Back and shoulder injuries from repetitive lifting are a leading cause of sick leave in warehouses, production lines and nursing. Active (motorised) exoskeletons add motors at hip or back; an on-device model recognises what the wearer is doing and when a lift starts, and adds torque in sync. A cloud portal aggregates usage and ergonomic load data. The aim is less strain per shift, not replacing workers.
Problem & value
- Today: manual handling of parcels, parts and patients is mitigated by job rotation, lift aids and training. Many tasks (mixed parcels, patient transfers, tight spaces) cannot be fully automated or fitted with fixed hoists. Musculoskeletal disorders (MSDs) remain a main occupational-health problem in the EU; EU-OSHA treats manual handling as a core risk factor [1].
- AI value: passive exoskeletons (springs, e.g. Ottobock Paexo) give constant support. Active suits use intent recognition so assistance is applied only during the lift and does not hinder walking or stooping. Connected analytics show how often and how heavily people bend and lift, and feed risk assessments (Gefährdungsbeurteilung).
- Evidence: vendors report reduced muscle load and fewer sick days in pilots, e.g. German Bionic reports deployments at DPD, Dachser and Fiege [2]. Independent long-term evidence on injury reduction is still limited; treat vendor figures as pilot results and measure on your own site.
Variants
- Logistics (parcel hubs, warehouses): back/hip assist for lifting and carrying; fleet of shared suits per shift.
- Manufacturing & assembly: shoulder/arm support for overhead work is mostly passive; active back assist for heavy part handling.
- Care (nursing, long-term care): assist during patient transfers and repositioning; extra hygiene, safety and acceptance requirements; may touch medical-device rules.
- Construction & tools: load-bearing for heavy tool work (e.g. Hilti EXO-O1 is passive shoulder support).
- Out of scope: medical rehabilitation exoskeletons for paralysed patients (e.g. Wandercraft, ReWalk) are medical devices under HLTH; military load-carriage suits are not covered.
How it works
- IMU + joint angle + motor current + force sensors (100-1000 Hz)
- on-device state estimation (gait / posture phase)
- intent recognition (lift onset, carry, walk, stairs; small NN or classical filter)
- assist controller (torque profile, safety limits, hardware watchdog)
- motors at hip/back + user override (button, voice or app)
- logging (pseudonymised)
- cloud portal
- ergonomic analytics, maintenance, fleet reports
- Real-time layer: a microcontroller or small SoC runs the control loop deterministically. The ML model (typically a compact CNN/LSTM/TCN, quantised, TFLite Micro or ONNX) only selects mode and timing; hard torque, angle and speed limits are enforced by non-ML safety logic. This separation matters for certification.
- Training: policies and recognisers are trained on motion-capture and IMU data from many wearers, increasingly with musculoskeletal simulation and RL for assist profiles (see ROBO-010).
- Human-in-the-loop: the wearer always has priority; the suit must be switchable off instantly and yield to the wearer's motion. Analytics are used by occupational safety, not for individual performance ranking.
Data preparation
For bought exoskeletons the data needs are small: the device ships with its own sensors and control software. What you need is workplace data to choose the right use: the tasks and lifting loads (from occupational safety risk assessments, HR and the safety officer's files), sick-day and musculoskeletal complaint statistics (HR, often in Excel or paper, personal and health-related, so aggregate or pseudonymise first, see PREP-022), shift plans and workplace layouts. Typical gaps: no quantified load data per task, scattered ergonomic assessments, unclear consent rules for wearer data, and no baseline to prove an effect.
Preparation steps: list the candidate tasks with weight, frequency and posture; collect a baseline of complaints and absence per workplace before the pilot; agree which device data (usage time, assist level, motion data) is stored, where, and who may see it; define pilot KPIs and a simple feedback form for wearers (PREP-021). Tools: Excel or a simple database, the manufacturer's portal, optionally a pseudonymised export. Effort is low, roughly 10 to 20 percent of the project; developers building their own control need motion-capture and EMG datasets and should plan far more (GUIDE-003).
