1636 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 6, JUNE 2013

Effects of Robotic Knee Exoskeleton on HumanEnergy Expenditure

Andrej Gams∗, Tadej Petric, Tadej Debevec, and Jan Babic

Abstract—A number of studies discuss the design and controlof various exoskeleton mechanisms, yet relatively few address theeffect on the energy expenditure of the user. In this paper, we dis-cuss the effect of a performance augmenting exoskeleton on themetabolic cost of an able-bodied user/pilot during periodic squat-ting. We investigated whether an exoskeleton device will signifi-cantly reduce the metabolic cost and what is the influence of thechosen device control strategy. By measuring oxygen consumption,minute ventilation, heart rate, blood oxygenation, and muscle EMGduring 5-min squatting series, at one squat every 2 s, we show theeffects of using a prototype robotic knee exoskeleton under threedifferent noninvasive control approaches: gravity compensationapproach, position-based approach, and a novel oscillator-basedapproach. The latter proposes a novel control that ensures synchro-nization of the device and the user. Statistically significant decreasein physiological responses can be observed when using the roboticknee exoskeleton under gravity compensation and oscillator-basedcontrol. On the other hand, the effects of position-based controlwere not significant in all parameters although all approaches sig-nificantly reduced the energy expenditure during squatting.

Index Terms—Adaptive control, exoskeletons, metabolic cost,oscillators, squatting.

I. INTRODUCTION

W EARABLE robots, also known as the exoskeleton de-vices, are being developed to either augment the abil-

ities of able-bodied humans or to substitute/improve conditionof people with impaired physical abilities [1]. Such deviceshave been the subject of research for many years and variousmechanism design and control issues have been discussed inthe literature, see, for example, a comprehensive overview byDollar and Herr [2]. Whether augmenting abilities or improv-ing disabilities, the purpose of assistive exoskeleton devices isto provide useful mechanical power by means of synchronousoperation with the person that is wearing it [3].

The human–robot interface is crucial for a safe and com-fortable interaction with a wearable robotic device [4]. Various

Manuscript received October 1, 2012; revised December 5, 2012 and January8, 2013; accepted January 12, 2013. Date of publication January 17, 2013; dateof current version May 15, 2013. Asterisk indicates corresponding author.

∗A. Gams is with the Department of Automation, Biocybernetics andRobotics, Jozef Stefan Institute Jamova cesta 39, 1000 Ljubljana, Slovenia andwith the Biorobotics Laboratory, Ecole Polytechnique Federale de Lausanne,Lausanne, Vaud, Switzerland (e-mail: andrej.gams@ijs.si).

T. Petric, T. Debevec, and J. Babic are with the Department of Automa-tion, Biocybernetics and Robotics, Jozef Stefan Institute Jamova cesta 39,1000 Ljubljana, Slovenia (e-mail: tadej.petric@ijs.si; tadej.debevec@ijs.si;jan.babic@ijs.si).

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TBME.2013.2240682

strategies for detecting user intentions were proposed. Highlyinvasive strategies, that detect user intentions through the centralor peripheral nervous system were demonstrated on primates tocontrol a robotic manipulator [5]. Slightly less invasive are ap-proaches based on surface electromyography (EMG) [6]. Theuser intentions can be correlated with his/her EMG data usingeither model-based approach [7] or model-free approach [8].EMG was also used for the control of the HAL-5 exoskele-ton device in prediction of the motion [9] and for adjusting theimpedance around the knee joint [10]. An effective fuzzy-neurocontroller to automatically control an exoskeletal shoulder robotwith EMG signals of the human shoulder muscles was proposedin [11]. Petric et al. [12] show how EMG measurements can beapplied for the control of periodic tasks.

Noninvasive methods, on the other hand, use different forceand contact sensors to control the device [2]. Several controlmethods have been applied using noninvasive methods, for ex-ample, [13] discusses a controller which uses inverse dynamicsof the exoskeleton as a positive feedback controller so that theloop gain approaches unity, but requires a very accurate modelof the system. Banala et al. [14], on the other hand, proposejust gravity compensation. Ronsse et al. [1], [15] propose amethod based on adaptive frequency oscillators. The method isonly applicable for periodic tasks. In this paper, we propose andevaluate a similar yet considerably modified method of control-ling a robotic knee exoskeleton.

