Abduction/Adduction Assistance From Powered Hip Exoskeleton Enables Modulation of User Step Width During Walking

Using wearable robotics to modulate step width in normal walking for enhanced mediolateral balance has not been demonstrated in the field. We designed a bilateral hip exoskeleton with admittance control to power hip abduction and adduction to modulate step width. Objective: As the first step to show its potential, the objective of this study was to investigate how human's step width reacted to hip exoskeleton's admittance control parameter changes during walking. Methods: Ten non-disabled individuals walked on a treadmill at a self-selected speed, while wearing our bilateral robotic hip exoskeleton. We used two equilibrium positions to define the direction of assistance. We studied the influence of multiple stiffness values in the admittance control on the participants’ step width, step length, and electromyographic (EMG) activity of the gluteus medius. Results: Step width were significantly modulated by the change of stiffness in exoskeleton control, while step length did not show significant changes. When the stiffness changed from zero to our studied stiffness values, the participants’ step width started to modulate immediately. Within 4 consecutive heel strikes right after a stiffness change, the step width showed a significant change. Interestingly, EMG activity of the gluteus medius did not change significantly regardless the applied stiffness and powered direction. Conclusion: Tuning of stiffness in admittance control of a hip exoskeleton, acting in mediolateral direction, can be a viable way for controlling step width in normal walking. Unvaried gluteus medius activity indicates that the increase in step width were mainly caused by the assistive torque applied by the exoskeleton. Significance: Our study results pave a new way for future design and control of wearable robotics in enhancing mediolateral walking balance for various rehabilitation applications.


Abduction/Adduction Assistance From Powered Hip Exoskeleton Enables Modulation of User
Step Width During Walking Abbas Alili , Graduate Student Member, IEEE, Aaron Fleming , Member, IEEE, Varun Nalam , Member, IEEE, Ming Liu , Member, IEEE, Jesse Dean , and He Huang , Fellow, IEEE Abstract-Using wearable robotics to modulate step width in normal walking for enhanced mediolateral balance has not been demonstrated in the field.We designed a bilateral hip exoskeleton with admittance control to power hip abduction and adduction to modulate step width.Objective: As the first step to show its potential, the objective of this study was to investigate how human's step width reacted to hip exoskeleton's admittance control parameter changes during walking.Methods: Ten non-disabled individuals walked on a treadmill at a self-selected speed, while wearing our bilateral robotic hip exoskeleton.We used two equilibrium positions to define the direction of assistance.We studied the influence of multiple stiffness values in the admittance control on the participants' step width, step length, and electromyographic (EMG) activity of the gluteus medius.Results: Step width were significantly modulated by the change of stiffness in exoskeleton control, while step length did not show significant changes.When the stiffness changed from zero to our studied stiffness values, the participants' step width started to modulate immediately.Within 4 consecutive heel strikes right after a stiffness change, the step width showed a significant change.Interestingly, EMG activity of the gluteus medius did not change significantly regardless the applied stiffness and powered direction.Conclusion: Tuning of stiffness in admittance control of a hip exoskeleton, acting in mediolateral direction, can be a viable way for controlling step width in normal walking.Unvaried gluteus medius activity indicates that the increase in step width were mainly caused by the assistive torque applied by the exoskeleton.Significance:

