In this study, we use two prototypes of the Utah Knee Exoskeleton, shown in Fig. 1. The Utah Knee Exoskeleton consists of an actuated knee joint and a passive ankle joint. The ankle joint uses a free hinge to support the weight of the device without applying torque at the ankle. The knee joint can apply torques of up to 55 Nm, despite the device’s low weight (2.6 kg including battery and physical interfaces). To achieve this high torque density (~21 Nm/kg), the Utah Knee Exoskeleton uses a geared brushless DC motor (EC-22 120 W and 3.8:1 gearbox, Maxon Motors, Switzerland) in combination with a torque-sensitive actuator, which consists of a five-bar linkage with an elastic element similar to [20]. This torque-sensitive actuation (TSA) changes the transmission ratio passively and continuously as a function of the output torque, allowing for a wider range of torque-velocity profiles than a fixed transmission ratio.

The DC motor is powered by a current driver (Everest Core, Ingenia, Spain), which uses an absolute encoder (RM08, RLS, Slovenia) for commutation. Absolute encoders also measure the deformation of the elastic element in the torque-sensitive transmission (iC-MU & MU2S, iC-Haus, Germany) and the angular position of the main output joint (iC-MU & MU18S, iC-Haus, Germany). An inertial measurement unit (IMU, MTi-1, Xsens Technologies B.V., The Netherlands) is attached to the wearer’s thigh to measure its global orientation and rotational velocity in the sagittal plane. The peripherals are interfaced through two 32-bit microcontrollers (PIC32, Microchip Technology, USA). Finally, an embedded computer (Raspberry Pi 4 Compute Module) communicates with the microcontrollers and to a remote computer over Wi-Fi for online data telemetry and control parameter tuning.

A set of semi-rigid interfaces transfers the assistive loads generated by the active knee exoskeleton to the user’s lower limbs. The interfaces comprise a thigh cuff, a shank cuff, and a footplate (Fig. 1). A free hinge connects the shank cuff to a three-quarter length footplate, to preserve natural motion at the ankle and toe joints. The ankle joint and footplate prevent the device from sliding down or twisting about the leg, which ensures that the torque at the knee is accurately applied. During fitting, the interfaces are adjusted to each person to minimize misalignment between the user’s biological knee joint and the powered knee joint axis.

The assistance provided by the active knee exoskeleton is determined by the controller represented in Fig. 3, which has feedforward compensations as well as a state-dependant assistive torque profile.

The feedforward compensation torque (1) τc=−(Iθ¨m+bTSAθ˙TSA+τfric)

lowers the apparent impedance of the exoskeleton to the wearer. θ˙TSA(rad/s) is the velocity of the elastic element’s joint whereas θ˙m, θ¨mrad/s,rad/s2 are the motor velocity and acceleration, used in conjunction with the feedforward inertia Ikgm2 and the TSA- and motor damping terms bTSA, bm(Nm/s). The TSA damping is set to zero when the elastic element flexes, and set to a negative value when it extends, which dampens the release of energy stored in the elastic element. The friction compensation, τfric, is conditional based on a speed threshold, θ˙mTHS(rad/s), and described through (2) τfric=τcoulifθ˙m>θ˙mTHS,kfricθ˙motherwise.

For the assistive torque profile, a state machine splits the wearer’s gait into a flexion and extension phase, depending on the thigh angle relative to the vertical θt and the thigh angular velocity θ˙t, which are illustrated in Fig. 3. The extension phase broadly overlaps with stance phase of the gait cycle, but may begin slightly before foot contact occurs. The transition conditions were based on nominal biomechanics data [21] and refined during pilot testing. While these values worked well for all subjects on our specific step mill, it may need to be adapted for different stair heights. The exoskeleton only applies assistive torque during the extension phase, as nominal biomechanics data of stair ascent indicates that minimal knee torque occurs during the flexion phase [21]. However, assistive torque is not necessarily applied at the transition to the extension phase.

A position-based, trapezoidal torque profile determines the desired assistive joint torque τd as a function of thigh angle, θt. The trapezoidal profile is described with five parameters: The start and end angles of the profile θr,s,θr,e the start and end angle of the profile’s plateau θp,s,θp,e, and the peak torque τp (Fig. 3). We tuned these five parameters to each subject during the training sessions. The torque assistance profile begins at θr,s, the thigh angle at which the weight acceptance phase of the stride begins. Setting θr,s too high may cause gait instability by applying the torque before weight transfers to the stance leg, or even before foot contact occurs. Too low a starting angle causes greater quadriceps activation, yielding a lost opportunity for effective support. Similarly, θp,s is set such that the assistive torque ramps up sufficiently fast to support the increased weight on the stance leg without jerking said leg. Afterwards, the end of peak torque, θp,e, and the end of support, θr,e, are set such that the wearer is eased into full stance without overextending their knee. Finally, the peak torque is gradually increased during the training session until the wearer notes that the change in torque feels undesirable. The additional weight carried by the participant during the experiment is assumed to proportionally affect the required (assistive) knee torque.

