Lightweight active back exosuit reduces muscular effort during an hour-long order picking task

Low back injuries are the most common workplace injury in the U.S., imposing substantial economic and personal burdens. Back-support exoskeletons and exosuits can reduce spinal loading during lifting, but adoption hinges on devices providing clear biomechanical benefits without hindering performance or comfort over extended, real-world tasks. Prior work shows assistance level, device stiffness, and adaptability affect back muscle activity and user experience, with passive systems often trading adaptability for simplicity and active systems risking added weight and complexity. The research question is whether a lightweight, soft, active back exosuit with adaptive impedance control can be highly usable and robust in dynamic, hour-long warehouse-like order picking tasks, while reducing muscular effort without disrupting movement. The study develops such a device and evaluates its usability, controller robustness, and biomechanical efficacy in a simulated real-world setting.

The paper reviews known usability challenges for occupational back exos (weight, complexity, joint misalignment, discomfort, restriction, movement disruption). Rigid exoskeletons deliver high moments but risk misalignment; soft exosuits avoid misalignment but typically provide lower moments and require careful anchoring. Passive systems can be tuned for specific tasks but adapt poorly across varied tasks and may restrict reaching and walking; active systems offer adaptability but often add weight and complexity. Prior controllers distinguish lifting vs lowering via state machines or switches; others scale assistance using EMG or grip/load signals. Field-like evaluations over longer durations are needed, especially for order picking, where only one prior study showed a 10.5% reduction in mean back extensor EMG with a passive exoskeleton but increasing discomfort over time. The authors’ previous work showed an active exosuit with adaptive assistance reduced peak back extensor EMG up to 15% in constrained tasks, suggesting potential if robustness can be shown in dynamic tasks.

Device and controller: The soft, textile-based back exosuit integrates a lightweight cable-driven actuator that tensions a ribbon spanning the back and hips, acting in parallel with erector spinae. Three IMUs (back and both thighs) measure kinematics; a load cell measures ribbon tension for closed-loop control. Total mass is 2.7 kg with overall torque density 11.1 Nm kg−1. Functional apparel includes backpack with actuation and controller units, shoulder and chest straps, and adjustable thigh wraps with a BOA system to accommodate sizes. The rigid actuation uses a brushless DC motor driving a 4:1 pulley to a ribbon spool; sensors include a hall-effect encoder, load cell (FUTEK LSB200), and three IMUs (BNO085). Two 28.8 V (in series) Li-ion batteries enable ~12 h operation at 200 lifts/h with 250 N peak force. The controller unit (ATSAME70N21 MCU) runs force control at 1 kHz; distributed 8-bit PIC microcontrollers on back and thighs handle IMU and analog signals over CAN to the motor driver (Elmo Gold Twitter). Controller algorithm: Assistance is commanded via an adaptive impedance function based on relative trunk angle and angular velocity. Relative angle is defined as θ_rel = θ_trunk + 0.5(θ_RT + θ_LT) − 12° − 0.3·abs(θ_RT − θ_LT), enabling assistance in both squat and stoop while minimizing assistance during walking. The controller applies higher impedance during lifting and lower during lowering, with non-linear (sine) scaling to accommodate changing trunk moment arm, delivering peak assistance (up to 250 N or ~30 Nm) beyond ~90° relative angle. A wide transition phase (±120°/s) uses a quadratic interpolation between lifting and lowering force commands to avoid mistriggers during ambiguous movements (e.g., picking under a shelf). Low-level force tracking achieves small RMSE relative to command. Usability test: Ten participants (6 novices, 4 expert users) completed a usability protocol including self-donning, checking range of motion, walking, transferring boxes, and doffing. Timed tasks measured time-to-don, time-to-lift, and time-to-doff across 5 repetitions. Participants completed the System Usability Scale (SUS) survey. Order picking experiment: Fifteen participants (11 men, 4 women; 31 ± 4 years; 73 ± 12 kg; 172 ± 13 cm; BMI 25 ± 5 kg/m^3) performed a 1-hour order picking task on two days in randomized counterbalanced order: with exosuit assistance and no suit. Each participant completed 320 cycles (640 lifts; two lifts per cycle) involving lifting a 10 kg box from a pallet under a 4-ft covered rack, transferring to an uncovered pallet 8 ft away, then walking to a cone and returning; half of the lifts occurred under a shelf. A metronome paced one transfer every 12 s. Short data breaks (1–2 min) occurred every 15 min. No instructions were provided on lifting style or speed. External measurement IMUs (Xsens MTi-3 on T8 and both thighs at 200 Hz) captured kinematics in both conditions. EMG (Delsys, 2148 Hz) recorded eight sites across four muscle groups: back extensors (lumbar longissimus; thoracic/lumbar iliocostalis), abdominals (rectus abdominis; external obliques), hip extensors (gluteus maximus; biceps femoris), and rectus femoris. MVICs were collected for normalization. Load cell data from the exosuit quantified applied moments and tracking error. Data processing and outcomes: EMG was band-pass filtered (50–450 Hz), rectified, and low-pass filtered to a 6 Hz linear envelope; amplitudes were normalized to MVIC. Peak EMG per event was defined as the 90th percentile (APDM method), with analyses segmented for lifting and lowering; median EMG was also computed. Primary outcome: peak EMG across four muscle groups; secondary: median EMG. Kinematic outcomes included peak trunk flexion angle and absolute trunk angular velocity per event; event durations were also computed. IMU metrics classified squat vs stoop-style lifts. Device performance included peak load cell moment and RMSE between commanded and measured force. ROM impacts were assessed in a subset (frontal/transverse planes). Event detection used thresholds on relative trunk flexion. Statistics: Linear Mixed Model (LMM) ANOVAs tested effects. For the picking task, a three-factor LMM included condition (exosuit vs no suit), muscle group (4), and epoch (four 15-min segments). Bonferroni corrections were applied for co-primary outcomes; alpha 0.01 for secondary measures. Tukey’s HSD was used for post hocs. Transformations addressed normality/linearity violations. Power analysis indicated 14 participants would detect condition differences in peak back extensor EMG with 80% power (α=0.05). Ethics approval: Harvard Medical School IRB (IRB18-0960); informed consent obtained.

