| Ekso GT | Ekso Bionics | FDA 2016 | Stroke and spinal cord injury | First clinical exoskeleton cleared for stroke rehabilitation |
| EksoNR | Ekso Bionics | FDA 2020 | Stroke, SCI, acquired brain injury, multiple sclerosis | Only device cleared for brain injury and MS |
| Indego Therapy and Personal | Parker Hannifin (later Ekso) | FDA 2016 (clinical and personal) | Spinal cord injury, T3 to L5 | Modular design weighs 12 kg, lightest in class at launch |
| HAL Lower Limb Type | Cyberdyne | FDA 2017 | SCI, stroke, neuromuscular disease | Bioelectric (sEMG) intent control |
| Atalante X | Wandercraft | FDA 2024, expanded 2025 | SCI C4 to L5, MS, hemiplegia | First FDA-cleared self-balancing exoskeleton |
| Trexo Plus | Trexo Robotics | Pediatric | Cerebral palsy, SMA | Pediatric gait training in clinic and home |
| MyoSuit | MyoSwiss (Switzerland) | CE | Generalized lower-limb weakness | Soft, lightweight, portable |
| Product | Manufacturer | Type | Body Region | Notable Adopters |
|---|---|---|---|---|
| Cray X | German Bionic | Active, powered | Lower back / lift assist | Logistics, automotive |
| Apogee and Apogee+ | German Bionic | Active, powered | Lower back / lift assist, plus posture coaching AI | Logistics, healthcare |
| MATE-XT and MATE-XT GO | Comau | Passive | Shoulders / overhead work | Stellantis, Iveco, Maserati |
| H-WEX (Waist EXoskeleton) | Hyundai | Passive | Lower back | Hyundai factories worldwide |
| H-VEX (Vest EXoskeleton) | Hyundai | Passive | Shoulders, overhead work | Hyundai assembly plants |
| CEX (Chairless EXoskeleton) | Hyundai | Passive | Legs and lower back, sit-stand | Hyundai standing-work stations |
| EksoVest and EVO | Ekso Bionics | Passive | Shoulders, overhead work | Ford, Toyota, Boeing |
| Levitate Airframe | Levitate Technologies | Passive | Shoulders, overhead work | BMW, Boeing, dental clinics |
| Paexo Shoulder, Back, Thumb | Ottobock | Passive | Shoulders, back, thumb | VW, Audi |
| Phoenix and ShieldX | SuitX (acquired by Ottobock 2021) | Passive and active variants | Lower limb, back, shoulders | Construction, automotive |
| Guardian XO | Sarcos / Palladyne AI | Active, full body | Whole body | Material handling pilots, Delta Air Lines, Sumitomo |
| Product | Developer | Status | Notes |
|---|---|---|---|
| Hardiman | General Electric | Canceled 1971 | First powered exoskeleton attempt; never operationally worn |
| BLEEX | UC Berkeley / DARPA | Research only | First untethered lower-limb exoskeleton (2004) |
| HULC (Human Universal Load Carrier) | Berkeley Bionics, licensed to Lockheed Martin | Discontinued | Carried 90 kg loads but increased metabolic cost in U.S. Army trials |
| ONYX | Lockheed Martin | Field tested | Lower-limb actuated knee assist; under 6 kg powered weight |
| TALOS (Tactical Assault Light Operator Suit) | U.S. Special Operations Command | Concluded 2019 | $80 million program; produced spinoff technologies but no fielded suit |
| Soft Exosuit | Harvard Biodesign Lab and DARPA Warrior Web | Research and licensing | Conor Walsh's textile-based hip and ankle exosuits |
| Elevate Ski Exoskeleton | Roam Robotics | Commercial | Pneumatic recreational exoskeleton for skiing |
| Ascend | Roam Robotics | Commercial | Knee assist for osteoarthritis and weak quadriceps |
| MO/GO | Skip Innovations and Arc'teryx | Commercial | Outdoor hiking exoskeleton pant launched in 2024 |
A modern exoskeleton's sensing stack typically includes some combination of the following.