To realize the total desired torque τ=τc+τd, a kinematic model of the drive train is combined with a dynamic model of the motor to compute the desired current I_d. This desired current is sent to the motor driver’s internal closed loop current control. We calculate the instantaneous transmission ratio of the drive train from the encoder data on both the motor and the elastic element [20]. The open-loop torque control method described above yields a steady state error of 1.2% in devices with a similar drive train design [20]. The low-level microcontrollers perform these calculations at a rate of 2 kHz. The controller parameters can be set via the embedded computer’s high-level control loop, which runs at 500 Hz. The embedded computer connects to a laptop via WiFi, so that these parameters can be tuned while the exoskeleton is running.

A convenience sample of eight participants with no mobility impairments participated in this study, as described in Table I. The Institutional Review Board at the University of Utah approved the study protocol (IRB_00120712). The subjects provided informed consent to participate in the study, as well as permission to publish photos and videos from the experiment.

Prior to data collection, participants familiarized themselves with the device during two 45-minute training sessions. During these sessions, the exoskeleton interfaces were adjusted to the user. After familiarization, the participants donned electromyography and motion capture sensors, shown in Fig. 2. We measured surface electromyography (sEMG), a proxy for muscular effort [22]. We placed four wireless sEMG sensors on each leg at the rectus femoris (RF), vastus medialis (VM), vastus lateralis (VL) (Trigno Quattro Sensor, Delsys Inc., USA), and gastrocnemius medialis (GM) (Trigno Avanti Sensor, Delsys Inc., USA). The ground units of the Trigno Quattro sensor also house an inertial measurement unit (IMU), which we use for data segmentation in post-processing.

We prepared the skin and placed the sensors according to SENIAM guidelines [23], with the exception of the gastrocnemius medialis sensor, which is located slightly more proximally to eliminate contact with the exoskeleton interface and reduce the potential for motion artifacts. We wrapped the sensors on the upper leg to minimize motion artifacts from interface deformation. The sEMG data is acquired through EMGWorks Analysis v4.8.0 (Delsys Inc., USA).

To measure joint kinematics, participants wore eight wireless inertial motion capture sensors on their lower limbs, pelvis, and sternum (Xsens MVN Awinda, Xsens Technologies B.V., Enschede, The Netherlands), drawn in orange in Fig. 2. A separate computer records the kinematic joint data through MVN Analyze 2020.0 (Xsens Technologies B.V., Enschede, The Netherlands). After sensor placement, subjects performed a calibration sequence. A dedicated trigger panel from DelSys synchronized the data acquisition between sEMG and motion capture systems [24]. The Utah Knee Exoskeleton recorded sensor data separately on its embedded computers.

Participants completed a task similar to event one of the Candidate Physical Abilities Test (CPAT) for firefighters in the United States [25], which requires candidates to ascend a step mill at a speed of 60 steps/min for three minutes, while wearing 34 kg of added weight and without touching the handrails. In our study, participants similarly ascended a step mill (StepMill 3, StairMaster, United States of America) at 60 steps/min for three minutes without using the handrails for weight bearing. The steps were 15 cm tall and 23 cm deep. However, because firefighter candidates are recommended to complete several weeks of fitness training before attempting the CPAT [25], and our participants were drawn from the general population with no prior training, we asked participants to wear 9.1 kg weighted vest rather than the full 34 kg. The participants completed the test twice: once with the exoskeleton providing assistive torque and once without the exoskeleton. Each trial began with a ninety-second warmup while the step mill slowly accelerated, followed by a three-minute data collection session. In the exoskeleton condition, the assistive peak torque slowly increased throughout the warm-up period, in order to ease the wearer into the assistance and avoid excessive co-contractions [26], [27]. In between trials, the participants rested for at least thirty minutes. The motion capture sensors were replaced and recalibrated between trials to accommodate donning or doffing the exoskeleton interfaces. Half the participants completed the exoskeleton trial first, while the other half started without the exoskeleton in an effort to average the effects of fatigue between conditions.

The data was analysed in Matlab R2022b (MathWorks, USA). sEMG data are exported using EMGWorks Analyze (Delsys Inc., USA), whereas the joint trajectories are exported through Xsens’ Matlab library [28]. Biological joint kinematics are high-pass filtered at 0.15 Hz using a second-order Butterworth filter. The sEMG data are band-pass filtered (20-450 Hz, second order Butterworth filter), rectified, and low-pass filtered using a 5 Hz second-order Butterworth filter. We compute peak activation and area under the curve for each stride, with the latter considered a proxy for total muscular effort. The calculated exoskeleton output torque is low-pass filtered with a 100 Hz second-order Butterworth filter, while output joint velocity data is low-pass filtered with a 40 Hz second-order Butterworth filter. Exoskeleton computed output torque and power are bodyweight normalized. All filter orders and frequencies are based on visual inspection of the raw and filtered data in both the temporal- and frequency domain.

After filtering, we segmented the data by stride. As foot contact data was not available, we segmented at the peak thigh angle relative to the vertical, which likely precedes heel strike slightly [21]. We then interpolated each stride to 1000 samples. Next, we normalized the filtered sEMG data with respect to the mean peak activation of the respective muscle in the no-exoskeleton condition. Finally, we calculated key outcome measures, such as minimum hip flexion angle, peak knee extension speed, and mean trunk lean.