Simultaneous frequency extraction and waveform learning al-low filtering of a measured signal with predicting and estimatingthe output in real time, without delays between the input and theoutput. Few systems allow simultaneous frequency extractionand waveform learning [16]. For example, Righetti et al. [17]propose a system with pool of adaptive frequency oscillators in afeedback loop. A similar approach was used for the first layer ofa two-layered movement imitation system [16]. A novel designof the first layer introduced in [18] uses a single oscillator fol-lowed by a complete Fourier series approximation with a built-inalgorithm to determine the Fourier coefficients. In the paper, wetest the applicability and the effectiveness of this approach on arobotic knee exoskeleton when performing periodic squatting.As a measure of effectiveness, we compare the metabolic costof an able-bodied user for the given task without and with assis-tance of the robotic knee exoskeleton. Additionally, we compareour approach with two other control approaches: gravity com-pensation approach and position control approach.

Metabolic cost of the user is one of the possible criterions forthe effectiveness of an assistive device. As stated in [2], a com-parison of metabolic power between performing a task with orwithout an exoskeleton is a good determinant as to whether there

0018-9294/$31.00 © 2013 IEEE

GAMS et al.: EFFECTS OF ROBOTIC KNEE EXOSKELETON ON HUMAN ENERGY EXPENDITURE 1637

is any energetic advantage of using the device. Other criterionsexist, for example, [14] used EMG to evaluate the performanceof a leg orthosis, while [19] used it to measure the effectivenessof an elbow exoskeleton. Changes in oxyhemoglobin and de-oxyhemoglobin concentration in the working muscles have alsobeen proposed [20].

The review by Dollar and Herr [2] also states that surpris-ingly few instances of qualitative evaluation of the effectivenessof the exoskeleton have been reported. Walsh et al. [21] es-timated the metabolic cost of a quasi-passive leg exoskeletonfor load carrying augmentation. Even though the exoskeletonlargely supported the payload during the single support phaseof walking, a 10% increase of the transport energy cost wasnoted compared to a conventional loaded backpack. The addedmass and the kinematic constraints imposed on the wearer wereconcluded to be the dominant cause. While the exoskeletondevice in [21] was mobile, several therapeutical and not stand-alone devices exist, i.e., the Locomat [22] or the Lopes [23]exoskeleton. A study of the energy expenditure while walkingon a treadmill and using the Locomat for assistance by Kreweret al. [24] showed an increase in the oxygen consumption, whichis one of the main physiological parameters associated with themetabolic cost. Ronsse et al. [1] published results of a study onthe Lopes nonportable exoskeleton. In a detailed study aboutthe effects of various levels of assistance from the device on themetabolic cost, they also reported an increase in the metaboliccost of the participants.

On the other hand, powered ankle exoskeletons have beenshown to reduce the metabolic cost of walking [25]. Similarly,Lenzi et al. [26] have shown a reduction of muscle effort us-ing a stationary powered exoskeleton for hip activation. Singlejoint exoskeletons, such as the ankle and knee exoskeletons,focus on the muscles that provide significant parts in the totalmechanical power needed in everyday activities, such as walk-ing, ascending, and descending stairs and getting up from achair [27], [28], [29]. Several notable knee exoskeletons andwearable orthoses were developed in the past [3], for example,Yobotics, Inc. from Cincinnati, Ohio developed a robotic kneeexoskeleton that assists in stair climbing and squatting duringthe load-carrying tasks [30]. Tibion Corporation from Californiadeveloped a knee orthosis for rehabilitation [31].

The goal of our research was to investigate: 1) whether an ex-oskeleton device effectively contributes to the energetics of thehuman motion; and 2) whether the choice of the control strat-egy affects the physiological responses. Three control strategieswere implemented and evaluated on a knee exoskeleton thatwas specifically built for this research study. In order to quanti-tatively determine the contribution of the exoskeleton, we mea-sured and determined heart rate, oxygen consumption, minuteventilation, blood oxygen saturation, EMG, and metabolic costof the participants wearing the exoskeleton.

The paper is organized as follows: In Section II, we describethe study design. In Section III, we describe the evaluated ex-oskeleton control approaches with an emphasis on the oscillator-based approach. Section IV gives the results while Section Vdiscusses on the possibilities and impact of the given study. Aconclusion is given in Section VI.