I. INTRODUCTION
H UMAN walking is one of the most critical activities toward independence in daily life.However, bipedal walking is an inherently unstable activity [1].To overcome that instability humans need to continuously control their balance in mediolateral (ML) and anteroposterior (AP) directions to achieve smooth and stable walking [2].It has been demonstrated that individuals can maintain balance during walking against perturbations in the AP direction largely via passive dynamics of the human body, while active motor control is required for managing the perturbations occurring in the ML direction [3], [4].When sensorimotor integration or muscle strength is disrupted, as observed in older adults, post-stroke patients, or people with multiple sclerosis, the ability to maintain mediolateral stability is usually negatively impacted [5], [6].Loss of balance in the ML direction can lead to falls and serious injuries [7], [8].Therefore, interventions or technologies that can address balance control in ML direction in walking are needed.
People maintain balance in the ML direction during gait via several strategies [9], including active control of step width [10].Rankin et al. [11] previously demonstrated that step width behavior during walking is likely the result of sensorimotor integration of stance leg position relative to the position of the center of mass on a step-by-step basis.It has also been shown that individuals primarily use the hip joint to modulate step width during gait, where the gluteus medius (GM) muscle activation is the most correlated to step-by-step behavior [12], [13].In that way, hip joints can also influence the center of mass position and velocity to achieve a stable gait [14].Humans also modify their step width behavior under various contexts that challenge balance to reduce the risk of falls.For example, individuals increase their step width when larger gait perturbations are expected to happen [15].Increasing step width also alleviates the consequence of imprecise previous foot placement (i.e., step width) during gait [4].Impairment to any of these components of step width modulation (e.g., GM weakness, impaired sensorimotor integration) hinders step width modulation as a feasible resource for individuals to maintain ML balance during walking [16].
Poor gait stability in several populations has been linked to changes in step width behavior.Older adults tend to walk with wider and more variable steps [14].Similarly, individuals with lower limb amputation or post-stroke patients exhibit a gait with larger step width compared to non-disabled peers [17], [18], [19], [20].It has been postulated that this behavior is a result of poorer sensorimotor integration, where individuals after stroke fail to accurately integrate the position of their stance limb and pelvis to then place their swing foot appropriately [21].Despite increased metabolic cost [22], [23], these populations prefer to walk with wider steps, implying the prioritization of stability and balance during the gait.This relationship between step width behavior and gait stability among several populations makes modulation of step width a potential target for assistive or rehabilitative systems to improve balance.
Various engineering systems have been developed to control the step width for clinical populations for study of balance control mechanisms or rehabilitation purposes.External, lateral stabilization tools have been developed by different groups to investigate step width and its variability [14], [24].Pennycott et al. [25] decreased individuals' step width by a robot-driven gait orthosis aimed to be used for balance rehabilitation of neurologically impaired people.Heitkamp et al. [26] directly modulated foot placement of post-stroke patients during gait through a cable-actuated force-field to investigate the potential assistive and rehabilitative effects.They demonstrated success in improving step-by-step foot placement using an assistive force-field for guided foot placement in post-stroke patients.Additionally, by using a resistive force field (i.e., moving patients' foot placement away from the appropriate location) participants improved step width behavior once the force-field was removed.These efforts have demonstrated the potential to successfully modulate stepwidth behavior; however, these immobile test beds are limited in their potential for use outside of the laboratory.
Another emerging technology to assist walking balance is wearable robotic powered exoskeletons [27].Since hip joints are important for gait stability [9], [13], several hip exoskeleton systems have recently been developed and tested.Monaco et al. [28] designed a novel control algorithm governing active flexion/extension hip exoskeleton to prevent elderly people and transfemoral amputees from falling after perturbations.A unified active control framework of 2-DOF hip exoskeleton acting in AP direction has been proposed by Qiu et al. [29] for both walking and balance assistance.They were able to show that the proposed system was able to mitigate push and pull perturbations.Contrary to hip exoskeletons acting in the AP direction, there are a very limited number of hip exoskeletons available that demonstrated balance assistance in the ML direction during the gait.Zhang et al. [30], [31] showed the feasibility of the proposed 4-DOF hip exoskeleton that reacts to standing balance perturbations and produces a compliant guidance force.Chiu et al. [32] developed a bilateral hip exoskeleton emulator to further investigate potential ways to assist users' behavior during walking or perturbation recovery.Although exciting, the goal of these studies was to use hip exoskeleton control to reactively assist to the recovery of walking stability after perturbations in gait.There have been limited number of studies to explore the potential of wearable robots to proactively assist the modulation of step width in natural strides for improved ML balance in regular walking.Thus, many research questions exist before enabling proactive step width manipulation via a hip exoskeleton, e.g., what should be the exoskeleton control and configuration?How do human wearers respond to the exoskeleton's action?
The objective of this study is to address these knowledge gaps before developing closed-loop control of wearable robots to enable step width regulation.In this study, we presented an active hip exoskeleton that powers hip abduction/adduction bilaterally via admittance control.By adjusting admittance control parameters during natural walking, we investigated the human step width response to our active hip exoskeleton.We showed success of our powered ML hip exoskeleton and control in increasing and decreasing the step width during walking without causing a change of step length in healthy individuals.In addition, we examined the activity of hip abductors (i.e., the gluteus medius) to understand the human's hip ML control efforts in changing of step width in the experiments.The results of this study implied the promise of using an active abduction/adduction hip exoskeleton and admittance control parameter change to modulate step width in natural strides.The outcome also informs the high-level control design of hip exoskeleton that can regulate the step width for proactive gait balance interventions in the future for individuals with balance deficits.