Implementation paths
Which path for whom
- SMB: A, because mature, certified exoskeletons from EU manufacturers exist at SMB prices and the use (lifting, overhead work) is standard.
- Enterprise: A, because the exoskeleton is a commodity aid; the effort lies in occupational safety, rollout and fit, not in building your own device.
- Developer: C, because open control and gait-intent models (ROS 2, open datasets) are needed to build or adapt actuation and assistance algorithms.
Path A — Buy (recommended)
- Active suits: German Bionic (Augsburg, EU; Cray X with the cloud platform Apogee), Ottobock (EU; Paexo passive range, plus Ottobock's industrial portfolio), Noonee (CH, passive wearable chair), Skelex (NL, passive), Sarcos (US) reportedly discontinued its Guardian XO exoskeleton hardware (company renamed Palladyne AI, now software-focused), so treat it as unavailable, Samsung and Hyundai/Wearable Robotics research wearables. Hilti EXO-O1 and Ottobock Paexo are passive and a cheaper alternative where active torque is not needed.
- Choose A when: you want a certified product, service and a fleet portal. Pilot 5-10 suits on one task type for 3 months, with works council involvement.
Path B — Build on cloud APIs
Rarely applicable for control. Cloud is used for analytics: ingest suit telemetry (MQTT/HTTPS) into an EU-region time-series store and BI tool, compute lift counts and load indices, and expose them through the vendor API (e.g. Apogee data export) into occupational-health dashboards. LLMs can summarise ergonomic reports for safety officers but must not process identifiable per-person data without a legal basis.
Path C — Self-host (open source)
For research, prototypes or university projects: ROS 2 with ros2_control, Pinocchio/OpenSim or MuJoCo for biomechanics, Edge Impulse or TensorFlow Lite Micro / ONNX Runtime on STM32 or NVIDIA Jetson Orin Nano class boards (HW1), time-series storage in TimescaleDB/InfluxDB with Grafana. Building a product-grade wearable robot this way still requires machine and PPE conformity work.
Path D — Custom training
Worth it only for exoskeleton manufacturers and research groups: collect labelled IMU/EMG/motion-capture data from diverse wearers (body size, sex, age, task), train intent and gait recognisers, tune assist profiles with RL in simulation, then validate with user studies. Data needs: dozens to hundreds of wearers, with consent.
State of the art (2026)
- Intent recognition from IMU-only signals on-device is mature for lifting, walking and stairs; assist profiles are increasingly learned (RL in musculoskeletal simulation, human-in-the-loop optimisation) rather than hand-tuned.
- Fleet analytics with ergonomic load indices are now sold with the hardware.
- Limits: unusual movements, carrying unbalanced loads, crowded spaces and different body types cause misclassification and mistimed assistance; battery life limits shifts; weight and heat reduce acceptance; assistance can shift load to other body parts. Real-world injury-reduction studies are still scarce.
- Not yet: general embodied foundation-model control of exoskeletons (ROBO-004) is research-stage.
Shared resources & combinations
- Edge ML stack (edge-runtime, time-series, anomaly-detection) is shared with ROBO-011 smart sensors.
- Simulation and RL tooling shared with ROBO-010 (sim-to-real).
- Posture analytics from cameras complement the suit: VIS-024 (pose and ergonomics), VIS-007 (PPE).
- Safety context: ROBO-016 when suits are used near AMRs and robot cells; ROBO-002 in intralogistics.
- Maintenance: PRED-006 for fleet health; HLTH-025 for care documentation of transfers.
Effort, cost & KPIs
- Path A: pilot in weeks to months incl. risk assessment, training and works-council agreement; per-suit purchase or lease plus portal subscription (check vendor quotes; do not assume prices).
- Path C/D: 12+ months for a validated prototype; specialised mechatronics, controls and ML team.
- KPIs: MSD-related sick days, lifting events and load index per shift, perceived exertion (Borg scale), voluntary usage rate, adoption after 3 months, incidents and near-misses, productivity per task, suit availability.
People & change
Involve management and the head of logistics or production, the occupational safety officer and the company doctor first; they decide where the aid helps. HR and the DPO come in because wearable sensors can capture motion and health-related data of identifiable employees. End users must try the devices voluntarily; fit and comfort decide acceptance.