TABLE IBASELINE PARTICIPANT CHARACTERISTICS

II. STUDY DESIGN

We performed a randomized cross-over designed study thatcompared the effects of the robotic knee exoskeleton on themetabolic cost of the squatting under three different controlapproaches.

A. Participants

Seven healthy young males volunteered to participate in thestudy. The participants were informed about the procedure be-fore the experiments and the free informed consent was obtainedfrom all. Table I shows baseline participant characteristics.

B. Procedures

The participants were asked to perform four 5-min seriesof periodic squatting. After a referential series without the ex-oskeleton (free), but with all other sensors and markers attached(see Section II-C), the participants performed the other threeseries, each using a different control approach (gravity compen-sation, position based, oscillator based) in a randomized order.The randomization of the knee exoskeleton series allowed forappropriate comparison between the tested control approaches.

Each participant was asked to individually warm up beforethe start. The squatting series began with a 2-min rest period,used for the measurement of the resting cardiorespiratory val-ues. The participant then performed periodic squatting for 5 min.Squatting was defined as going from a fully extended posture(0◦ in the knees) to a low yet still comfortable squat (at leastslightly over 90◦), and back. The participant was asked to main-tain the frequency of the squatting at 1 squat/2 s, where either ametronome or the predefined movement of the knee exoskele-ton (see Section III-B) was used for the reference. In order toprevent possible injuries from falling on the back, because thedevice was attached to the frame in the ankles, the participantswere asked to put their hands on top of a railing (see Fig. 1)and grab the railing in case of emergency. The participants wereinstructed before the experiments and constantly reminded dur-ing the experiment not to grab the bar but to only lightly touchit with their palms open. Further, to exclude the possible in-fluence of the bar and the light touch phenomenon [32] on themeasurements, the participants had to lightly touch the bar in allexperiments, including when they performed squatting move-ments without the help of the knee exoskeleton (free). Eachsquatting series ended with a 2-min rest period in an upright po-sition to measure the recovery kinetics of the cardiorespiratoryvalues.

Before the next squatting series the participants rested forat least 1.5 h after the first series (free) and at least an hour

1638 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 6, JUNE 2013

Fig. 1. Kinematics of the robotic knee exoskeleton. The knee joint angle ismarked with q.

after each of the following series. Following the last series, theparticipants were asked to rank the control approaches from themost to the least helpful.

C. Measures and Methods

The measured physiological variables were oxygen consump-tion (V O2), minute ventilation (VE ), heart rate (HR), and bloodoxygen saturation (SpO2). V O2 and VE were measured usinga portable metabolic cart (K4 b2 , Cosmed, Italy) while HR andSpO2 were measured with the WristOx 3100 Pulse Oximeter(Nonin Medical, USA).

We determined the energetics of squatting using metabolicequivalent of task (MET), which is a physiological measureexpressing the energy cost of physical activities [33]. METindexes the intensity of activities, where MET value of 2 re-quires twice the energy an average person consumes at rest.MET is calculated from oxygen consumption so that 1 MET =3,5 mlO2 /kg/min.

To assess an approximation of the overall muscle effort duringsquatting, we measured EMG of the Vastus Lateralis muscle onboth legs, rectified the signals, and finally integrated them overthe sliding time slice that corresponded to the duration of onesquatting motion (2 s). The EMG signals were recorded usingthe surface electrodes and the DataLog (Biometrics Ltd.) dataacquisition device. The integration of the EMG was carried outin MATLAB.

Joint angles and torques of the exoskeleton were deter-mined using the encoders and the motor-current measurements,respectively.

We also measured kinematical and dynamical aspects of mo-tion using the contactless motion capture system and forceplates. These are not discussed in this paper. Initial results ofthese aspects are presented in [3].

One-way ANOVA statistical analyses were performed to de-termine significant differences in responses of separate squattingseries. Multiple comparison procedure using the Tukey–Kramercriterion was applied to examine possible significant differences

in the power consumption of the device. The alpha level of sig-nificance was set a priori at 0.05.

D. Robotic Knee Exoskeleton

We designed a prototype robotic knee exoskeleton mecha-nism for both legs of the user. Each exoskeleton leg has oneactive rotational degree of freedom in the knee and a passiveball joint in the ankle. The ankle joint of the exoskeleton wasattached to the supporting frame, so the user did not have tocarry the mechanism. The ankle joint was relatively loose andallowed the device to rotate with the leg of the user.