A. Participants
Ten non-disabled, neurologically intact adults (6 males, 4 females; age = 25.2 ± 4.6 years; mass = 70.8± 11.6 kg; height = 1.72 ± 0.06 m; mean ± s.d.) participated in this study.The study was conducted with the approval of the Institutional Review Board of the University of North Carolina at Chapel Hill (133028), with all participants providing informed consent.

B. Equipment: ML Hip Exoskeleton
A fully powered and compliant 2-DOF robotic ML hip exoskeleton developed by Neuromuscular Rehabilitation Engineering Laboratory (NREL) was used in this study.During the design process, we utilized a modular motor design and employed a low-level admittance control mechanism, which is identical to the one employed in [33].The device was equipped with two brushless EC flat motors (Maxon, USA) acting on a hip in the frontal plane (Fig. 1(a)).Passive hinges enabled free movement of the legs in the sagittal plane.The exoskeleton system uses a harmonic gearbox (CSD-20-100-2A-GR-SP674, Harmonic Drive, MA, USA) providing a 100:1 transmission ratio.The hip exoskeleton can produce an output torque of 34 N•m, and a peak torque of 57 N•m.The hip abduction/adduction angle was measured by the encoder (Maxon, USA) attached to the motor.The interaction torque (τ int ) between human and exoskeleton was measured by the load cell situated on the arm connecting the actuator to the limb.The hip exoskeleton was worn around the waist and fastened to the thighs of the participants through cloth cuffs as shown in Fig. 1(a).

C. Admittance Controller for Hip Exoskeleton Actuators
To ensure safe physical human-exoskeleton interaction, we used admittance control [34] for each actuator to power hip abduction/adduction, as shown in Fig. 1(b).Admittance controllers used virtual impedance parameters, such stiffness (K), damping (B), equilibrium angle (θ eq ), and inertia (M) to model the dynamics of human-exoskeleton system.Admittance control, according to the observed interaction torque (τ int ), determines the angular velocity reference signal for the low-level controller to produce required motor torque.(1) describes the dynamics for interaction torque τ int (t), where θ represents the measured exoskeleton hip abduction/adduction angle, and θ(t) and θ(t) are its derivatives standing for motor velocity and acceleration respectively: The control was implemented on a real-time PC using Ether-Cat (TwinCAT 3.1) protocol executing at 1000 Hz.
In this study, we used two parameters of the admittance controllers (K and θ eq ) to study the human step width response to different exoskeleton mechanics.M and B were fixed at 0.2 and 0.3 values respectively.The equilibrium angle was utilized to define the direction of the applied torques in terms of abduction or adduction of the hip joint as shown in Fig. 1(b).We aimed to explore a large range of step-width behavior, both in the abduction and adduction directions, thus we selected an equilibrium angle beyond the widest (θ eq,Ab ) and narrowest (θ eq,Ad ) ML hip joint excursions that shown in literature [35].
Therefore, we opted to use double the reference value (∼ 7°) to ensure that the necessary torque would be applied for both increasing and decreasing the step width of hip exoskeleton users.Once these angles were established, they were set to a constant value (15°) for the whole gait cycle to systematically investigate various stiffness values.Essentially, we modified different stiffness coefficient (K) for either direction to produce different interaction behavior.
When K is equal to zero, i.e., zero impedance (ZI) mode, the ML hip exoskeleton aims to match the hip joint movement of the human wearer without any extra torque applied.When stiffness K increases, the abduction/adduction torque applied in the direction of the predefined equilibrium angle increases.