The controller layer has evolved through three broad generations.
| AI Approach | Function | Representative Result |
|---|---|---|
| Convolutional neural networks (CNNs) on sEMG plus IMU | Motion classification, gait phase recognition | 99.3% classification accuracy across walking, running, stair climbing in ankle exoskeleton studies[5] |
| Bidirectional LSTM on sEMG | Online gait-phase estimation 120 ms ahead of mechanical motion | Robust stance and swing detection for hip and knee exoskeletons[18] |
| Deep CNN on IMU plus plantar pressure | Five-phase gait cycle recognition | Used in commercial soft ankle exoskeletons for cerebral palsy[19] |
| Model-free reinforcement learning (TD3, PPO) | End-to-end controller for walking, sit-to-stand, squat assistance | Squat assistance and crutch-force minimization for lower-limb exoskeletons[20] |
| Sim-to-real deep learning on musculoskeletal models | Versatile policy that generalizes across activities and users | Hip exoskeleton reduces metabolic cost by 24.3% (walking), 13.1% (running), 15.4% (stair climbing) on first deployment[6] |
| Recurrent musculoskeletal coupling | Robust torque control under uncertain interaction forces | Stable walking control under variable patient-suit force coupling[20] |
| Foundation-model-style transformer policies | Multimodal sensor fusion across morphologies | Active research direction inspired by RT-1, RT-2, and Octo for general manipulation robotics |
Intent prediction is the task of inferring what a wearer is about to do, ideally before they begin to do it, so that the suit can match assistance to motion rather than fight it. EMG signals lead the corresponding mechanical contraction by about 100 to 200 milliseconds, which is enough lead time to drive a motor and complete a coordinated assist. Convolutional and recurrent neural networks have been applied to this problem with steadily improving accuracy. A 2025 study published in Scientific Reports used three IMUs and eight textile-electrode sEMG sensors with a CNN to classify ankle exoskeleton motions with 99.263% accuracy, surpassing prior gel-electrode studies and improving long-term wearability.[5]
Deep learning has largely replaced threshold-based finite-state machines for gait phase detection. A 2024 study published in MDPI Sensors used a bidirectional LSTM on multi-muscle sEMG data during exoskeleton-assisted walking and identified stance and swing phases with high accuracy 120 ms in advance of the kinematic transition, exploiting the natural electromechanical delay of skeletal muscle to enable predictive rather than reactive control.[18]
Several groups have demonstrated that model-free reinforcement learning can produce robust exoskeleton controllers without requiring hand-engineered gait models. Algorithms such as Twin Delayed Deep Deterministic Policy Gradient (TD3) and Proximal Policy Optimization (PPO) are commonly used. A 2023 paper in the Journal of NeuroEngineering and Rehabilitation reported a deep RL controller for a lower-limb rehabilitation exoskeleton trained on a coupled musculoskeletal model and three independent neural networks; the controller delivered reliable walking assistance under uncertain interaction forces between user and suit.[20] A separate 2024 arXiv paper used RL to minimize crutch loading forces in a lower-limb medical exoskeleton, an explicit safety objective rarely captured by classical controllers.[21]
The 2024 Nature paper from Steve Collins's group at Stanford demonstrated that reinforcement learning over a high-fidelity musculoskeletal simulator could produce a single hip-exoskeleton controller that performed across walking, running and stair climbing without per-subject calibration, reducing measured metabolic cost on first deployment.[6] The result was widely cited as the first convincing demonstration of sim-to-real transfer for a wearable robot. A 2026 follow-up bioRxiv preprint reexamined the published metabolic results, arguing that some reported energy savings exceed physiological limits relating mechanical power to metabolic cost during gait, which has triggered an ongoing methodological discussion.[7]
At the time of writing in early 2026, several research groups are adapting transformer-based foundation-model architectures from general robotics, including ideas from Google DeepMind's RT-1 and RT-2 and the Open X-Embodiment Octo policy, to exoskeleton control. The motivation is the same as in manipulation robotics: a single policy pretrained across many users, sensors and devices may generalize better than per-device controllers that have to be tuned for each new wearer.
Automotive assembly lines have been the largest single industrial buyer of exoskeletons, primarily because so many assembly tasks involve repetitive overhead work, sustained awkward postures, or heavy lifting. Major automotive deployments include the following.
Boeing has used the EksoVest at the Everett, Washington plant for overhead 787 assembly work. Airbus has tested the Skelex 360-XFR, a passive shoulder support, at its Hamburg plant. Logistics operators including DHL, GXO Logistics and Iron Mountain have deployed German Bionic Cray X and Apogee active back-support systems for repetitive lifting in distribution centers.[23] Delta Air Lines and Sumitomo conducted pilot deployments of the Sarcos Guardian XO full-body exoskeleton for heavy material handling, although Sarcos paused commercialization of the Guardian XO in 2022 in favor of focusing on AI software for robotics under the renamed Palladyne AI brand.[24]
Construction adoption has been slower than automotive due to the unstructured nature of jobsites, but passive shoulder and back-support exoskeletons from SuitX, Ottobock and HeroWear are increasingly used by drywallers, electricians and heating ventilation and air-conditioning installers for sustained overhead work and material handling.
