Exoskeletons — Industrial, Medical, Military

1. Scope and definition

An exoskeleton is a wearable mechatronic structure that operates in parallel with the human musculoskeletal system, applying torques or stored elastic energy at one or more joints to reduce metabolic cost, off-load muscle activation, restore lost function, or amplify load capacity. Unlike a prosthesis (which substitutes a missing limb) and unlike a manipulator (which the operator commands from outside), an exoskeleton co-locates with the wearer and shares sensorimotor authority — the human remains in the control loop and the device only ever assists.

The field straddles four communities that emerged largely independently and have only recently converged on shared component, control, and standards stacks: industrial / occupational (overhead work, lifting, posture support), medical / rehabilitation (gait training and ambulation in stroke, spinal-cord injury, multiple sclerosis), military / load-carriage (dismounted infantry, logistics handling), and a small but growing consumer / sport / mobility sector. This note covers all four with emphasis on real products, the mechatronic stack underneath, the evidence base, and the standards landscape.

Cross-reference field anchors: [[Robotics/legged-robotics]], [[Robotics/impedance-control]], [[Robotics/prosthetics]], [[Robotics/humanoid-balance]], [[Engineering/biomechanics]], [[Engineering/ergonomics-human-factors]], [[Engineering/bioinstrumentation]].

2. Classification axes

The taxonomy is multi-axis; a single device fixes a point in all five spaces.

2.1 Energy source

• Passive — no batteries, no actuators. A spring, gas strut, cam mechanism, or elastic band stores energy from gravity-loading motion (e.g. trunk flexion) and releases it back during the return phase, or pre-loaded by a fitting therapist. Mass 1.5–5 kg. Examples: Ottobock Paexo, Levitate AIRFRAME, Laevo V2.5.
• Quasi-passive — clutched springs plus low-power electronics. A solenoid or escapement switches an elastic element in and out of the load path based on gait phase. Net positive work ≈ 0 on level ground; only redirects already-existing energy. MIT Walsh & Herr 2007.
• Active — motors deliver net positive mechanical work. Battery (typically 600–1500 Wh) drives BLDC + planetary or SEA actuators through a torque controller. Mass 4–25 kg for limb devices; 60–90 kg for full whole-body augmentation.

2.2 Structure

• Rigid — articulated metal/composite linkages with discrete pin joints (with self-aligning compensation). Best load capacity; worst conformance; classic form.
• Soft (exosuit) — textile + cable architecture, no rigid frame. Bowden cables transmit motor torque to apparel-like attachment points. Light, conformal, but limited to tension forces; cannot resist compression or stabilise against fall. Originated Harvard Wyss (Walsh group), commercial via ReWalk Restore.
• Hybrid — rigid frame for load path, soft interface for skin contact (most rehab exos).

2.3 Anatomical coverage

• Joint-specific — single joint (knee, hip, ankle, elbow, lumbar). Smallest, lightest, target a specific task. Dominant in industrial (back, shoulder).
• Segment — upper limb only, lower limb only, trunk only.
• Full-body — Sarcos Guardian XO is the canonical example; full-body rehab exos like Wandercraft Atalante self-balance without crutches.

2.4 Use case

• Industrial / occupational — repetitive task assist, posture support, manual lifting; 8-hr shift duty cycle.
• Medical / rehabilitation — gait training, ambulation for non-ambulatory patients, post-stroke recovery, paediatric mobility.
• Military — load carriage on march, logistics handling at base/depot; very few combat applications fielded.
• Consumer / sport — mountaineering knee assist, golf-ball-retrieval (Honda Walk Assist), ski exos (Roam Robotics, now Arc’teryx MO/GO).

2.5 Control authority

• Fully autonomous gait — exo drives the limb, user passive (early SCI ambulation paradigm; Ekso GT in adaptive-assist mode 2014).
• Assist-as-needed — exo measures user effort and supplies only the deficit (Atalante adaptive mode; modern rehab default).
• Augmentation / amplification — user always drives, exo amplifies (Sarcos XO operator-leading control).
• Resist mode — strength training (occupational; some Lokomat protocols).

3. Industrial / occupational exoskeletons

Industrial musculoskeletal disorders (MSDs) cause about 30 percent of US worker-comp claims (Bureau of Labor Statistics 2022) — overhead work loads the rotator cuff and supraspinatus, repeated lifting loads the lumbar discs (L4-L5, L5-S1 intradiscal pressure). Industrial exos reduce muscle activation EMG by 30–60 percent on the targeted group in lab studies. Field adoption took off 2017–2020 with the Ottobock acquisition of SuitX (2021) and Hyundai Motor Group’s CEX/VEX programme.

3.1 Overhead and shoulder support

• Ottobock Paexo Shoulder (2018, Germany) — 1.9 kg, cable + cam-spring mechanism, gravity-compensation torque profile shaped to sin(α) shoulder flexion angle. Used by Volkswagen (assembly), Boeing (aircraft fit-out), Audi.
• SuitX ShoulderX (2017, now Ottobock) — 4.5 kg, cam-spring, three-position adjustability.
• Levitate AIRFRAME (2017, US) — 1.8 kg, cam mechanism with no battery; Boeing 2019 field study showed deltoid EMG dropping from 28 to 11 percent of maximum voluntary contraction at α = 90°.
• Comau MATE-XT (2018, Italy; IUVO/IIT spinout) — 2.8 kg, carbon-fibre passive spring.
• Ekso EVO / EVO-S (2020, Ekso Bionics; the S is the smaller-frame version) — 2.0 kg, single-spring vest-based.
• Hilti EXO-O1 (2022, co-developed with Ottobock) — purpose-built for construction overhead drilling; tool-belt integration.

3.2 Lumbar / back support

• Laevo V2.5 (Netherlands, 2017→2022) — passive chest-pad-and-thigh-pad mechanism, leaf springs, 2.6 kg. Most-deployed back exo in European industry.
• Ottobock Paexo Back (2019) — 3.0 kg, spring elements.
• SuitX BackX (2017, now Ottobock) — 3.5 kg, hip-pivot spring.
• Innophys Muscle Suit Every (Japan, 2014) — 3.8 kg, McKibben pneumatic actuator with manual hand-pump pre-pressurisation (no battery, no compressor). Sold ~30 000 units in Japan agriculture and elder care.
• German Bionic Cray X (2020) — active lumbar exo; lithium-ion 8 hr run; CAN-bus telemetry; “Apogee+” 2023 variant adds posture coaching.

3.3 Whole-body industrial augmentation

• Sarcos Guardian XO (2020, US; Sarcos → Palladyne AI in 2023) — 24 powered DOF, ~98 percent power-assist ratio at lift joints, 90 kg payload felt as ~1 kg. Hot-swap Li-ion packs in 2 hr blocks. Customers: Delta Air Lines (cargo), GM (assembly), US Navy shipyards. Lease pricing target ~USD 100 k/yr.
• Sarcos Guardian XT — upper-body teleoperated variant on a lift base; not strictly wearable; companion product.
• Hyundai CEX / VEX (2019, South Korea) — chairless / overhead-work assistive variants deployed at Hyundai and Kia assembly lines.