D. Experimental Setup
A motion capture system (VICON, Oxford, U.K.), consisting of 12 cameras sampled at 100 Hz, was used to track foot positions while participants walked on an instrumented treadmill (Bertec Corp. Columbus, OH, USA).As shown in Fig. 1(a), participants wore a safety harness attached to an overhead rail, which did not provide body weight support.Twelve light-retroreflective markers were placed on the participant's feet (the second toe calcaneus, first and fifth metatarsals, lateral and medial malleolus, and heel of each foot).Ground reaction forces (GRFs) were recorded at a sampling rate of 1000Hz by an instrumented split-belt treadmill synchronized with the motion capture system to detect the gait events.Detected gait events were visually inspected during the processing of the data to avoid missing heel strikes and toe offs.
It has been demonstrated that the gluteus medius (GM) muscles play a crucial role in modulating step width [36], [37] when humans voluntarily adjust their step width.Additionally, these Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
muscle groups exhibit the highest level of activity during ML foot placement [13].These findings served as the foundation for our investigation into the behavior of GM muscles.Thus, electromyographic (EMG) signals of the GM muscles during the treadmill walking was measured bilaterally by bipolar surface electrodes (Motion Lab Systems, Louisiana, USA) sampled at 1000Hz.SENIAM [38] guidelines were used for the placement of the electrodes after cleaning the skin surface with alcohol.

E. Experimental Protocol
We investigated the effects of the active torque application of the hip exoskeleton in the abduction/adduction direction for the entire gait cycle on human's step width.The decision to apply abduction/adduction torque throughout the entire gait cycle was based on research findings that demonstrated the role of the GM in controlling step width both in stance and swing phases [11], [37].The experiment was carried out in one day and divided into three sessions: acclimation, abduction assistance, and adduction assistance.
To ensure efficient human-device interaction, first, all participants were asked to walk on the treadmill while the hip exoskeleton was in a zero impedance (ZI) mode at their selfselected walking speed.The self-selected walking speed was determined by gradually increasing the treadmill speed in increments of 0.1 m/s until the participant verbally confirms that it was their preferred speed.The participant's self-selected treadmill speed was set constant throughout all sessions to avoid any impact of walking speed on step width [39].
We allowed participants to acclimate walking with the hip exoskeleton in the ZI mode for 2 walking trials (3 minutes each).Then participants were asked to walk for three trials, wearing the device, for both abduction and adduction assistance conditions.Each abduction or adduction walking trial was 5.5 minutes long with the ZI mode control for the first thirty seconds.In each trial, participants experienced 5 levels of stiffness K: 0, 20, 40, 60, and 80 N•m/rad.We applied each stiffness level for 1 minute of the walking trial for both left and right joints and randomized the order of K levels for each individual trial.Ultimately, each participant experienced the same level of K three times for each of the abduction or adduction assistance sessions.These levels of K were chosen empirically based on pilot study work.

F. Data Processing, Evaluation Metrics, and Statistics
The marker positions were low-pass filtered by a fourth-order Butterworth filter with a cutoff frequency of 10 Hz.GRF signals were low-pass filtered with a cutoff frequency of 25 Hz.Heel strike events were identified using the GRF of each limb with a threshold of 30 N [40].We computed step width and step length based on maker positions.The step width was calculated as a mediolateral distance between heel markers at each heel strike [36].Similarly, step length was defined as an anteroposterior distance between heel markers at each heel strike.Only the last 40 seconds of the step width data relevant to each K level was analyzed for the steady-state behavior to avoid the transient effects.
The calculation of linear correlation coefficient (Pearson's r) is conducted using JMP, Version 17.0.0(SAS Institute Inc., USA).The data used for the analysis consists of the averaged steady state step width values for all participants across the modulated stiffness values.
We were also interested in investigating how quickly humans react to the K change when this change led to significant step width modification.First, we selected the K changes that led to significant step width adjustment.To quantify how quick is the human reaction, we defined the reactive strike number right after a K change using Cusum Control Chart (CuSum) method [41], [42].Using cusum function in MATLAB (MathWorks, USA) allowed us to identify the first strike that shows the significant shift of mean from the step width derived from the strikes before K change.The number of these identified strikes after the K change was defined as the reactive strike number.The details about cusum algorithm and its implementation is shown in Appendix.
To evaluate muscle activity, we computed the EMG envelope of bilateral GM.EMG envelope was obtained by filtering raw EMG signals with a high-pass filter (cutoff at 20Hz, 4th order), followed by rectification and low-pass filtering (cutoff at 50Hz, 4th order).We then segmented EMG for both the left and right GM based on the timing of left and right heel strikes.We normalized EMG envelope in time for each gait cycle (heel strike to heel strike of the same limb).To compare EMG activity across participants, we normalized EMG data based on our zero-impedance condition.Specifically, we averaged EMG data from all zero impedance trials for a given participant and took the maximum value from this averaged data for each limb muscle.We then divided all EMG data by this maximum value for each muscle and repeated this procedure for each participant [43].We evaluated the participants' level of effort in hip abduction/adduction control during the stance phase by integrating EMG activity of GM muscles from 0-40% of the gait cycle [36].In this study we focused on the analysis of the stance phase GM EMG activity because GM activity is the highest in stance phase for pelvic stabilization [11], [44].We averaged the integrated EMG activity for the last 40 gait cycles from each impedance condition.We then averaged this number across the repetitions of impedance conditions so that each condition contained one averaged value for each participant and used later for statistical analysis.
For the abduction and adduction assistance sessions, we performed one-way repeated measures ANOVA to determine if K-levels significantly influenced the step width or stance phase GM activity.p values less than 0.05 were interpreted as significant for all analyses.Post-hoc Tukey tests were performed if ANOVA analysis showed significance.