We tested key metrics of the kinematics and sEMG data for statistically significant differences between the exoskeleton and no-exoskeleton conditions using paired t-tests. While all participants completed the same task, differences in height or body mass could make the task relatively more or less challenging. In addition, we tuned the exoskeleton individually for each subject. To assess the impact of these individual differences on performance, we used Pearson’s correlation coefficient to test for correlations between participant height and mass, and the peak level of assistive torque. We used Pearson’s correlation coefficient to test for correlations between the peak torque and power and changes in peak and integral sEMG activation and knee extension speed. The alpha level for all statistical tests is set to α = 0.05. When applicable, we applied the Bonferroni correction for multiple comparisons.

Fig. 4 displays the output torque and power trajectories of both exoskeletons across the gait cycle for each participant. The trapezoidal position-based torque profile manifests as a smoother and more bell curve-like profile in the time domain. The mean peak assistive torque ranged from 0.38 Nm/kg (participant 4) to 0.71 Nm/kg (participant 1) based on tuning, which corresponds to 42% – 79% of the nominal peak biological torque for unweighted stair ascent [21]. On average, the torque profiles (including compensations) start around 4% of the gait cycle and end at 40%.

The mean peak bodyweight normalized power ranged from 1.2 W/kg (participant 4) to 2.9 W/kg (participant 1), which is 56% and 133% of the peak biological knee power during unweighted stair ascent, respectively [21]. As described in Section II, the controller only provided positive power during the stance phase, with negligible power generated during swing. Exoskeleton power oscillated somewhat during peak assistance, particularly for participants with higher torques.

Finally, the bottom row of Fig. 4 indicates that both exoskeletons generated an average bodyweight normalized energy of approximately 0.4 J/kg over a stride.

The joint kinematics are presented in Fig. 5. Column (a) plots the joint kinematics averaged over all participants for each condition. At the knee level, Fig. 5 shows an increased knee extension velocity at the start of the gait cycle. Each participant increased their knee extension speed. On average, the knee extended 52% faster (p = 3.4 × 10⁻¹³, n_tests = 2), with no significant difference between legs (no exoskeleton condition: p = 0.53, exoskeleton condition: p = 0.50, n_tests = 2). Moreover, the knee extension phase shortens by approximately 10% of the gait cycle in the exoskeleton condition. On the other hand, more time is spent at maximum knee extension before the leg is lifted to the next step.

In the trunk kinematic data, Fig. 5(c) shows that participants lean further forward when wearing the exoskeleton during stair ascent by an average of 9°, although this change is not statistically significant (p = 0.08, n_tests = 1).

Regarding hip kinematics, the most notable difference is that the minimum hip flexion angle decreased by an average of 9° in the exoskeleton condition, p = 0.02, n_tests = 2. As with the knee, there is no significant difference between the left and right side for the minimum hip flexion angle (no exoskeleton: p = 0.67, exoskeleton: p = 0.73, n_tests = 2). Finally, the main variation in ankle kinematics between conditions appears during the swing phase. In the exoskeleton condition, peak plantarflexion is held for a shorter time while peak dorsiflexion is held longer, from the end of swing until weight acceptance.

The sEMG data are shown in Fig. 6. The right rectus femoris data of participant 4 and participant 6 are excluded due to substantial motion artifacts. In Fig. 6(a), a good agreement can be observed between the mean activation patterns of the left and right muscles in the same condition. Temporally, the mean peak of each activation pattern occurs roughly 5% earlier with the exoskeleton than without.

When wearing the exoskeleton, the participants reduced activation of all three measured quadriceps muscles. The mean integral of the sEMG signal, a proxy for total muscle activation, significantly decreased during the exoskeleton trial compared to the no-exoskeleton trial (RF: −31%, p = 0.0015; VM: −28%, p = 0.0011; VL: −27%, p = 1.410⁻⁴). Similarly, the peak sEMG value, analogous to the peak muscle activation, reduced significantly for all quadriceps muscles while participants wore the active knee exoskeletons (RF: −41%, p = 1.2 × 10⁻⁴; VM: −26%, p = 0.0048; VL: −29%, p = 1.9 × 10⁻⁴).

In contrast, the gastrocnemius muscle may activate more when using the exoskeleton. The mean integral gastrocnemius medialis activation increases slightly when participants wear the active knee exoskeletons, although this change is not significant (GM: +3%, p = 0.71). Similarly, the peak activation of the GM increased by 11% during the exoskeleton trials, although this change was not significant (p = 0.26).

There was a trend toward subjects with greater mass or height receiving larger peak torques, but these correlations were not significant (p = 0.079, p = 0.16; respectively). There were no statistically significant correlations between the peak torque or peak power values and the change in integral sEMG, peak sEMG, or knee extension speed. This suggests that differences in individuals may have been accounted for through changes in the control parameters, or that a larger study is needed to detect these differences.