3.4 Sit-stand support

• Noonee Chairless Chair (Switzerland, 2014) — passive trunk-and-leg structure that converts into a “seat” when the wearer squats; locks for sustained semi-seated postures. Pilots at Audi (2015, A4/A5 final assembly), BMW, Daimler.
• Audi Chairless Chair pilots — Audi’s own evaluation in Ingolstadt and Neckarsulm plants 2015–2018; mixed outcome (queue-time and donning-time issues outweighed posture-support benefit for short tasks).

3.5 Industrial standards interlock

ISO 13482:2014 — personal-care robots — is the principal industrial-exo safety standard, covering pinch-points, restraint, EMC, and physical human-robot contact. ASTM F48 (Exoskeletons and Exosuits committee, formed 2017) is producing the granular sub-standards F48.01–F48.06 (terminology, ergonomics, test methods, performance, human factors). ANSI/RIA TR R15.806 covers North-American industrial deployment.

4. Medical / rehabilitation exoskeletons

4.1 Lower-limb rehab and personal ambulation

• Ekso NR (Neurological Rehabilitation) — current Ekso Bionics flagship clinical exo. Indicated for stroke, spinal-cord injury, MS, traumatic brain injury. Variable-assist mode meters torque to user effort. Replaces the earlier Ekso GT (FDA 2016, K161443). Used in over 300 rehabilitation hospitals worldwide.
• ReWalk Personal 6.0 — Lifeward (formerly ReWalk Robotics, name change 2024). First FDA-cleared personal-use exoskeleton (K131798, 2014). Four powered DOF (hip + knee bilateral); requires Lofstrand crutches for balance. Tilt-sensor intent: forward-lean triggers step. Indicated T7–L5 SCI. Major milestone December 2023: CMS announced Medicare coverage for personal exoskeletons after a decade of reimbursement gridlock.
• Cyberdyne HAL Lower Limb (Japan) — PMDA-cleared as medical device 2013; FDA-cleared 2017 (K162791). Cybernic Voluntary Control reads sEMG over rectus femoris, hamstrings, glutei, soleus, tibialis. Single-leg and double-leg variants. Rental business model in Japan (no outright sale of medical variant).
• Indego Personal / Indego Therapy — originally Vanderbilt Center for Intelligent Mechatronics under Michael Goldfarb; commercialised by Parker Hannifin; spun out as Indego LLC. FDA cleared 2016 (K171334). Modular design: separate hip, thigh, and shank modules that ship in a backpack. Indicated T3–L5 SCI for Personal; broader for Therapy.
• Wandercraft Atalante / Atalante X (France, 2019) — the first self-balancing commercial exoskeleton; the user does not need crutches because the exo’s twelve powered DOF maintain ZMP stability autonomously. Twelve DOF (hip 3 + knee 1 + ankle 2 per leg). Used at Mass General Brigham, NYU Langone, Hôpital Raymond-Poincaré.
• Rex Bionics REX (New Zealand, 2010) — also self-balancing without crutches, but slow (max ~0.1 m/s) and very heavy (~38 kg). Niche use for SCI assessment and demos.
• Hocoma Lokomat Pro (Switzerland, 2001) — treadmill-based with body-weight support harness; bilateral 2-DOF orthoses (hip + knee, ankle passive). Hip peak 70 N·m, knee peak 40 N·m. The “guidance force” knob tapers 100 → 0 percent across a treatment programme. Augmented Performance Feedback module visualises real-time joint tracking error. Owned by DIH Medical (NASDAQ: DHAI).
• Hocoma G-EO System (Reha Technology, 2010) — end-effector gait trainer with stair-mode; foot platforms drive distal kinematics rather than full-leg orthosis.
• ABLE Human Motion ABLE Exo (Spain, 2022, CE-marked) — modular lower-limb exo for SCI ambulation training; founder spin-out from UPC Barcelona.
• SuitX Phoenix (2016) — modular low-cost SCI exo; CE 2017; design intent USD 30 k vs the USD 80–150 k incumbent price.

4.2 Upper-limb rehab and assistive

• Myomo MyoPro (US, 2012) — orthosis with sEMG-triggered elbow + hand assist. Reads bicep/tricep surface EMG and amplifies user-driven motion. Indicated for stroke, brachial plexus injury, spinal-cord injury affecting arm. Outpatient prescription pathway; CMS HCPCS L8702 code.
• Hocoma ArmeoPower (2011) — shoulder + elbow + wrist clinical exo with gravity compensation and game-based therapy software.
• Saebo MAS (2010) — passive arm gravity-compensation device for stroke training.
• Harmonic Bionics Harmony SHR (2020) — bilateral shoulder rehabilitation exo; out of UT Austin Ashish Deshpande lab.
• Bioservo Carbonhand (Sweden, 2016) — fingertip-pressure-triggered grip glove; FDA-cleared and CE-marked.
• Neofect RAPAEL Smart Glove — sensor glove (more telerehab than exo, included for completeness).

4.3 Soft exosuits

• Harvard Wyss / Walsh exosuit lineage — research originating with the Asbeck/Walsh soft-exosuit work showing 14 percent reduction in metabolic cost of walking (Science Robotics, 2017; also Asbeck et al. IJRR 2015). Spawned ReWalk Restore.
• ReWalk Restore (US, 2019, FDA-cleared) — soft cable-driven exosuit for stroke gait. Single-sided device (paretic leg); Bowden cables drive ankle dorsiflexion and plate-flexion via textile interface. Mass ~5 kg.
• Roam Robotics → Arc’teryx MO/GO (2024) — pneumatic soft exo for hiking knee support, productised through Arc’teryx outdoor brand.

4.4 Paediatric

Paediatric exos must accommodate growth and tighten fit tolerances; therapy windows are also developmental, with neuroplasticity arguments for early intervention.

• Trexo Robotics (Canada, 2017) — paediatric gait-training exo mounted on a rolling walker frame; intended for ages 2–18 with cerebral palsy, SCI, or genetic neuromuscular conditions. Family use at home plus clinic licensing.
• Atlas 2030 / ATLAS 2030 (Marsi Bionics, Spain) — the first paediatric ambulatory exoskeleton; designed for spinal muscular atrophy (SMA) children aged 4–10; CE-marked. Founder Elena García Armada (CSIC Spain).