A. Step Width and Step Length With a Steady K Value
Fig. 2 shows the influence of stiffness K of active hip exoskeleton abduction or adduction assistance on step width and step length.In average participants walked with wider steps during the abduction assistance session due to the applied abduction Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.torque to both limbs over the whole gait cycle (Fig. 2(a)).The step width increased almost linearly (Pearson's r = 0.665, p < 0.001) with the stiffness K in abduction assistance before saturation reached with K = 60 Nm/rad (i.e., participants produced the maximum step width).One-way ANOVA showed that change of K value in hip abduction significantly influences step width across participants, compared to ZI mode (p < 0.001).Furthermore, post-hoc Tukey tests showed that the average step width relevant to K = 20 condition was significantly smaller than the K = 60 and K = 80 levels (p < 0.001).Nevertheless, the calculated step length results did not show any statistical significance for various values of K, as shown in Fig. 2

(c)(blue error bars).
When the stiffness K of hip exoskeleton, acting in the adduction direction, increased, in average participants walked with narrower steps over the whole gait cycle as shown in Fig. 2(b).Due to the limited range of step width for human to walk with narrow width (compared to the range when walking wider), the step width hit the saturation value (smallest step width that human can walk comfortably) when K was only 20.Therefore, the step width did not reduce significantly when K was greater than 20.Nevertheless, statistical analysis revealed that, in comparison to walking in ZI mode (K = 0), in average participants walked with narrower steps with a stiffness value set more than zero (p < 0.03).Finally, similar to the step width increase session results, participants did not change their step length significantly for different stiffness K, as shown in Fig. 2(c)(red error bars).3(a)) and adduction (Fig. 3(b)) assistance when the stiffness K was changed from ZI mode (K = 0).As in Fig. 3, the step width reaction ocurred almost immediately after the stiffness changed.Using the CuSum method discussed in Section II, we quantified number of reactive heel strikes when stiffness K changed from 0 to other studied levels (Fig. 4).In general, when stiffness K increased Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.from the zero, it only took less or equal to 4 strikes in the transient reaction for step width to reach steady state value.Note that we only focused on the stiffness change from K = 0 because the step width showed significant difference in steady state when K increased from 0 in this study (Fig. 2

C. GM EMG Activity
The GM activity did not vary significantly with abduction and adduction assistance over changed stiffness levels for both limbs.Fig. 5 shows the averaged pattern of the gluteus medius activity for all participants during abduction or adduction assistance.Only the left side is shown as an example.Table I summarizes bilateral integrated GM during the stance phase.Both results showed that no significant change of GM activation pattern was identified across studied K-Level conditions for either abduction or adduction assistance.