• Usability: Excellent overall SUS score 92.8 ± 5.3; users rated the exosuit as not complex (1.20 ± 0.42/5), easy to learn (4.90 ± 0.32/5), and felt confident using it (4.80 ± 0.42/5). Average time-to-don 35.6 ± 2.5 s, time-to-lift 45.3 ± 2.8 s, time-to-doff 7.2 ± 0.4 s.
• Device characteristics and ROM: Lightweight 2.7 kg with high torque density (11.1 Nm kg−1). Minimal movement restriction: no significant restriction in lateral plane ROM; transverse plane ROM reduced by 4.3° (4.5%) (T(9)=3.14, p=0.012) compared to without suit.
• Controller robustness: Only 14 mistriggers out of 9600 lifts (0.1%) across 15 participants; all occurred under-shelf picks and all but one in a single participant. Simulations showed the ±120°/s transition window was necessary to keep mistriggers below 1% across mass orientations. Load cell indicated peak assistance 18.6 ± 5.4 Nm across pallet positions with RMSE ~0.33 ± 0.7 Nm between commanded and measured forces.
• EMG reductions: Compared to no suit, peak EMG decreased by 18% (back extensors), 11% (hip extensors), and 22% (rectus femoris); condition × muscle group F(3,3506)=40.0, p<0.001. Median EMG decreased by 20% (back extensors), 13% (hip extensors), and 20% (rectus femoris). No significant condition effect for abdominal EMG, suggesting no undue co-activation. Reductions were consistent across four 15-min epochs; no condition × epoch interaction (p>0.05), while EMG amplitudes generally decreased over time (epoch main effects significant).
• Kinematics and performance: Slightly higher peak trunk flexion with exosuit (100.4 ± 28.4°) vs no suit (99.8 ± 28.7°), F(1,882)=5.9, p=0.015. No significant differences in peak trunk velocity or lift/lower durations. Trend toward more squat-style lifts with the exosuit. Assistance delivered more than 80% of peak over a wide trunk flexion range (53–127°) due to the sine impedance strategy.

The soft, active exosuit with adaptive impedance delivered practical assistance during complex, hour-long order picking without hindering movement, demonstrating both high usability and robustness. The wide transition window minimized mistriggers in dynamic, under-shelf tasks, maintaining user trust and consistent assistance delivery using only integrated motion sensors to preserve usability. Biomechanically, the exosuit reduced back extensor EMG by 18–20% and also decreased hip and knee extensor EMG, indicating beneficial load sharing across joints. Compared to reports of passive devices in similar tasks that showed smaller mean EMG reductions and increased discomfort over time, the active, sine-shaped impedance profile provided substantial assistance over a broad range of trunk flexion, and low hysteresis with tight force tracking likely contributed to the greater EMG reductions observed. Consistent reductions across 320 lifts suggest early adaptation with stable benefits. Overall, the results support the feasibility of deploying an adaptive active exosuit in real-world warehouse operations to mitigate back loading while maintaining natural movement.

A lightweight (2.7 kg), soft, active back exosuit with an adaptive impedance controller was developed and shown to be highly usable (SUS 92.8), robust (0.1% mistriggers over 9600 lifts), and biomechanically effective in an hour-long order picking simulation, reducing peak and median back extensor EMG by 18–20% without adversely affecting lifting kinematics. The device’s active control strategy delivered broad, direction-sensitive assistance suitable for dynamic tasks. Future work should conduct field studies over full shifts to assess long-term usability and injury risk mitigation, integrate robust load-detection to adapt assistance during lowering, and refine controller transitions or activity recognition to deliver faster, context-aware assistance (including support for lunging lifts) without increasing mistriggers.

• Assistance asymmetry: Designed to assist lowering less than lifting to reduce restriction, potentially sacrificing biomechanical benefit during lowering where lumbar moments may be similar to lifting.
• Transition window: The wide ±120°/s transition reduces mistriggers but can delay peak assistance onset, which may reduce biomechanical benefit compared to faster assistance.
• Relative angle approach: The angle definition reduces assistance during walking (desirable) but also diminishes support for lunging lifts or asymmetric thigh flexion patterns.
• Sensor simplicity: Only motion sensors were used; absence of load detection (e.g., grip/EMG/pressure) limits context-aware force adaptation.
• Biomechanical measurements: No direct measures of muscle forces or spinal compression; inferences are based on EMG and kinematics.
• EMG analysis: Temporal EMG features within lifts were not analyzed due to variable movement timing.
• Potential biomechanical alterations: Passive/inertial properties of the exosuit may alter user biomechanics; a transparent-control condition was not included.
• Usability protocol: Did not include battery insertion/removal, which could affect real-world usability perceptions.