Four velcro straps on each leg were used to attach the mecha-nism to the user. The coupling was made more rigid by additionalcrossbars behind the thigh, in front of the knee, and behind thecalf.

Each leg of the device was actuated with a 200 W, 36-V dcelectrical motor (Maxon RE50, 370955), followed by a trans-mission with a gear ratio 113. The combination allowed a theo-retical maximal continuous torque of 47.23 N·m, which matchesthe mean value of the maximal knee flexion/extension torque forthe bodyweight squats as reported in [34]. An external powersource was used. The device can be seen in Fig. 1, which showsa participant performing squatting movement with the attacheddevice.

III. EXOSKELETON CONTROL APPROACHES

We evaluated the effects of three control approaches on themetabolic cost of squatting. The control approaches were gravitycompensation approach, position-based approach, and a noveloscillator based approach.

A. Gravity Compensation Control Approach

In general, the gravity compensation approach compensatesthe weight of the mechanism itself and the load it is carrying, sothat the applied joint torques statically support the mechanismholding a load in its current configuration. In our case, therobotic knee skeleton was supporting itself and the participantbody weight. Only a part of the participant’s body weight wascompensated due to the limited torque output of the device.

When the robot kinematics is known in advance, the calcu-lation of the gravity compensation torque is usually analyti-cally and computationally undemanding. The kinematics of therobotic knee exoskeleton is defined as shown in Fig. 1. We as-sume that the thigh and the shin are roughly the same lengthd. The simplified gravity compensation torque τ is then definedwith

τ = mgd sinq

2(1)

where m is the mass above the knee and g is the length of thegravity vector. The term q/2 ensures that the device outputsthe maximal torque when the knee is fully flexed. Note thatonly the knee joints of the mechanism are actuated and that theparticipant has to maintain stability with the ankles.

Since the mechanism cannot output enough torque to com-pletely compensate the participant’s weight and only a part of

GAMS et al.: EFFECTS OF ROBOTIC KNEE EXOSKELETON ON HUMAN ENERGY EXPENDITURE 1639

Fig. 2. Control schemes for all three control approaches. τh denotes the human torque, τ is the device torque, and q is the actual knee joint angle. Left: Gravitycompensation approach control scheme. Middle: Oscillator-based approach. q is the estimated knee joint angle. Right: Position control approach control scheme.Note the control loop with the referential joint angle qd .

the weight is compensated, (1) derives into

τ = τnmi

mnsin

q

2. (2)

Here, τn = 47.32 N·m is the nominal continuous motor torque,mi is the mass of the ith participant, and mn = 85 kg is thenominal mass of the participants, i.e. the mass of the heaviestparticipant. In this case, the output torque is zero (τ = 0) whenq = 0, i.e., when the leg is fully extended. The output torque isequal to nominal torque (τ = τn ) when the leg is fully flexed(q = π). The gravity compensation control scheme is shown inFig. 2, left.

With this approach the device is always trying to extend theknee, even when the user is trying to perform a squat. Sincethe mechanism is back drivable, it reacts to external torque, i.e.,the mechanism complies if an external torque is greater that theoutput torque of the motor, for example, when the participantuses its weight to push it down.

B. Position-Based Control Approach

As the name suggests, the position control approach uses apositional feedback loop, as shown in Fig. 2, right. The controlleris given by

τ =

⎧⎪⎪⎨

⎪⎪⎩

τnmi

mn

qd − q

qn|qd − q| < qn

τnmi

mnsign(qd − q) |qd − q| ≥ qn .

(3)

This way the motor outputs the maximal torque when the kneejoint position error reaches the nominal value qn , set at qn = 10◦.Since the output is the torque and the feedback is position, thedevice acts in a compliant way. This control approach tries toensure that the joint angle position follows the desired position,in the experiments given with a sinusoidal curve. The humanmust, therefore, synchronize to the periodic movement of thedevice and not vice versa.

C. Oscillator-Based Control Approach

The oscillator-based approach relies on an adaptive frequencyoscillator to determine the frequency of the squatting movementand estimates the required helping torque. Instead of a pool ofadaptive oscillators in a feedback loop as in [17], and usinga logic algorithm to determine the basic frequency Ω as in[16], we used a single oscillator in the loop, combined with anadaptive Fourier series [18]. By determining the basic frequencyof squatting Ω, we predicted the joint position q, and reproduced

the desired torque waveform with a torque estimator. Fig. 2,middle, shows the control scheme.

