2.1.1. Impairment method

To impair walking gait as well as the elevating limb recovery response, 12–14 lb (5.4–6.4 kg) leg weights were attached to the shank segment of the knee exoskeleton, as pictured in Figure 1(a). Methods for evaluating the efficacy of these weights in impairing walking gait and stumble recovery are discussed in the Data Analysis subsection.

2.1.2. Improvement method

A closed-loop walking controller was designed to assist the knee joint during level ground walking (Figure 1(b)), with the intent of correcting or reducing the simulated impairment introduced by the leg weights. The controller is comprised of a finite-state machine that transitions between two states: stance and swing. To prevent knee buckling during the initial stance, knee angle is regulated to zero degrees (set as the angle when the participant was standing upright with legs straightened, which was reset before each trial) using the following PD controller:

(1) $$ \tau ={k}_p\left({\theta}_{desired}-\theta \right)+{k}_d\left({\dot{\theta}}_{desired}-\dot{\theta}\right), $$

where $ \tau $ is the commanded torque, $ {\theta}_{desired} $ is 0 deg, $ \theta $ is the current knee angle, $ {\dot{\theta}}_{desired} $ is 0 deg/s, $ \dot{\theta} $ is current knee angular velocity, and $ {k}_p $ and $ {k}_d $ are experimenter-defined proportional and derivative gains, respectively. This assistance is removed 300 ms into stance to facilitate the initiation of swing phase by allowing knee flexion. To assist swing-phase motion, a position controller tracks a predefined knee angle trajectory (shown in Figure 1(b)) using Eq. 1, where $ {\theta}_{desired} $ and $ {\dot{\theta}}_{desired} $ are the desired knee angle and knee angular velocity from the spline and its derivative at the current time in swing phase. The elapsed time of swing state is determined by a swing-stance ratio and the stance time of the current stride. State transitions are determined by heel FSR loading and thigh angular velocity thresholds (Figure 1(b)). The following variables were tunable for each user: $ {k}_{p, stance} $ , $ {k}_{d, stance} $ , $ {k}_{p, swing} $ , $ {k}_{d, swing} $ , peak swing knee angle, swing-stance ratio, and thigh angular velocity threshold.

A stumble recovery controller was designed to assist the knee in the execution of the elevating strategy. Upon detection of a stumble indicated by contact with the obstacle recorded from the toe FSR, a feedforward torque pulse intended to assist the flexion and subsequent extension of the elevating response is implemented. Specifically, the pulse is a 450-ms (approximate length of elevating strategies from a previous stumble study conducted at the same walking speed (King et al. (Reference King, Eveld, Martínez, Zelik and Goldfarb2019))) spline with approximately 14.5 Nm peak torque in flexion and extension. This spline is shown in Figure 1(b), in which the flexion torque and extension torque each span the same duration (225 ms). Note that if the participant took longer than 450 ms after perturbation to foot-strike, the controller commanded zero torque until foot-strike; if the participant’s foot-strike occurred before 450 ms, then the controller entered Stance Mode, as shown in the finite-state machine in Figure 1(b).

2.2.1. Experimental protocol

Three young, healthy adult participants (one female, two males; mean height; 1.77 m, mean mass: 77.9 kg; mean age: 26 years) were recruited to participate in the exoskeleton stumble recovery experiment. Participants were screened with the following inclusion criteria: Does weight added to the shank impair gait in a manner reflective of a target population, and does a knee exoskeleton walking controller improve the impaired gait? The first inclusion criterion was used to identify candidates who might simulate an impaired population. The second criterion was used to identify candidates whose simulated impairment could be corrected with the torque limits of the device. Therefore, the intent of this screening was to ensure that the participants started with a (steady-state) walking baseline that was representative of the set of individuals who would use a lower-limb exoskeleton for walking; namely, they have (1) an impaired gait without an exoskeleton and (2) a less impaired gait when using an exoskeleton. Quantitative measures to define this inclusion criterion are given in the Data Analysis subsection.