IV. DISCUSSION
In this study, we presented a novel way to modulate human step width during normal walking via a hip exoskeleton, powering abduction/adduction by admittance control.Our hip exoskeleton application is fundamentally different from existing exoskeleton design and control.The vast majority of existing exoskeletons have been focused on assisting sagittal plane motions for cyclic stepping [28], [33], [45], [46], while ML balance is ensured by human users (e.g., using crutches) [47].Although very limited studies have tried to use exoskeletons to react to a balance perturbation (i.e., control after a perturbation occurs), to our knowledge, no wearable robot approach yet exists to proactively modulate the step width for enhanced ML walking stability.Hence, in this study, we not only presented a new configuration and control of a bilateral hip exoskeleton, potentially for proactive ML balance augmentation, but also contributed new knowledge that adjusting the stiffness in the abduction/adduction admittance control of a hip exoskeleton can successfully modulate the human user's step width without influencing the step length during normal walking.This result implies that stiffness tuning can be a viable option for controlling step width, instead of relying on torque-tracking approaches that require a high-level torque reference command.
In this study, we examined how humans respond to varied hip exoskeleton control parameters applied in ML direction.We observed an almost linear trend between altered stiffness and average step width for abduction until the step width reached a limit.Such a linear response in step width enables design of simple, linear controllers for closed-loop control of human step width via adjusting stiffness in admittance control during walking in the future.Given that hip abduction/adduction assistance did not change the step length, the saturation of the step width was likely due to the biomechanical limits for participants to increase or decrease the step width with invariant step length during normal walking.Nevertheless, for adduction sessions, we did not clearly observe the linear relationship between stiffness K and step width.This is likely because the adjusting step of K-value (i.e., 20 Nm/rad) was too large while the biomechanical limit for reduced step width was small.A higher resolution of K-value change (e.g., increasing K by 5 Nm/rad increments) could potentially be used to further explore whether the control stiffness and step width maintains a linear relationship within the step width limit for hip adduction assistance.Finally, the average step length of the hip exoskeleton users did not change for either abduction or adduction assistance at various K levels, supporting previous studies that suggest independence between anterior-posterior and mediolateral foot placements [10].These results suggest that separate controllers can be designed for 4 DOF hip exoskeletons controlling step width and length independently.
Another factor that makes our presented approach for augmenting ML balance practical is that we did not observe gait disruptions (i.e., trips or use of handlebars) among all participants during change of exoskeleton control K, even when the K value changed significantly (see supplementary video for reference).Such human response implied that we can adjust stiffness K during normal walking to achieve step width modulation safely.Participants smoothly adapted to the change in K and stabilized at the new step width within 4 strikes for both the abduction and adduction sessions.This response time may be inadequate to enable precise strike-to-strike step width control (or foot placement), which might be needed when walking on irregular ground surfaces, e.g., a rocky trail.Nevertheless, the response time is sufficient to proactively regulate the step width for enhanced ML balance during walking exercise.However, in this study, we only considered using stiffness K to change the step width.In addition, K value change is applied to the entire gait cycle.Further research is needed to understand if the ML hip exoskeleton can change step width within a single step by adjusting other admittance control values (such as equilibrium

TABLE I RESULTS OF NORMALIZED STANCE PHASE GM EMG ACTIVITY
position), changing stiffness K only during a certain phase of the gait, or using closed-loop feedback control.
The activity of GM during the abduction assistance indicated that increases in step width behavior were primarily due to changes in exoskeleton assistance level rather than changes in participant behavior.A previous study has demonstrated that volitional increases in step width are primarily generated by increased GM activity during the stance phase [36].Nevertheless, our result indicated that, despite increased step width, GM activity did not change.This finding suggests that ML hip exoskeletons can provide assistive support for step width increase without imposing additional demand on the GM muscles.It also shows that ML hip exoskeleton's abduction assistance can be used with minimal resistance from non-disabled users.For the adduction assistance mode, GM activity did not change either.However, research showed that non-disabled individuals walking with narrower steps decreases EMG of GM [36].This finding suggests that hip exoskeleton users may resist narrowed step width.This is potentially caused by perceived threat to their balance due to reduced base of support in ML direction and potential tripping by their own limb.To further investigate the cause of this resistive response to exoskeleton abduction torque, we might need to systematically examine the human user's subjective response on the sense of efforts and balance confidence.
The presented hip exoskeleton, control, and human response may lead to many exciting applications.First, our exoskeleton design and study results could likely be utilized to enhance ML balance control in populations with declined sensorimotor functions or hip muscle weaknesses, such as post-stroke patients [48], older adults [49], or individuals with transfemoral amputations [50], [51].It is essential to emphasize that the structure of our hip exoskeleton enables the implementation of alternative control schemes, offering flexibility in adjusting control parameters and their timing.Additionally, the capability to control the motors unilaterally presents a potential opportunity to address the asymmetrical gait observed in certain populations mentioned earlier.Our hip exoskeleton may also be used as an assistive or therapeutic device during their gait training in clinics or home in the future to address their walking balance.The proposed hip exoskeleton and control can be employed as a research platform to study ML balance control mechanisms.As a research tool, it can be used to further study the connection between step width modulation and balance to better comprehend the neuromechanical aspects of motion control during walking [52] inside or outside of the lab environments.These will be our future research directions.
Finally, besides the limitations and future research discussed above, other limitations were also identified in our presented study.First, we carried out the study on a treadmill with a fixed speed and lab environment.It would be interesting to test the Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
human response when walking speed varies or over level ground.In addition, the research focuses on human response in step width, step length, and GM activity only, we did not measure other biomechanics parameters in gait, such kinematics and kinetics of knee and ankle.Further investigations are required to test additional muscle groups, in addition to the gluteus medius, to determine whether compensation by other muscle groups has occurred or not.More systematic analysis of human biomechanics while using our hip exoskeletons is needed in the future.