Works council: yes. Devices that record usage, motion or load data are objectively suited to monitor behaviour or performance, so §87(1) no. 6 BetrVG applies, besides health and safety co-determination; §90(1) no. 3 requires early information (GUIDE-004). Conclude a works agreement: no individual performance evaluation, limited retention, no use for disciplinary action. Health-related data is a special category under Art. 9 GDPR, so run a DPIA if device data is stored per person and sign an AVV with the vendor (GUIDE-005).
Support AI literacy (Art. 4 AI Act) with short instruction on how the assist adapts and when to switch it off (GUIDE-006). Typical concerns: pressure to work faster, surveillance, health effects. Address them with voluntary use, a trial phase with a volunteer pilot group of 5 to 10 people, champions from the shop floor and transparent KPIs (GUIDE-008). Humans stay in control: the wearer can adjust or switch off the assistance at any time.
Risks, regulation & compliance
- Product law: a workplace exoskeleton is typically machinery (Machinery Directive 2006/42/EC, replaced by Regulation (EU) 2023/1230 from 20 January 2027) or PPE; care and rehabilitation suits with a medical purpose fall under the MDR (EU) 2017/745. The manufacturer chooses the route; employers should check the CE declaration and intended purpose [3][4].
- EU AI Act: depends. If the AI is a safety component of a product regulated under Annex I (machinery, medical device), it is high-risk under Art. 6(1) with the application dates tied to that product law; check the current timetable. A pure intent recogniser inside non-safety logic is minimal risk. Art. 4 AI literacy applies to employers [5].
- Art. 5 prohibition warning: inferring workers' emotions (stress, fatigue as mood, frustration) from physiological or biometric signals is prohibited in the workplace except for medical or safety reasons (Art. 5(1)(f)). Do not add emotion inference to suit analytics. Fatigue/strain estimates must be about physical load for safety and not used to rate attitude.
- GDPR and employment law: motion, heart-rate and load data are personal data of employees and may reveal health (Art. 9). Use pseudonymised or team-level aggregates, purpose limitation (ergonomics, not performance control), DPIA, retention limits and works-council co-determination (in Germany BetrVG §87(1) no. 6 for technical systems that monitor employees) [6].
- Occupational safety: risk assessment under the OSH Framework Directive 89/391/EEC; exoskeletons are a supplementary measure and do not replace technical and organisational measures.
- Other risks: falls or entrapment on power failure, wrong assist timing, skin pressure, dependence and muscle deconditioning, uneven fit for women or small bodies. Mitigate by fitting, training, opt-in voluntary use, free-wheel mode on fault and medical clearance where needed.
Getting started
- Identify two or three high-load tasks via incident data and ergonomic assessment (e.g. RULA/NIOSH); check whether a passive or active suit fits.
- Involve the works council, occupational physician and data protection officer before the pilot; agree on data scope.
- Run a 3-month pilot with 5-10 volunteers; log exertion scores, sick days, usability and incidents; compare with a control group.
- Request the CE declaration, intended purpose, safety manual and data-processing agreement from the vendor.
- Decide on rollout by task and shift; plan cleaning, charging and fitting logistics for a shared fleet.
Sources
- EU-OSHA — Work-related musculoskeletal disorders — https://osha.europa.eu/en/themes/musculoskeletal-disorders
- German Bionic — Cray X and Apogee platform — https://www.germanbionic.com
- Regulation (EU) 2023/1230 on machinery — https://eur-lex.europa.eu/eli/reg/2023/1230/oj
- Regulation (EU) 2017/745 (MDR) — https://eur-lex.europa.eu/eli/reg/2017/745/oj
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Regulation (EU) 2016/679 (GDPR) — https://eur-lex.europa.eu/eli/reg/2016/679/oj
- Ottobock — Paexo industrial exoskeletons — https://paexo.com/en
- Hilti — EXO-O1 exoskeleton — https://www.hilti.group
- ISO 13482 — Safety requirements for personal care robots — https://www.iso.org/standard/53820.html