The protocol involved sessions on two separate days, an acclimation day and a testing day. All experimental protocols were approved by the Vanderbilt Institutional Review Board, and all participants gave their written informed consent.

During the acclimation day, the knee exoskeleton was fit to the participant’s right leg. Specifically, a thigh segment, shank segment, AFO, and shoe were selected that best fit the participant, and the height of the lateral shank upright was adjusted such that the center of the exoskeleton knee joint best approximated the axis of rotation of their biological knee. Then the participant walked on the treadmill at 1.1 m/s while the experimenter tuned the walking controller. Finally, the participant was introduced to a few stumbles (with the same setup and protocol discussed subsequently, including sensory occlusion techniques) in order to be familiarized with the experimental protocol prior to testing day.

On testing day, the participant underwent a series of stumble trials using a custom obstacle perturbation system (King et al. (Reference King, Eveld, Martínez, Zelik and Goldfarb2019)). The experimental setup and protocol are pictured in Figure 2. For each trial, participants walked on a force-instrumented treadmill at 1.1 m/s. After a random number of strides (between 25 and 45), a 35 lb (16 kg) steel block was released from the obstacle delivery apparatus onto the treadmill belt such that it contacted the participant’s foot at an experimenter-defined percentage of swing phase. Participants were instructed to try to recover when the perturbation occurred and return to steady-state walking, after which the treadmill would be decelerated to end the trial. Therefore, each trial took between 30 s and 2 min, depending on the random number of strides before the stumble. Note that various sensory occlusion techniques (e.g., noise-canceling headphones and inferior vision-blocking goggles) and a cognitive distraction task (Serial Sevens) were employed to limit the expectation of the stumble. Additionally, the obstacle delivery apparatus was shown to deliver the obstacle without the perception of the participants in a previous validation study. See King et al. (Reference King, Eveld, Martínez, Zelik and Goldfarb2019) for full protocol and system details. For this experiment, perturbations were targeted to occur in the early swing phase (20–40% swing), such that elevating strategies would be employed to recover. Handrails were removed from the treadmill so they could not be used to recover, but participants wore a full-body harness to prevent contacting the treadmill in the event of a fall.

Figure 2. Stumble experiment setup and protocol. The obstacle stumble perturbation system described by King et al. (Reference King, Eveld, Martínez, Zelik and Goldfarb2019) is employed here (left). Stumbles trials were conducted for four experimental cases (right).

The stumble perturbations were introduced for four experimental cases, shown in Figure 2: (1) without wearing the exoskeleton (No Exo); (2) wearing the weighted-shank exoskeleton with the controllers disabled (Weighted); (3) wearing the weighted-shank exoskeleton with the walking controller enabled (Walk Only); and (4) wearing the weighted-shank exoskeleton with the stumble recovery controller enabled (Stumble Recovery). The No Exo case was intended to capture how an unimpaired, young, healthy adult responds to early swing perturbations, which served as the healthy control response. The Weighted case was intended to capture the response of the participant impaired by the weighted-shank exoskeleton without device assistance, which served as the baseline response of the impaired individual. The Walk Only case was intended to capture how an impaired individual responded if they were wearing an exoskeleton that assisted walking but did not have a stumble recovery feature. Finally, the Stumble Recovery case was intended to capture how an impaired individual responded if they were wearing an exoskeleton that assisted walking and also provided stumble recovery assistance when detected. Note that for both the Stumble Recovery and the Walk Only cases, weights were still attached to the shank.

The No Exo trials were conducted first. The order of exoskeleton trials (Weighted, Walk Only, Stumble Recovery) was randomized. Additionally, trials in which the left limb (i.e., the limb not wearing the exoskeleton) was stumbled were interspersed randomly; these trials were included so that not all trips would occur on the right side, in order to limit expectation of the stumble event and any anticipatory behaviors associated with knowing only the right side could be tripped. These trials were not included in the analysis, as the exoskeleton controller tested in this work was designed for the elevating limb. The participant was not informed of the case of each trial. Each participant was introduced to 3–6 trials for each case. Additionally, before and after the stumble perturbation trial sets, 60-s walking trials (without perturbation) were conducted for the No Exo, Weighted, and Walk Only cases.