V. CONCLUSION
In this study, we successfully demonstrated the ability of our hip exoskeleton and its admittance control acting in the mediolateral direction to modulate step width in non-disabled participants.We showed that the step width of the users can be directly modulated by changing the admittance control parameters, namely stiffness K, once the added torque direction (either abduction or adduction) was defined by fixing the equilibrium angle (θ eq ) in the controller.After the stiffness K increased from zero to a studied value, we found that human participants started to react immediately after a K change and showed significant step width change within 4 strikes.In addition, besides the change of step width, the immediate stiffness change did not disrupt the gait and step length.When examining the activity of bilateral gluteus medius, which plays an important role for ML stability in gait, we found the activity does not show significant difference regardless of the level of stiffness value (and therefore torque applied).This observation indicated that the increase in step width was caused mainly by the exoskeleton rather than human efforts.This study outcome may pave a new way for designing closed-loop control of step width and foot placement in ML directions via our hip exoskeleton for assistive or rehabilitation purposes in the future.

APPENDIX
The CUSUM algorithm is used to detect small incremental changes in the mean of a process.Given a sequence s 1 , s 2 , s 3 , . . ., s n with target average m s and target standard deviation σ s , upper (U i ) and lower (L i ) cumulative process sums are defined as: The variable n is the number of standard deviations from the target mean that makes a shift of mean detectable.A process violates the CUSUM criterion at the sample s j if it obeys U j > cσ s or L j < −cσ s .
In this study, we defined the sequence as the set of step width measurements taken before the K changed (i.e., the 20 consecutive heel strikes leading up to the change), merged with all the step width data corresponding to a new K value.We defined the mean and standard deviation of the step width data samples taken before the K change as m s and σ s .To set the control limit c for the subsequent analysis, we chose a value of 3, and set n to 1.

Fig. 1 .
Fig. 1.(a) ML Hip Exoskeleton applied abduction/adduction torque during the whole gait cycle.Participants were instructed to walk comfortably.(b) Admittance control diagram used to assure safe human-device interaction.The equilibrium angle parameter is used to define the direction of the applied torque according to modulated stiffness values.

Fig. 2 .
Fig. 2. Influence of stiffness K of hip exoskeleton abduction (blue) and adduction (red) assistance on step width and step length.(a) Step width increased across K levels for abduction assistance mode.(b) Step width decreased across K levels for adduction assistance mode.(c) Step length did not change for the hip exoskeleton assistance regardless the assistance direction and applied stiffness.The data points represent the mean and error bars show the standard error mean, and asterisks ( * ) indicate significant (p<0.05)results of post-hoc tests.

Fig. 3
Fig. 3 shows the transient behavior of the participants in terms of the step width change at the moment of a stiffness change.Stiffness values were modified at the 20th heel strike in the figure.Normalized and averaged step width data across participants were shown for both abduction (Fig.3(a)) and adduction (Fig.3(b)) assistance when the stiffness K was changed from ZI mode (K = 0).As in Fig.3, the step width reaction ocurred almost immediately after the stiffness changed.Using

Fig. 3 .
Fig. 3. Transient behavior of normalized step width across changes of stiffness K from ZI model for (a) abduction and (b) adduction assistance.

Fig. 4 .
Fig. 4. Transient responses of participants' step width to modulated stiffness levels is shown for abduction assistance and adduction assistance.Error bars represent the standard error mean.

Fig. 5 .
Fig. 5. Modulated step width did not influence the activity of the gluteus medius.The average pattern of left leg GM for (a) abduction and (b) adduction assistance modes are plotted from heel-strike to heel-strike across all participants.