4.5 Clinical evidence base

Evidence quality varies by indication. Cochrane review (Mehrholz 2017) on electromechanical-assisted training after stroke pooled 36 RCTs and concluded that combined exo + over-ground physiotherapy increases independent walking probability versus over-ground alone, particularly within three months of stroke; effect on speed and endurance is smaller and less consistent. For SCI ambulation, evidence is dominated by single-arm cohort studies (Esquenazi 2012; Spungen 2013) showing feasibility and modest cardiovascular benefit; functional-independence-measure improvement is harder to demonstrate. The Lokomat-vs-overground training debate continues — meta-analyses (Mehrholz 2013, 2017) marginally favour electromechanical training for non-ambulatory patients and over-ground for already-ambulatory ones. MS use cases (Ekso, Wandercraft) are growing but RCTs are scarce.

5. Military and first-responder

Dismounted infantry routinely carry 30–60 kg of pack, weapon, armour, water, ammunition, and battery; logistics handlers in depot routinely lift 20–40 kg cartons repetitively. Military exo programmes have not generally validated as combat systems but several have transitioned to logistics and industrial roles.

• Berkeley BLEEX (Kazerooni, UC Berkeley, 2004) — first full lower-body powered military exo prototype; hydraulic actuators; led to Lockheed’s HULC.
• Lockheed Martin HULC (Human Universal Load Carrier) — productionised BLEEX; Army field trial 2009–2012; cancelled 2012 after metabolic-cost field measurements showed worse economy than unaided walking with the same load — the device’s own ~14 kg mass plus inefficient hip/ankle dynamics negated the load-transfer benefit.
• Lockheed Martin ONYX (2018; not the same as the BAE Onyx) — passive-mostly knee assist for repeated stair climbing under load; uses electromechanical clutched-spring at knee. Targeted firefighter and infantry stair-and-incline scenarios.
• Lockheed Martin FORTIS — passive arm-tool-support exo; pivoted from military to industrial / shipyard (Huntington Ingalls, Newport News).
• Sarcos Guardian XO — partnered with US Army TACOM (Tank-Automotive Command) on logistics depot evaluations; sustainment-mission-focused rather than combat.
• Mawashi Uprise (Canada, 2014, with DRDC Defence Research and Development Canada) — passive load-bearing exo that transfers backpack load to ground; structural skeleton parallels the wearer’s spine and legs.
• Russian Ratnik-3 — the powered exoskeleton component of Russia’s “future soldier” infantry kit; status uncertain; reports of prototype trials but no confirmed fielded device.
• SOCOM TALOS (Tactical Assault Light Operator Suit, 2013–2019) — ambitious full powered-armour programme; ended 2019 with a “lessons learned” report. Power-density and integration challenges drove the cancellation; subsystems (sensors, ballistic plates) folded into other programmes.

The military lesson learned across BLEEX/HULC/TALOS: passive load-transfer works, active full-body augmentation does not yet pay for its own mass and battery in over-ground walking. Logistics and base/depot applications remain the realistic near-term military market.

6. Components, actuation, and transmission

6.1 Actuator topologies

• Series Elastic Actuator (SEA) — motor + reducer in series with a calibrated spring; force read by spring deflection. Pratt & Williamson, IROS 1995, MIT Leg Lab. Low impedance for safe contact; cheap force sensing. Used in Ekso, Indego, BLEEX, MIT Knee. See [[Robotics/impedance-control]] for the underlying theory.
• Harmonic Drive — Harmonic Drive LLC (founded 1964 USM/Hasegawa Gear) compact strain-wave gear, near-zero backlash, ratio 50:1 to 320:1 in one stage; staple of robotic exos and prosthetics.
• BLDC motor + planetary gearbox — typical 100–300 W frameless or housed motors with 1:10 to 1:100 planetary; lighter and cheaper than HD but with backlash. Common in industrial active exos.
• Pneumatic Artificial Muscles (PAM) / McKibben muscles — braided sleeve over inflatable bladder; contracts when pressurised. Festo Fluidic Muscle (DMSP series) is the commercial reference. Innophys Muscle Suit Every uses this principle with hand-pumped pre-pressurisation. Compliant by construction but slow and difficult to control precisely.
• Hydraulic — Ekso’s earliest BLEEX-derived prototypes used hydraulic actuation; abandoned because of mass, leak, and noise penalties. Sarcos Guardian XO retains an electric-hydraulic hybrid with proprietary micro-pumps.
• Cable-driven (Bowden) — Harvard Wyss soft exosuits route motor torque through Bowden cables to remote textile attachment points. Motor mass can sit on the trunk; only thin cables travel down the leg. Limited to tension only.
• Shape Memory Alloy (SMA) and chemical actuation — experimental; not commercial. Slow response and low cyclic life.

See [[Robotics/motors-electric]] for motor selection, [[Engineering/electric-motors]] for BLDC fundamentals, and [[Engineering/gears-power-transmission]] for transmission design.

6.2 Sensors

• IMU (6-DOF or 9-DOF) at thigh, shank, foot — gait phase via finite-state machine or HMM/LSTM. Common parts: Bosch BMI260, InvenSense ICM-20948, STM LSM6DSV. See [[Robotics/sensors-pose-motion]].
• Joint encoders — absolute magnetic (AS5048, RLS Aksim) at each motor axis; resolution to ~14 bit.
• Surface EMG — bipolar Ag/AgCl skin electrodes; signal 10 μV to 5 mV, bandwidth 20–500 Hz. Cyberdyne HAL is the canonical EMG-driven exo. Sweat and electrode migration are the chronic field issues. See [[Engineering/bioinstrumentation]].
• FSR / capacitive insole pressure — Tekscan F-Scan, Pedar, Moticon — 4–99 cells, detect heel-strike and toe-off and centre-of-pressure.
• Force / torque sensors at human-suit interface — strain-gauge or piezo; used in active industrial and in Guardian XO admittance control.
• Depth camera — Atalante uses a depth camera for stair detection; some research platforms use Intel RealSense or Orbbec for terrain-aware control.

6.3 Power

• Lithium-ion packs dominate; 600–1500 Wh for active lower-limb rehab and industrial; 2 kWh+ for Guardian XO. Operational time 2–4 hr (active rehab) up to 8 hr (industrial active with low-duty assist).
• Hot-swap packs are essential for shift work (Guardian XO, Cray X).
• Tethered power is acceptable and common for stationary rehab use (Lokomat is permanently powered from mains).
• Regenerative energy harvesting at the knee — Donelan et al. (2008) demonstrated 5 W average harvesting during walking, but commercial exos generally do not include it; the added mass and complexity rarely pays back the recovered energy.
• Power-electronics topology: standard BLDC FOC drives with 24–48 V DC bus; see [[Engineering/power-electronics]].

6.4 Anthropometry and fit

Exos must accommodate a range of users — typically 5th to 95th percentile by stature, with parametric adjustment of thigh and shank segment lengths, hip width, and torso girth. Joint axis misalignment ε produces parasitic motion ε·θ at the interface, creating shear and skin pressure. Knee is the worst case because the anatomical knee is polycentric (instant centre of rotation migrates posteriorly during flexion via the four-bar cruciate ligaments) — a pure pin joint will mis-track by 10–20 mm through full range. Wandercraft Atalante, Hocoma Lokomat, and ABLE use linkage-based self-aligning knees (four-bar or six-bar polycentric joints) to track the migration.