The adaptive frequency oscillator-based joint position esti-mator is governed by

φ = Ω − K · e · sinφ (4)

Ω = −K · e · sin φ (5)

e = q − q (6)

where φ is the phase of the oscillator, q is the input signal (theknee joint angle), and K (positive constant) is the couplingstrength (see [18] for parameter settings). e is the input intothe oscillator, determined with a difference between the actual(q) and the estimated joint position (q). Since there is only oneoscillator, the frequency Ω and phase φ are clearly defined. Theestimated joint position q is determined by the adaptive Fourierseries

q =M

∑

c=0

(αc cos(cφ) + βc sin(cφ)) (7)

with M defining the size of the series. The amplitudes associatedwith the terms of Fourier series are determined by

αc = η cos(cφ) · e (8)

βc = η sin(cφ) · e (9)

where η is the learning constant and c = 0 . . . M . We used thefollowing parameter values: η = 2, K = 20, and M = 10. Theparameters were determined empirically.

In general, this structure acts as a state observer, in the sensethat it acts as a filter to smoothen the input signal. Unlike theconventional filters, this adaptive oscillator is able to predictand estimate the output in real time, without delays betweenthe input and the output. The only constraint is that the signalneeds to be periodic or pseudoperiodic [18]. Approaches fordetermining the onset of periodic behavior, such as walking, arebeing explored [35].

This method estimates and predicts the knee joint angle q,and outputs an estimated torque τ given by

τ = Kp (u − τ) (10)

where Kp is a positive constant and u is given by

u =

⎧⎪⎪⎨

⎪⎪⎩

τnmi

2mnsin

q

2q ≥ 0

τnmi

mnsin

q

2q < 0.

(11)

1640 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 6, JUNE 2013

TABLE IIAVERAGE OXYGEN CONSUMPTION (VO2 ; ml/min) OF THE PARTICIPANTS BEFORE (REST), DURING EACH MINUTE AND

FOLLOWING (REC) ALL FOUR SQUATTING SERIES

TABLE IIIAVERAGE MINUTE VENTILATION (VE ; l/min) OF THE PARTICIPANTS BEFORE (REST), DURING EACH MINUTE AND

FOLLOWING (REC) ALL FOUR SQUATTING SERIES

TABLE IVAVERAGE HEART RATE (HR; beats/min) OF THE PARTICIPANTS BEFORE (REST), DURING EACH MINUTE AND

FOLLOWING (REC) ALL FOUR SQUATTING SERIES

Fig. 3. (a): Relative changes in the average VO2 of the participants during different squatting series using the exoskeleton, compared to free series. + denotesstatistically significant difference between free and the other approaches. (b) Relative changes in the average VE of the participants during different squattingseries using the exoskeleton, compared to free series. * denotes significant difference between free and both gravity and oscillator approaches. (c) Relative changesin the average HR of the participants during different squatting series using the exoskeleton, compared to free series. * denotes significant difference between freeand both gravity and oscillator approaches.

When the knee joint velocity is q > 0, i.e., in the first half ofthe squat (when going down), the helping torque is lower thanwhen going up. This is also the main difference in comparisonto the gravity compensation approach.

Contrary to the position control approach, where the partici-pants have to synchronize to the predefined periodic movement,here synchronization takes place between the neuromechanicaloscillator, which is actually driving the joint and the artificialoscillator providing the assistance, similar as in [15]. The ar-tificial oscillator also continuously adapted to the frequencyand trajectory of the participants’s squatting motion, e.g., theoscillator-based control approach always tries to adapt to themovement of the participant and not vice versa.

IV. RESULTS

A. Physiological Responses

Tables II–IV list the measured V O2 , VE , and HR responses,respectively. Fig. 3 shows relative changes in the measured V O2 ,VE , and HR during the series with the knee exoskeleton com-pared to the Free series. The relative differences are expressedas percentage change from the Free series. Fig. 4 shows theaverage SpO2 values and standard deviations.