2.2.2. Data collection & processing

For each trial, ground reaction forces (GRFs) were collected at 1,000 Hz via a lateral, split-belt, force-instrumented treadmill (Bertec, Columbus, USA). Full-body kinematic data were collected at 200 Hz via infrared motion capture (Vicon, Oxford, GBR), which included feet, shank, thigh, pelvis, trunk, forearm, and upper arm segments. GRF and motion capture data were filtered with a zero-phase, third-order, low-pass Butterworth filter with cutoff frequencies of 15 and 6 Hz, respectively. Inverse kinematics were computed using Visual3D (C-Motion, Germantown, USA) to estimate joint-level kinematics for each trial.

2.2.3. Data analysis

Trials were excluded from the analysis based on a few criteria. First, if the participant stepped on the obstacle or over the obstacle (i.e., missed targeting), the trial was not deemed a swing-phase perturbation and was excluded. Second, if the participants’ foot hit the obstacle substantially medially, this caused the block to rotate which is an outlier relative to remaining stumbles (i.e., perturbation impulse is altered, and in some cases block rotated enough such that it did not need to be cleared); thus, if the block rotated more than 45 degrees about the vertical axis after contacting the participants’ foot, the trial was excluded from the analysis.

For each trial, the perturbation event was identified as the instant at which the foot contacted the obstacle, which was determined via a transient peak in the anterior–posterior GRF measured by the treadmill. The swing percentage of each perturbation was estimated as the time from the toe-off immediately preceding the perturbation to the instant of perturbation, relative to the average of 20 swing times prior to the perturbation.

To determine participant inclusion in the experiment, (i.e., participants for whom adding unilateral leg weights would approximate the gait of an impaired population, and for whom the exoskeleton could correct this impaired gait), overall temporal symmetry was computed using gait information from the average of 20 strides during the walking trials for each case. Overall temporal symmetry is defined here as the ratio of the right limb’s temporal swing-stance symmetry to the left limb’s temporal swing-stance symmetry (Patterson et al. (Reference Patterson, Parafianowicz, Danells, Closson, Verrier, Staines, Black and McIlroy2008)). Temporal swing-stance symmetry is defined as the ratio of swing time to stance time for a given limb, where swing time is the time from toe-off to foot-strike, and stance time is the time from foot-strike to toe-off. In a study examining the gait asymmetry of community-ambulating stroke survivors, the normative range for overall temporal symmetry was defined as 0.9 to 1.1, mild asymmetry was defined to be 1.1 to 1.5, and severe asymmetry defined as greater than 1.5 (with a value over 1.0 indicating a preference to rely on the non-paretic limb) (Patterson et al. (Reference Patterson, Parafianowicz, Danells, Closson, Verrier, Staines, Black and McIlroy2008)). Note that gait asymmetry has been identified as a factor for predicting falls for stroke patients (Wei et al. (Reference Wei, Liu, Chang and Liu2017)). These ranges were used to determine if the participants were adequate candidates for the impairment and improvement techniques chosen for this study to represent an impaired population who might use a lower-limb exoskeleton. Specifically, if the Weighted case induced an asymmetry in the mild range, and the asymmetry was restored to the normative range with the Walk Only case, then the participant’s gait (1) was considered to be substantially impaired with the weighted-shank exoskeleton, and (2) could be corrected with the torque assistance available from the current device. Note that half of the individuals screened for this inclusion criterion passed. Two participants did not pass the first criterion when screened, as they forcibly overcame the inertial asymmetry imposed by the leg weights following acclimation. One participant just passed the first criterion, but did not pass the second, likely due to exoskeleton torque limitations.