Donning/doffing time is a major adoption barrier:

• Industrial passive: 30 s to 2 min, single user
• Industrial active: 2–5 min, single user with practice
• Rehab clinical: 15–30 min, two-person team for SCI patient
• Sarcos Guardian XO: ~5 min after training

7. Control architecture

7.1 Three-level hierarchy

The standard exoskeleton control architecture is layered:

[High-level: intent / mode selection]
    EMG decoding | EEG/BCI | tilt-sensor | mode buttons | terrain classifier
        |
[Mid-level: trajectory / impedance reference generation]
    gait-phase FSM | reference torque from clinical database | admittance ref
        |
[Low-level: joint torque or position control]   @ 1–4 kHz
    SEA torque loop | impedance controller | position PID
        |
[Motor commutation: FOC]                         @ 20–40 kHz

7.2 Low-level: torque vs position vs impedance

Pure position control is rigid and unsafe for human contact — any disturbance is rejected as if it were error. Pure torque control (open-loop) is safe but cannot enforce trajectory. Impedance control (Hogan, MIT, 1985) defines a virtual mass-spring-damper relationship between position error and applied force:

$\tau =K_p({\theta }_{\text{ref}}-\theta )+K_d({\dot{\theta }}_{\text{ref}}-\dot{\theta })+{\tau }_{\text{ff}}$

with low stiffness $K_p$ for compliant assist-as-needed and high stiffness for full-guidance rehab. Admittance control is the inverse causal relation — measured force commands position. SEAs (Series Elastic Actuators) make both natural to implement because the spring inherently provides a force/position mapping. See [[Robotics/impedance-control]].

7.3 Mid-level: gait phase estimation

Finite-state machine over stance ↔ swing, typically subdivided into eight phases:

1. heel-strike
2. loading-response (foot-flat)
3. mid-stance
4. terminal-stance (heel-off)
5. pre-swing (toe-off)
6. initial-swing
7. mid-swing
8. terminal-swing

Transitions triggered by ground-reaction-force thresholds from insole FSRs and/or IMU-derived shank angular-velocity zero-crossings. Phase determines which reference torque profile (from a clinical database, often normalised to percent of gait cycle) the exo follows. Modern systems use HMMs or LSTMs over IMU and FSR features for noise-robust phase estimation; see [[Math/time-series-and-hmm]].

7.4 High-level: intent inference

Modality	Latency	Robustness	Used by
EMG (surface)	50–100 ms pre-motion	sweat/placement sensitive	Cyberdyne HAL, Myomo MyoPro
IMU + phase detection	100–200 ms	very robust	Ekso, Indego, Atalante
Force at interface	10–30 ms	misreads if loose	industrial active, Sarcos XO
Pressure insole	50–80 ms	excellent for gait timing	Lokomat, ReWalk
EEG / BCI	200–500 ms	research only	CYBATHLON pilot work
Tilt-sensor lean	100–200 ms	simple, robust	ReWalk Personal

7.5 Self-balancing control (Wandercraft Atalante)

The first commercial self-balancing exo. Twelve powered DOF; Zero Moment Point (ZMP) maintained by hip and ankle torque modulation, similar to humanoid balance control (see [[Robotics/humanoid-balance]] and [[Robotics/legged-robotics]]). The user does not need crutches. The control reduces a 12-DOF underactuated system through hybrid zero dynamics (Westervelt/Grizzle framework, Michigan). Foot placement and torso pitch are the primary balance regulators; perturbation rejection is handled by a foot-placement adjustment policy.

7.6 RL- and ML-trained controllers

Recent trend (2023–): RL controllers trained in physics simulators (Isaac Gym, MuJoCo) and transferred to hardware. USC’s Sangbae Kim collaboration with Samsung, the Steve Collins ankle exoskeleton work at CMU, and Ottobock’s “Genium” knee adaptation pipeline are landmarks. Human-in-the-loop optimisation (Zhang et al., Science 2017; Slade et al., Nature 2022) tunes assist profiles by directly measuring metabolic cost during walking and gradient-descending the parameters — converging on a per-user optimum in ~20 minutes. See [[Robotics/rl-for-control]] and [[Math/reinforcement-learning-theory]].

7.7 Admittance vs impedance control trade-offs

The choice between impedance and admittance causality has practical safety and performance consequences:

• Impedance control (force-as-output, motion-as-input) — robust on stiff environments (rigid walls, tight cuffs, dense tissue), unstable on very soft environments because small motion errors translate to large force errors. Natural for SEAs.
• Admittance control (motion-as-output, force-as-input) — robust on soft environments (loose textile cuffs, foam padding) and stiff actuators (hydraulic, big BLDC + harmonic drive), unstable against rigid contact because measured force is essentially impulse-like at contact transitions. Natural for Sarcos Guardian XO’s high-stiffness hydraulic stack.

Hybrid switched controllers detect contact state and switch causality (Ott et al., 2010) — the underlying instability theorem (Colgate, 1988) shows the two are causal duals: stable in mutually exclusive environment-stiffness regimes.

Tuning bandwidth rule of thumb: closed-loop force/position bandwidth should be ≥3× the human voluntary movement bandwidth (~3 Hz at the joint, so ≥10 Hz closed loop). Industrial passive exos have effectively infinite bandwidth (no loop), which is why they are the safest class. Active exos with controllers below ~5 Hz feel “laggy” and provoke user fight.

8. Metabolic cost and outcome evidence

The honest performance question — does the exo make walking easier? — is measured by indirect calorimetry (VO₂, energy expenditure per metre walked). Key results:

• Asbeck/Walsh soft exosuit (Science Robotics, 2017) — 14 percent reduction in metabolic cost of walking at 1.5 m/s on level ground; the first published result clearly above the “cost of carrying the device” threshold.
• Steve Collins ankle exoskeleton (CMU, Nature 2015) — ~7 percent metabolic reduction with an unpowered clutched-spring ankle exo (the result that proved quasi-passive devices can pay back their own mass).
• Slade et al. (Nature, 2022) — human-in-the-loop optimisation of an ankle exosuit; ~17 percent metabolic reduction over walking without device; published from the Collins / Stanford lab.
• Awad/Galiana et al. (Science Translational Medicine, 2017) — post-stroke gait improvement with Wyss soft exosuit: 23 percent symmetry improvement on paretic side in selected chronic-stroke participants.
• HULC field study (US Army, 2010) — negative result, the device increased metabolic cost on most subjects, leading to programme cancellation. A cautionary baseline against vendor lab data.

Clinical functional outcome (Barthel Index, FIM, 10-Metre Walk Test) shows weaker effect sizes than metabolic cost; the field’s challenge is translating physical-economy gains to patient-relevant outcomes.