Statistical analysis has shown that for V O2 , shown in Fig. 3(a)and Table II, there is a significant difference (p < 0.01) be-tween not using the knee exoskeleton (free) and using the kneeexoskeleton in the third, fourth, and fifth minutes and in the

GAMS et al.: EFFECTS OF ROBOTIC KNEE EXOSKELETON ON HUMAN ENERGY EXPENDITURE 1641

Fig. 4. Average participant SpO2 during the squatting exercises. No changescan be observed with or without the knee exoskeleton, or between the controlapproaches of the exoskeleton. The error bars indicate standard deviation.

Fig. 5. Average participant MET during squatting without the robotic kneeexoskeleton (free) and with the device attached with all three control approaches.Error bars represent the standard errors of the means (s.e.m.). + denotes asignificant difference (p < 0.05) between free and other approaches.

recovery phase, regardless of the used control approach. Statis-tically, no difference can be observed between the three controlapproaches.

VE , Table III and Fig. 3(b), is significantly different (p <0.01) between free and both the gravity and the oscillator controlapproaches in fourth and fifth minutes and in the recovery phase,while no significant difference can be found between free andthe position approach.

Table IV shows absolute HR results while Fig. 3(c) showsrelative HR results compared to free, that were significantlydifferent (p < 0.01) between free and both gravity and oscillatorcontrol approaches from the second minute onward, while nosignificant difference was found between free and the positionapproach.

No differences were noted in SpO2 , as shown in Fig. 4. Theresults show that from a cardiorespiratory point of view, usingthe robotic knee exoskeleton for squatting significantly reducesresponses if the device is using either the gravity compensationor the oscillator-based control approach. On the other hand,no significant difference can be determined between these twomethods. Using the knee exoskeleton in position control modesignificantly reduces oxygen consumption (VO2) with respectto not using the device (free), while other parameters are notsignificantly affected.

Fig. 5 shows the average metabolic rate as expressed byMET for the squatting task without the robotic knee exoskeleton

Fig. 6. Integrated EMG signals as a measure of the overall muscle effortduring squatting. The width of the time slice used for integration correspondsto one squatting motion and equals to 2 s.

(free) and with all three control approaches. All three controlapproaches are significantly different (p < 0.05) in 3–5 min,while no significant difference can be observed between theapproaches.

Fig. 6 shows the overall muscle effort during squatting for oneof the participants. While the results for this particular partici-pant appear similar to physiological responses of all participants,statistically the results were inconclusive.

B. Device and Control Approach Responses

An important aspect of the interpretation of the physiologicalresponses is the behavior of the robotic knee exoskeleton underdifferent control approaches.

Fig. 7 shows the exhibited joint torque τ as a function ofthe knee joint angle q for all three control approaches for asingle participant. The plots are very similar for all participants.In the case of gravity compensation control approach (leftmostplot), the torques τ are linearly dependent on the joint angleq, which implies that the approach does not depend on themotion of the participant. The exact opposite is the positioncontrol method (rightmost plot), where a user, given a predefinedmovement trajectory, is unlikely to synchronize to the deviceperfectly. As shown in the position plot, the joint torques τ arespread over the whole range of joint movement, which indicatesthat the participant did not reach a steady state. On the otherhand, the oscillator plot (middle) shows that as a function ofjoint angle q, joint torques τ are in the form of a hysteresis.This indicates that the oscillator-based approach output torqueτ can successfully adapt to the human movement. For example,when the participant extends the knee, the torque produced bythe oscillator-based method is almost identical to the torqueproduced by the gravity-based method. On the other hand, whenthe participant flexes the knee (going down), the torque producedby the oscillator-based method is reduced, which eventuallyreduces the user’s effort, as there is no need to press hard on thedevice to flex. A slightly spread torque on the hysteresis in theoscillator plot is a consequence of the predictive nature of theoscillator (for details see [18]).

The success of the oscillator-based approach relies on anaccurate and rapid extraction of the movement frequency. Thetop plot in Fig. 8 shows the results of frequency extraction fromthe measured joint angle q. The extracted frequency quickly

1642 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 6, JUNE 2013

Fig. 7. Joint torque τ as a function of the robotic knee exoskeleton joint angle q for all three control approaches. The data of the same participant are given in allthree plots. Results of complete 5 min squatting series are depicted.

Fig. 8. Frequency extraction results of a randomly chosen participant duringthe “oscillator series.” Top: extracted frequency Ω. Bottom: difference betweenactual joint angles q and estimated joint angles q, denoted by e. The dashedblack lines mark 5o .

adapted to the input signal and stable motion was achieved, i.e.,the target frequency was 1 squat/2 s (0.5 Hz or π rad/s). In thebottom plot of Fig. 8, we can see the error e between the inputsignal q and the estimated signal q (6). After reaching the steadystate, at roughly 15 s, the error oscillates around 0o and remainsbelow 5o , depicted with dotted black lines.