Various outcome metrics were computed to assess the extent to which the weighted-shank exoskeleton and stumble recovery controller respectively impaired and improved the participants’ stumble recovery response. Both local (elevating limb kinematics) and global (overall fall risk) metrics were considered. These outcome metrics are diagrammed in Figure 3 along with representative trajectories. To evaluate local impairment/improvement, the range in knee angle, knee angular velocity, thigh angle, and thigh angular velocity from the instant of perturbation to the first foot-strike of the recovery limb were computed for each trial. To evaluate global impairment/improvement, trunk angle and trunk angular velocity at recovery foot-strike were computed (calculated relative to the value at perturbation). Additionally, the elapsed time from perturbation to the instant at which trunk angular velocity reversed direction from forward rotation to backward rotation (i.e., how soon induced forward angular momentum was arrested) was computed. These trunk metrics reflect the participants’ ability to restore trunk motion during the recovery. Finally, the recovery foot’s anterior–posterior center-of-mass (COM) position relative to the pelvis’ anterior–posterior COM position at recovery foot-strike (calculated relative to the value at perturbation) was computed for each trial. Trunk angle, trunk angular velocity, and foot positioning at recovery foot-strike have previously been reported to discriminate falls from recoveries (Pavol et al. (Reference Pavol, Owings, Foley and Grabiner2001); Owings et al. (Reference Owings, Pavol and Grabiner2001); Grabiner et al. (Reference Grabiner, Bareither, Gatts, Marone and Troy2012); Pater et al. (Reference Pater, Rosenblatt and Grabiner2016)) and many works cite trunk control and foot positioning as keys to successful recovery (Grabiner et al. (Reference Grabiner, Koh, Lundin and Jahnigen1993); Eng et al., Reference Eng, Winter and Patla1994; Schillings et al. (Reference Schillings, van Wezel, Mulder and Duysens2000); Forner Cordero et al. (Reference Forner Cordero, Koopman and van der Helm2003); Pijnappels et al. (Reference Pijnappels, Bobbert and van Dieën2005a)); thus changes in these metrics can be used as indicators of fall risk or improved recovery.

Figure 3. Outcome metrics to assess local and global impairment/improvement. Data from a representative comparison pair (perturbations at similar swing percentages from Case A and Case B) are plotted. The difference in each metric (Case A relative to Case B) was computed for each comparison pair for analysis in Figures 5, 6, and 7. Specifically, for Figure 5, Case A is Weighted and Case B is No Exo; for Figure 6, Case A is Stumble Recovery and Case B is Weighted; for Figure 7, Case A is Stumble Recovery and Case B is Walk Only.

To examine the extent of stumble recovery impairment, the local and global outcome metrics from Weighted trials were compared to No Exo trials. Since stumble recovery responses vary depending on when in swing phase the perturbation occurs (Eng et al. (Reference Eng, Winter and Patla1994); Schillings et al. (Reference Schillings, van Wezel, Mulder and Duysens2000); Shirota et al. (Reference Shirota, Simon and Kuiken2014b); Eveld et al. (Reference Eveld, King, Vailati, Zelik and Goldfarb2021b)), it was important to compare outcome metrics from like perturbations (i.e., occurring at the same swing percentage, or as close as possible) in order to interpret the effect of the intervention with other perturbation factors controlled. Thus, trials that were closest in swing percentage were paired using the following procedure: each successful Weighted trial was paired with a No Exo trial that was closest in swing percentage. If there were no No Exo trials within 3% swing of said Weighted trial, then that Weighted trial was not used in the analysis. Conversely, if there were more than one No Exo trial within 1% swing of said Weighted trial, then both (or more) No Exo trials were paired with the Weighted trial to comprise two (or more) comparison pairs. For each comparison pair, the difference in each outcome metric (Weighted relative to No Exo) was computed.

To examine the extent of stumble recovery improvement, the identical comparison pair approach was used for (i) Stumble Recovery relative to Weighted and (ii) Stumble Recovery relative to Walk Only.