8.1 Measurement methodology

Metabolic cost is measured by indirect calorimetry: a portable mask (Cosmed K5, Cortex MetaMax) samples expired O₂ and CO₂ and applies the Brockway equation:

${\dot{E}}_{\text{met}}=16.58{\dot{V}}_{O_2}+4.51{\dot{V}}_{C{O}_2}[\text{W}]$

where the volumes are in mL/s. Steady-state requires 3–5 minutes per walking condition; within-subject comparison (with/without exo) typically uses 6–8 minute bouts in a randomised order with rest. Statistical power requires roughly N = 10 subjects to detect a 7 percent difference at α = 0.05; N = 6 for 14 percent. Field protocols (Slade 2022) use the same indirect-calorimetry rig but in over-ground walking with GPS-tagged terrain.

8.2 Why metabolic cost matters more than peak torque

A common vendor pitfall is reporting peak assistive torque (Newton-metres at the joint) and using that as the figure of merit. Metabolic cost is the integrated effect: it incorporates the device’s own mass, its negative work (resistance during the unassisted half-cycle), control-induced gait alterations, and inefficient inter-segment energy transfer. HULC’s negative metabolic result with respectable peak torque is the classic warning case. Newer protocols treat peak torque as a design parameter and metabolic cost as the acceptance criterion.

9. Standards, regulation, and safety

9.1 Standards landscape

• ISO 13482:2014 — personal-care robots; the umbrella standard for physical-assistant exoskeletons. Covers mechanical integrity, restraint, EMC, and HRC.
• ASTM F48 — Exoskeletons and Exosuits committee, US-led, formed 2017. F48.01 terminology, F48.02 testing (gait, lift, fall), F48.03 design (anthropometric ranges), F48.04 ergonomics, F48.05 task performance, F48.06 cybersecurity. The granular standards landscape that the field actually relies on.
• ANSI/RIA TR R15.806 — North American industrial exo deployment guidance.
• IEC 60601-1 — general safety of medical electrical equipment; applies to medical exos.
• IEC 60601-1-2 — EMC for medical electrical equipment.
• IEC 62366-1 — usability engineering for medical devices.
• IEC 62304 — medical-device software lifecycle (IEC 62304 + ISO 14971 risk management).
• IEC 62443 — industrial cyber for connected systems (Cray X, HAL telemetry).
• ISO 14971 — risk management for medical devices.
• See also [[Robotics/safety-standards]].

9.2 Regulatory pathways

• FDA 510(k) — substantially-equivalent pathway used by Ekso GT (K161443), Indego (K171334), HAL (K162791), MyoPro, ReWalk Restore (K183300).
• FDA De Novo — ReWalk Personal 6.0 (K131798, 2014) — first-of-kind personal exoskeleton classification.
• EU CE under MDR (formerly MDD) — Atalante, Phoenix, ABLE Exo, ATLAS 2030 all CE-marked.
• Japan PMDA — HAL was world’s first medical-cleared exoskeleton (2013).

9.3 Safety failure modes

• Falls during use — gait-rehab exos require harness + spotter (Lokomat is permanently overhead-harnessed). Personal-use exos (ReWalk, Indego) require crutches and trained user. Wandercraft Atalante’s self-balancing removes crutches but is constrained to clinic-controlled environments.
• Pinch points — at hip pivot and behind knee; particular concern in textile interface routing.
• Pressure sores — sustained interface load at iliac crest, sternum, underarm. Industrial exos enforce session duration limits (typically max 2 hr continuous).
• Joint misalignment shear — ε > 10 mm causes skin shear; >25 mm causes breakdown over hours.
• Thermal limits on actuators — high-torque BLDC at the hip can reach >70 °C casing under sustained load; thermal limit triggers torque ramp-down.
• Emergency stop — both physical (lanyard) and software (button on controller and on remote). Should fail safe to free-swing (zero impedance), not freeze (which can pin the user in a fall).
• Battery safety — Li-ion thermal runaway; pack design under UN 38.3 + IEC 62133. Hot-swap reduces risk by removing the need for in-suit fast charging.
• Cybersecurity — connected exos (Cray X, HAL telemetry, ReWalk app) are subject to FDA premarket cyber requirements (post-2023 guidance) and IEC 62443. Compromised telemetry could leak protected health information; compromised firmware could in principle drive unsafe motion.
• Phantom resistance — passive industrial exos engineered for a specific working range; outside that range the spring fights the user.

10. Cost and reimbursement

Category	Unit price (USD, 2024)	Reimbursement
Industrial passive	3 000 – 6 000	Employer-paid; some workers’ comp
Industrial active (Cray X, Hyundai)	20 000 – 30 000	Employer-paid
Rehab clinical (Ekso NR, Indego, Lokomat)	80 000 – 150 000 install	Clinic capex; reimbursed indirectly via PT codes
Rehab personal (ReWalk, Indego Personal)	80 000 – 150 000 unit	CMS coverage from Dec 2023 for select SCI; VA covers ReWalk for select SCI vets; private insurance variable
Self-balancing (Atalante)	150 000 – 250 000	Clinic capex
Whole-body (Sarcos Guardian XO)	~100 000 / yr lease	Enterprise pay
Upper-limb assistive (MyoPro)	30 000 – 60 000	HCPCS L8702; growing private payer coverage
Paediatric (Trexo, ATLAS 2030)	30 000 – 80 000	Clinic and family-share funding

The December 2023 CMS Medicare coverage decision for personal exoskeletons was the long-pending watershed; subsequent state Medicaid programmes followed during 2024.

11. Players and ecosystem

11.1 Public companies

• Ekso Bionics (NASDAQ: EKSO) — industrial (EVO) + medical (NR); merged with Parker’s Indego portfolio commercial activity 2022.
• Lifeward (formerly ReWalk Robotics, name change 2024; NASDAQ: LFWD) — Personal 6.0, Restore, ReStore Pediatric.
• Cyberdyne (TSE 7779) — HAL family; Japan-listed; rental-based clinical business.
• DIH Medical (NASDAQ: DHAI) — owns Hocoma (Lokomat, ArmeoPower) since 2019.
• Palladyne AI (NASDAQ: PDYN) — parent of Sarcos / Guardian XO since 2023 rebrand.
• Ottobock SE & Co (Germany, IPO planned) — acquired SuitX 2021; Paexo family; broader prosthetics business.
• Parker Hannifin (NYSE: PH) — Indego, now spun out as Indego LLC.

11.2 Private and research-driven

• Wandercraft (France, 2012) — Atalante; consumer “Personal Walking Device” announced 2024.
• German Bionic (Germany, 2017) — Cray X; industrial active back.
• Innophys (Japan, 2013) — Muscle Suit Every.
• Comau (Italy, owned by Stellantis) — MATE-XT via IUVO/IIT spinout.
• Marsi Bionics (Spain) — ATLAS 2030 paediatric.
• Trexo Robotics (Canada, 2017) — paediatric.
• ABLE Human Motion (Spain, 2018) — UPC Barcelona spin-out.
• Myomo (US, 2006) — MyoPro upper-limb.
• Bioservo Technologies (Sweden, 2010) — Carbonhand grip glove.
• Levitate Technologies (US, 2014) — AIRFRAME shoulder.
• Laevo (Netherlands, 2013) — V2.5 back exo.
• Noonee (Switzerland, 2014) — Chairless Chair.