Frequency extraction results in Fig. 8 also indicate successfulsynchronization between the neuromechanical oscillator (par-ticipants’s nervous system) and the adaptive frequency oscilla-tor (4)–(9). Note that a stable synchronization of the adaptiveoscillator was achieved by all participants, while not all of theparticipants were able to maximize the assisting torque underthe position-based approach. This is indicated by the scattereddata in position plot in Figs. 7 and 9, which show a square ofjoint position error (qd − q) for two participants, chosen to showthe possible extremes. Note that the scale on the y-axis of thebottom plot is 2 order lower. The plot for participant A showsthat the error between the desired joint position (qd)2 and theactual joint position q was considerable, which resulted in thehigh assisting torques. Participant B, on the other hand, did notexploit the device torque—the error was minimal and assistingtorque, linearly dependent on the error (3), was small.

V. DISCUSSION

Although squatting is a common task, periodic squatting canonly serve as a rough approximation of walking. Mainly because

Fig. 9. Squared position error between desired qd and actual q joint angles fortwo participants. Participant A exploited the device torque for squatting, hencethe higher error values. Participant B, on the other hand, did not exploit thedevice torque for assistance in squatting.

Fig. 10. Measured force under a single participant without the robotic knee ex-oskeleton (free) throughout the whole experiment (magenta color) and envelopesof the measured forces for all four squatting series for the same participant.

it is metabolically considerably more demanding than walking,yet kinematically quite simpler. However, many activities in ev-eryday life demand similar movement, for example, getting upfrom a chair, picking something up from the floor or ascend-ing/descending stairs. The latter also requires dissipation of theenergy (negative work) when going down, which imposes ad-ditional demands on the design of a robotic exoskeleton. Fromthe point of view of the listed tasks, periodic squatting can beconsidered as walking over a (very) rough terrain. The highmetabolic demand also allows efficient testing as fatigue sets inrelatively fast. Fig. 10 shows the summed measured force re-sponse of Kistler force plates under a participant when squattingwithout the device (in magenta color) and the envelopes of all

GAMS et al.: EFFECTS OF ROBOTIC KNEE EXOSKELETON ON HUMAN ENERGY EXPENDITURE 1643

four squatting series. Fatigue is clearly seen in the envelope ofthe free series measured forces. The presented envelope plotsresemble physiological responses plots, for example, in Fig. 5.The force data depicts the force under the participant and notthe force the knee exoskeleton exerts to the ground.

Our results show that using the robotic knee exoskeleton sig-nificantly reduces physiological responses during squatting ifan appropriate control approach is applied. Moreover, the aver-age MET is reduced under all control approaches. Based on theresults one could conclude that an exoskeleton mechanism willdefinitely reduce the metabolic cost. While studies on metaboliccost when using an exoskeleton mechanism for carrying loads,e.g., [21], [36], or a stationary mechanism for walking [1] haveshown an increase in the cost, some critical differences in thedesign of the studies can account for the different results. A partof the difference can be attributed to the fact that the participantsdid not have to carry the exoskeleton with them, as squattingdoes not demand lifting of legs and no single-leg support phasetakes place. Second, the movement was kinematically and dy-namically not as complex as bipedal walking and, therefore, thehelping torque that the device exerted was more easily targetedto produce the maximal effect on the user. The participants alsolightly touched a safety railing during the experiments. Thiscould potentially affect the balance and, therefore, influencethe physiological responses [32]. The effect was minimized byobserving the same safety protocol in all four squatting series,including the free series and the light touch, therefore cannotbe the reason for the observed differences in the physiologicalresponses.