Military exoskeleton programs have repeatedly run into the metabolic-cost problem: a powered suit must add less metabolic burden than the load it is helping carry, otherwise the soldier is more tired wearing the suit than without it. The Lockheed Martin HULC, derived from BLEEX, was tested by the U.S. Army Natick Soldier Research, Development and Engineering Center in the early 2010s and showed increased metabolic cost compared to soldiers carrying equivalent loads with conventional rucksacks, leading to the project's termination.[4] The U.S. Special Operations Command's TALOS program, launched in 2013 by Admiral William McRaven and intended to produce a powered armored suit for special operators, was wound down in 2019 after spending roughly $80 million; it produced useful spinoffs in helmets, optics and ballistic materials but no operational exoskeleton.[4]
The Lockheed Martin ONYX, a partial lower-limb exoskeleton with knee actuators, AI-driven assistance modulation and a published unloaded weight under 6 kilograms, was tested with the U.S. Army's 10th Mountain Division and won a Popular Science Best of What's New award in 2018.[25] The U.S. Army Soldier Center has subsequently funded smaller, more modest exoskeleton programs aimed at specific tasks (logistics handling, evacuation litter carrying) rather than full powered armor for combat infantry.
A notable trend in military research is convergence with the commercial industrial market: rather than building bespoke combat suits, programs increasingly procure modified versions of devices originally designed for warehouse, construction or rehabilitation use.
A distinct branch of wearable robotics, the soft exosuit, was pioneered by Conor Walsh's lab at the Harvard Biodesign Lab in collaboration with the Wyss Institute and supported in part by the DARPA Warrior Web program. Rather than rigid frames, soft exosuits use textile webbing, Bowden cables and small motors at the waist or backpack to deliver assistive forces through the wearer's own skeleton.[26] Walsh's team has shown that soft hip exosuits can reduce the metabolic cost of walking and running, and a 2024 collaboration with Boston University demonstrated that a soft hip exosuit eliminated freezing-of-gait episodes for a participant with Parkinson's disease during indoor walking, a result published in Nature Medicine.[26][27] The technology has been licensed to multiple commercial spinouts, including Verve Motion (industrial back support) and Imago Rehab (stroke rehabilitation).
The exoskeleton industry has seen significant consolidation as it transitions from research curiosity to commercial product.
Market size estimates vary widely. As of 2025, analysts including Fortune Business Insights, MarketsAndMarkets, Grand View Research and Mordor Intelligence published global market values between roughly USD 0.6 billion and USD 3.4 billion, with consensus around USD 2 to 3 billion. Forecast compound annual growth rates over the 2025 to 2030 horizon range from roughly 19% to over 40%, reflecting both the early stage of the industry and uncertainty about which application segments will scale fastest. Most analysts expect medical and industrial applications to dominate near-term revenue, with augmentation, recreation and military as longer-term growth opportunities.[29]
| Analyst | 2025 Market Size | Forecast End Year | Forecast Value | CAGR |
|---|---|---|---|---|
| MarketsAndMarkets | USD 0.56 billion | 2030 | USD 2.03 billion | 19.2% |
| Fortune Business Insights | USD 3.52 billion (2026) | 2034 | USD 64.2 billion | 43.7% |
| Mordor Intelligence | USD 3.37 billion | 2030 | USD 13.5 billion | 32.1% |
| Straits Research | USD 3.44 billion | 2033 | USD 30.8 billion | 31.5% |
| FactMR | USD 2.1 billion | 2035 | USD 13.4 billion | 23.6% |
North America held the largest share in 2025, driven by FDA-cleared medical adoption and large-scale industrial pilots, while Asia Pacific (especially Japan, South Korea and China) is forecast to grow fastest.[29]
Despite progress, several challenges still constrain widespread deployment.
The exoskeleton field is at an inflection point comparable to where surgical robotics was in the early 2000s or where autonomous driving was in the mid-2010s. Hardware has matured to the point where reliable, FDA-cleared powered suits are commercially available; AI-driven control is rapidly closing the gap between research demonstrations and clinical or industrial usability; and a small but growing number of consumer-facing recreational and home-care products is opening the category to non-clinical buyers. Whether the industry will follow the path of surgical robotics, where one player (Intuitive Surgical) eventually captured most of the market, or the path of consumer wearables, where many specialized brands serve overlapping niches, is an open question that will likely be answered over the second half of the 2020s.