11.3 Research labs

• Harvard Wyss Institute (Conor Walsh group) — soft exosuits; Wyss → ReWalk Restore commercialisation pipeline.
• Carnegie Mellon (Steve Collins, now Stanford) — ankle exo metabolic studies; human-in-the-loop optimisation; Nature 2015 and 2022 papers.
• MIT Biomechatronics (Hugh Herr) — quasi-passive lower-limb, BiOM ankle.
• MIT Leg Lab (Pratt origin) — series elastic actuators 1995.
• EPFL LSRO (Lausanne) — TWIICE personal exo, MyoSuit.
• IIT Genoa (Italian Institute of Technology) — IUVO spinout, Comau MATE collaboration.
• Vanderbilt CIRT (Michael Goldfarb) — Indego origin lab.
• KAIST (Daejeon) — WalkON Suit CYBATHLON winner 2016, 2020, 2024.
• University of Tsukuba (Yoshiyuki Sankai) — HAL origin.
• UC Berkeley Hel Lab (Homayoon Kazerooni) — BLEEX, HULC, SuitX origin.

11.4 CYBATHLON

The CYBATHLON, hosted by ETH Zurich every four years starting 2016, runs an exoskeleton race (powered lower-limb ambulation) and a powered upper-limb event among other disciplines. The exo race has been won by KAIST (WalkON Suit) three times — 2016, 2020, 2024. Acts as both showcase and benchmark for personal-mobility-exo state of the art.

11.5 Market sizing and forecasts

Global market estimates vary by analyst methodology but converge on a roughly USD 0.6–1.0 B base in 2022 and USD 5–8 B by 2030 (Grand View Research 2024, Markets & Markets 2024, Fortune Business Insights 2023). The medical segment grows fastest in revenue terms, driven by the CMS coverage decision and ageing demographics; industrial grows fastest in unit terms, driven by sub-USD-10k passive devices in warehouse and last-mile-delivery operations. Japan is the most penetrated medical market per capita due to early Cyberdyne PMDA clearance and elder-care policy support; Germany and the Netherlands lead industrial adoption in Europe; the US leads in clinical rehab installs but lagged in personal-use reimbursement until 2023.

11.6 Supply chain and key suppliers

The exoskeleton supply chain mirrors that of robotics broadly:

• Actuators — Maxon Motor (Switzerland, BLDC frameless), Allied Motion / KDE, Moog (servo), Harmonic Drive LLC (strain-wave gears), Nabtesco (cycloidal RV), SCHAEFFLER (planetary).
• Sensors — Bosch Sensortec (IMU), STMicroelectronics (IMU), Honeywell (load cells), AS5048 / RLS Aksim (encoders), Tekscan / Pedar (insole pressure), Cometa / Delsys (sEMG).
• Power — Samsung SDI, LG Energy Solution, A123 Systems (Li-ion cells); Texas Instruments and Analog Devices (BMS chips).
• Compute — STM32 family (low-level control), NXP i.MX (mid-level), NVIDIA Jetson Orin Nano/Nano Super (vision and perception in research exos), ESP32-S3 (BLE telemetry).
• Composites and structures — Toray and Hexcel (CFRP for limb segments), Solvay / Cytec (specialty resins), in-house CNC of 7075 / Ti-6Al-4V for joint housings.
• Textiles and interfaces — BOA (closure dials), Fillauer / Coyote (medical-grade silicones), Cordura (high-abrasion nylon for industrial straps).

See [[Robotics/motors-electric]], [[Robotics/sensors-pose-motion]], [[Engineering/materials-composites]], [[Robotics/power-systems]] for sub-system depth.

12. Trends and outlook (2024–2026)

• Soft suits supplanting rigid for occupational and low-demand mobility — lighter, conformal, lower donning time. Rigid still dominates rehab and load augmentation where load paths must bypass user joints.
• Single-joint targeted devices over full-body — clinical evidence increasingly favours targeted, modular interventions (ankle-only, knee-only) over full-leg orthoses; matches the metabolic-cost evidence (single-joint ankle exos already exceed soft full-leg suits in some studies).
• AI / RL-trained controllers — Isaac Gym sim-to-real pipelines; human-in-the-loop metabolic optimisation; per-user adaptation in 20-minute calibration sessions.
• Reimbursement unlock — December 2023 CMS personal-exo coverage is the inflection point. Watch private payer follow-through and state Medicaid expansion through 2025–2026.
• Gig-worker focus — last-mile delivery (Amazon DSP, FedEx, UPS), warehouse pick-and-stow (Amazon, Walmart), and food-service back-of-house are testing single-task passive exos at modest unit prices.
• Consumer / sport leakage — Roam/Arc’teryx MO/GO hiking exo; Honda Walk Assist; mountaineering knee exos. Validates that consumer-priced passive exos can find a market without insurance reimbursement.
• Standards maturation — ASTM F48 standards filling in through 2025; harmonisation with ISO 13482 revision under way.
• Self-balancing personal-use — Wandercraft announced a personal walking device in 2024 derived from Atalante stability stack; the goal is removing crutches from personal SCI ambulation.

12.1 Open problems

Despite three decades of development, the field has several unsolved problems where commercial-grade solutions remain elusive:

• Stair and uneven terrain in personal-use exos. Atalante self-balances on level ground; stair traversal in personal-use (non-clinic) settings remains unreliable. Foot placement under sub-perfect surface conditions (carpet edges, thresholds, snow, gravel) breaks ZMP balance assumptions. Vision-driven terrain classifiers (Atalante uses depth cameras) help but do not yet match human reactive foot placement.
• EMG drift across sessions. Surface EMG electrode placement, skin impedance, and sweat-induced drift make day-to-day calibration unstable. Re-calibration takes 5–15 minutes per session, which is a major adoption barrier for HAL-class devices. Implantable EMG or pattern-classifier auto-recalibration is research-stage.
• Falls under perturbation in self-balancing exos. Recovery from a strong shove or trip uses humanoid balance methods (capture point, foot placement modulation, hip-strategy push recovery) but is currently dominated by safety harnesses in clinic. Truly unsupervised SCI ambulation in the wild needs perturbation rejection robust to multi-step disturbances.
• Long-duration interface tolerance. Pressure-sore-free wear beyond 4–6 hours is still beyond the state of the art for active rehab exos. Industrial passive exos do better but typically still impose 2-hour session limits.
• Cost-to-mass-produce ratio. Sub-USD-10k active exos at industrial volumes remain elusive; the bill of materials (motors, harmonic drives, IMUs, batteries, structural composites) bottoms out at roughly USD 4–6k even at scale. Soft exosuit BOMs are lower (cables, motors, textile) but with reduced functional scope.
• Standardised outcome metrics. Industrial: EMG reduction is the dominant metric but does not capture musculoskeletal-injury reduction over years. Medical: 10-Metre Walk Test and 6-Minute Walk Test are routine but do not capture community ambulation, fatigue, or quality-of-life. ASTM F48 is filling part of this gap.

13. Worked example: passive shoulder exo design

Design problem: passive gravity-compensation shoulder exo to assist overhead drilling at α = 90° elevation. Effective arm + tool mass m = 5 kg at r = 0.32 m from shoulder.

Step 1 — static moment. Gravity moment at the shoulder: ${\tau }_g(\alpha )=mgr\sin \alpha$ At α = 90°: τ_g = 5 × 9.81 × 0.32 × 1 = 15.7 N·m. (Note: a more careful biomechanics calculation includes the arm’s own mass distribution; see [[Engineering/biomechanics]].)

Step 2 — target PAR. Target power-assist ratio 60 percent (passive limit before user “fights”). Required exo moment: τ_e = 0.60 × 15.7 = 9.4 N·m. Residual user moment 6.3 N·m.

Step 3 — spring-cam profile. A linear extension spring with constant k provides force F(x) = k·x at lever arm L(α). To shape the resulting moment τ_e(α) = F · L(α) to match sin(α), use a cam with profile L(α) shaped so that k·x(α)·L(α) ≈ τ_e,target · sin(α). Cam machining tolerances and pre-load setting let one mechanism cover the 30–130° working range.

Step 4 — verification. EMG-on-deltoid measurement is the standard validation; reductions of 15–20 percentage points of MVC at α = 90° indicate 50–60 percent effective PAR.

14. Worked example: gait-phase FSM for a lower-limb rehab exo

Inputs: heel-strike (HS) from insole FSR threshold (200 N rising), toe-off (TO) from FSR falling below 50 N, shank angular velocity ω_s from IMU.

States and transitions:

• S1 loading-response ← HS event; duration ~100 ms; target torque profile: extensor-dominant ramp.
• S2 mid-stance ← ω_s crosses zero from negative; duration ~250 ms; target: stiff knee, slight hip extension.
• S3 terminal-stance ← FSR centre-of-pressure shifts forward of forefoot; duration ~150 ms; target: push-off plantar-flexion if ankle powered.
• S4 pre-swing ← TO event; duration ~100 ms; target: hip flexion initiation torque.
• S5 initial-swing ← internal timer; duration ~150 ms; target: knee flexion to clear ground.
• S6 mid-swing ← ω_s peaks; duration ~150 ms; target: limb advancement, knee extension begins.
• S7 terminal-swing ← knee angle approaching extension; duration ~100 ms; target: knee extension brake to prevent hyperextension at HS.
• S8 loop ← HS event → S1.

Robustness: multi-sensor confirmation of HS (FSR + IMU shank-deceleration) reduces false transitions during disturbances. HMMs and LSTM classifiers replace hard thresholds in newer designs.

14.1 Worked example: torque-current sizing for an active knee exo

Design specification: assist-as-needed knee flexion/extension for a 75 kg adult during sit-to-stand and stair ascent. Peak knee torque required from biomechanics literature (Bobbert et al., 2008; Riener et al., 2002): sit-to-stand ~1.1 N·m/kg, stair ascent ~1.0 N·m/kg. For 75 kg user, peak required ~80 N·m; targeting 50 percent power-assist ratio means exo delivers ~40 N·m peak.

Step 1 — motor + reduction. A flat BLDC like the Maxon EC-4pole 30 (200 W class) has a stall torque of about 0.2 N·m. To reach 40 N·m at the joint with margin, use a 200:1 strain-wave reduction (Harmonic Drive CSF-25 family). Continuous joint torque at the gear’s nominal rating: 0.1 N·m × 200 = 20 N·m continuous, 0.4 N·m × 200 = 80 N·m peak.

Step 2 — speed budget. Sit-to-stand knee angular velocity ~120 °/s = 2.1 rad/s at the joint. Motor speed at this joint speed: 2.1 × 200 / (2π) × 60 = 4000 rpm motor — within the EC-4pole’s 8000 rpm rating.

Step 3 — current budget. Torque constant k_t ≈ 25 mN·m/A. Peak phase current for 0.4 N·m motor torque: 16 A. 48 V bus, FOC drive (TI DRV8323 + LMG / GaN halves), I²R losses dominate at low speed.

Step 4 — thermal. Continuous I²R at 16 A × 0.1 Ω phase resistance × 2/3 (FOC) ≈ 17 W copper loss per motor. Two motors per leg = 34 W. Casing rises ~30 K above ambient under sustained walking; sit-to-stand duty cycle (~20 percent on) keeps continuous below 1/3 of peak.

Step 5 — battery. Average motor power per leg ~50 W under walking; two legs = 100 W; plus 30 W electronics + sensors + actuators-at-rest = 130 W. 1.2 kWh pack → 9 hr nominal; under realistic intermittent use → 4–5 hr field life.

See [[Engineering/electric-motors]] and [[Engineering/power-electronics]] for the motor and drive design depth.

14.2 Worked example: human-in-the-loop assist optimisation

Following Zhang et al. (2017) and Slade et al. (2022): given an active ankle exoskeleton with a parametric assist profile (peak torque magnitude τ_peak, peak timing t_peak as percent of stride, rise time t_rise, fall time t_fall), tune the four parameters per user to minimise metabolic cost.

Step 1 — define the search space. Bound each parameter physiologically: τ_peak ∈ [0, 0.5 N·m/kg], t_peak ∈ [40, 60 percent stride], t_rise ∈ [10, 30 percent stride], t_fall ∈ [10, 30 percent stride]. Encode as a 4-vector.

Step 2 — measurement. Indirect calorimetry with 6-minute bouts (Brockway equation, see §8.1). Discard first 3 minutes (transient). Each evaluation costs ~10 minutes wall-clock.

Step 3 — optimiser. CMA-ES (Covariance Matrix Adaptation Evolution Strategy) — black-box, gradient-free, handles noise well. Population size 8, generations 4 → 32 evaluations → 5–6 hours of subject time.

Step 4 — convergence. Per-user optimum converges in 4–6 generations (Slade 2022 reported median 20 minutes on a faster Bayesian-optimisation-based pipeline). Typical metabolic reduction: 10–17 percent over walking without device; 5–10 percent over a generic (non-personalised) assist profile.

Step 5 — generalisation. Per-user optima cluster but do not coincide. Population-mean assist profile delivers about half the benefit of personalisation. This is the principal argument for in-clinic or in-app calibration as a deployment step rather than factory tuning.

See [[Math/gradient-descent-variants]] and [[Robotics/rl-for-control]] for the optimisation algorithm depth, and [[Engineering/system-identification]] for the model-based alternative.

15. Common pitfalls

• Over-promising metabolic gain. Lab data on treadmill at fixed speed does not translate to field. HULC’s cancellation came from honest field measurement.
• Ignoring donning/doffing time in adoption planning. A 15-minute donning protocol is a deal-breaker for industrial shift work.
• Joint-axis misalignment assumed to be a fitting problem when it is actually a design problem. Self-aligning linkages cost mass but cure root-cause shear.
• Battery sizing for marketing peak, not field duty cycle. A 4-hour rated battery often delivers 2–2.5 hours under realistic intermittent assist demand.
• Conflating impedance and admittance control. They are causal duals; choice depends on whether the sensor is force or position; mis-matching causes instability under stiff or soft environments respectively. See [[Robotics/impedance-control]].
• Skipping clinical evidence. A 510(k) clearance establishes substantial equivalence, not efficacy. RCT data is what payers want for personal-use coverage.
• Cybersecurity afterthought. FDA 2023 guidance treats medical-device cyber as premarket-required, not post-market patch.
• Anthropometric ranges that exclude the bottom 5th and top 5th percentiles — common in early industrial designs and a deployment killer.

15.1 Deployment lessons from the field

• Boeing / Levitate AIRFRAME (2018–). Deployed in 787 fit-out at Everett. EMG reduction validated; the surprise lesson was that workers self-selected away from the device for short overhead tasks (<30 s) because donning time exceeded the task time. Adoption succeeded on long-duration overhead drilling and riveting where the device sat on for full shifts.
• Ford / EksoVest (later EVO). Pilot 2017, full-line deployment 2018 across multiple US and EU plants. Adoption succeeded because the device became a normalised tool issued at shift start; rejection failures came from poor fit on the 5th-percentile female and 95th-percentile male users — Ford pushed back on Ekso to expand the sizing range, which became EVO-S.
• Hocoma Lokomat in stroke rehab. Twenty-plus years of deployment shows the dominant adoption driver is not therapist preference but patient throughput — a Lokomat enables one therapist to supervise where two were needed before. Hospitals justify the USD 250 k capex on labour-cost basis, not patient outcome.
• Sarcos Guardian XO at Delta Air Lines. Pilot 2022, evaluation 2023. The lesson reported in press: operators loved the device on heavy-item moves (90 kg) but found it overkill for the long tail of 5–20 kg moves that dominate baggage operations. Sarcos’s pivot to Palladyne AI / Guardian XT (teleoperated mobile arm) acknowledges this — the wearable form factor is not the right answer for most warehouse work.
• Wandercraft Atalante in SCI rehab. Adoption succeeded in clinics where it replaces 2-person body-weight-supported treadmill training with 1-person Atalante sessions, doubling therapist throughput. Failure modes are rare but salient: pelvis-fit mismatch on bony or atrophied patients causes interface pain that ends sessions.
• ReWalk Personal post-CMS coverage. Coverage decision December 2023 unlocked the previously stalled personal-use market. Early indication 2024–2025 is that prescription volume grew but training-time and home-environment-survey requirements (raised by CMS to limit risk) keep the funnel narrower than expected.

16. Citations and key references

• Pratt, G. A., Williamson, M. M. (1995). Series elastic actuators. IROS ‘95. — Foundation for compliant HRC actuation used across exos.
• Hogan, N. (1985). Impedance control: An approach to manipulation, Parts I–III. Journal of Dynamic Systems, Measurement, and Control 107(1): 1–24. — Core HRC framework.
• Kazerooni, H., Steger, R. (2006). The Berkeley Lower Extremity Exoskeleton (BLEEX). J. Dyn. Sys., Meas., Control 128(1): 14–25.
• Walsh, C. J., Endo, K., Herr, H. (2007). A quasi-passive leg exoskeleton for load-carrying augmentation. International Journal of Humanoid Robotics, 4(3): 487–506.
• Sankai, Y. (2010). HAL: Hybrid Assistive Limb based on Cybernics. In Robotics Research, STAR vol. 66, Springer.
• Esquenazi, A., et al. (2012). The ReWalk powered exoskeleton to restore ambulatory function to individuals with thoracic-level motor-complete SCI. Am J Phys Med Rehabil, 91(11): 911–921.
• Asbeck, A. T., De Rossi, S. M. M., Holt, K. G., Walsh, C. J. (2015). A biologically inspired soft exosuit for walking assistance. IJRR, 34(6): 744–762.
• Collins, S. H., Wiggin, M. B., Sawicki, G. S. (2015). Reducing the energy cost of human walking using an unpowered exoskeleton. Nature, 522(7555): 212–215.
• Zhang, J., et al. (2017). Human-in-the-loop optimization of exoskeleton assistance during walking. Science, 356(6344): 1280–1284.
• Mehrholz, J., et al. (2017). Electromechanical-assisted training for walking after stroke. Cochrane Database Syst Rev.
• Awad, L. N., et al. (2017). A soft robotic exosuit improves walking in patients after stroke. Science Translational Medicine, 9(400).
• Slade, P., Kochenderfer, M. J., Delp, S. L., Collins, S. H. (2022). Personalizing exoskeleton assistance while walking in the real world. Nature, 610(7931): 277–282.
• Toxiri, S., et al. (2019). Back-support exoskeletons for occupational use. IISE Trans Occup Ergon Hum Factors, 7(3-4): 237–249.
• Yan, T., et al. (2015). Review of assistive strategies in powered lower-limb orthoses and exoskeletons. Robotics and Autonomous Systems, 64: 120–136.
• Delp, S. L., et al. (2007). OpenSim: Open-source software to create and analyze dynamic simulations of movement. IEEE Trans Biomed Eng, 54(11): 1940–1950.
• FDA databases: K131798 (ReWalk Personal), K161443 (Ekso GT), K171334 (Indego), K162791 (HAL), K183300 (ReWalk Restore).
• ISO 13482:2014 — Robots and robotic devices — Safety requirements for personal-care robots.
• ASTM F48 series — Exoskeletons and Exosuits (terminology, ergonomics, testing, performance).
• IEC 60601-1 / 62366-1 / 62304 — medical-device general safety, usability engineering, software lifecycle.

Adjacent

• [[Robotics/legged-robotics]] — bipedal balance and gait control underpin lower-limb exos.
• [[Robotics/impedance-control]] — the core human–robot contact framework.
• [[Robotics/prosthetics]] — companion field; SEAs, EMG decoding, socket pressure carry across.
• [[Robotics/humanoid-balance]] — ZMP and capture-point control underpin Atalante’s self-balancing.
• [[Engineering/biomechanics]] — joint-torque modelling, inverse dynamics, gait analysis.
• [[Engineering/ergonomics-human-factors]] — workplace integration, adoption, donning-time economics.
• [[Engineering/bioinstrumentation]] — surface EMG acquisition and conditioning.
• [[Robotics/safety-standards]] — ISO 13482, ASTM F48, IEC 60601 stack.