The data obtained from using different control approacheshave shown that even in a very targeted task, such as squatting,the reduction in some of the physiological responses greatlydepends on the type of the knee exoskeleton control. Significantdifferences in all the measured physiological responses occurwhen the device is used with either the gravity compensationor the oscillator-based approach. On the other hand, the useof the knee exoskeleton with the position control approach, asdefined in this paper, does not statistically significantly reducethe heart rate and/or the minute ventilation. Position controlapproach requires the participant to adapt the motion in order totake full advantage of the device, see Fig. 9. A few adaptationtrials could potentially lead to adaptation of the participants,which could have an effect on the physiological responses. Thispoints out at an undesired property of the control approach,which requires training. This was not the case with the othertwo approaches. Oxygen consumption and MET results showthat no matter which control approach is used an exoskeletonsignificantly reduces both. Overall muscle effort, assessed withEMG integration in a sliding window, were inconclusive. Nocommon trends were observed in the responses for differentparticipants. This can be in large part attributed to the propertiesof surface EMG measurements prone to high levels of the noiseand depend on many factors, such as skin conductivity, sweating,and placement. Nevertheless, Fig. 6 shows responses whichcompletely coincide with other physiological responses.

Fig. 11 shows the system average power/weight ratio for allthree control approaches. The bars represent mean values and

Fig. 11. System average power/weight ratio per method in Watt/kilogram. Thebars show the mean values and the error bars indicate the standard deviation.+ denotes significant difference between control approaches.

TABLE VPARTICIPANT CONTROL APPROACH RANKING

the error bars indicate the standard deviation. While one-wayANOVA shows significant differences between the approaches,a posthoc Tukey–Kramer multiple comparison procedure showsthat position-based approach is significantly different from theoscillator based and gravity compensation approaches. This co-incides with the physiological responses.

Even though the number of participants was 7, the cross-overstudy design and one-way ANOVA analysis allow us to clearlyand objectively define the differences between the squatting se-ries. When asked to subjectively rank the control approachesfrom the most to the least helpful, four out of seven partic-ipants ranked the gravity compensation method as the mosthelpful, while two thought the oscillator-based method helpedthem the most, and one preferred the position-based method.Participant control approach ranking is presented in Table V.The results of the survey roughly coincide with the measuredphysiological responses, and the measured energy consumptionof the device, which both found that gravity compensation andoscillator-based approaches provoked similar responses, whileposition-based method was slightly different.

VI. CONCLUSION

Collectively, our data show that a robotic knee exoskeletondevice can significantly reduce the metabolic cost of a periodictask, in our case squatting. Moreover, we have also confirmedthe hypothesis that a choice of an exoskeleton control approachsignificantly affects the users physiological responses.

Our proposed oscillator control approach, based on an adap-tive oscillator combined with an adaptive Fourier series, al-lows the user to easily and in a natural way synchronize withthe robotic knee exoskeleton and, thus, significantly reduce themetabolic cost of squatting. Both oscillator based and gravitycompensation control approaches induced similar responses.

With a relatively simple task of squatting, where the assistingtorque could easily target the acting muscles, the metabolic costresults have shown a great promise for the robotic exoskeletons.

1644 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 6, JUNE 2013

REFERENCES

[1] R. Ronsse, T. Lenzi, N. Vitiello, B. Koopman, E. van Asseldonk,S. De Rossi, J. van den Kieboom, H. van der Kooij, M. Carrozza, andA. J. Ijspeert, “Oscillator-based assistance of cyclical movements: Model-based and model-free approaches,” Med. Biol. Eng. Comput., vol. 49,pp. 1173–1185, 2011.

[2] A. Dollar and H. Herr, “Lower extremity exoskeletons and active or-thoses: Challenges and state-of-the-art,” IEEE Trans. Robot., vol. 24,no. 1, pp. 144–158, Feb. 2008.

[3] J. Babic, T. Petric, T. Debevec, and A. Gams, “Kinematic adaptationsduring repetitive squatting motions using robotic knee exoskeleton,” inProc. 21th Int. Workshop Robot. Alpe-Adria-Danube Reg., 2012, pp. 313–317.

[4] S. De Rossi, N. Vitiello, T. Lenzi, R. Ronsse, B. Koopman, A. Persichetti,F. Vecchi, A. Ijspeert, H. Van der Kooij, and M. Carrozza, “Sensing pres-sure distribution on a lower-limb exoskeleton physical human-machineinterface,” Sensors, vol. 11, no. 1, pp. 207–227, 2011.

[5] J. M. Carmena, M. A. Lebedev, R. E. Crist, J. E. O’Doherty, D. M.Santucci, D. F. Dimitrov, P. G. Patil, C. S. Henriquez, and M. A. L.Nicolelis, “Learning to control a brainmachine interface for reaching andgrasping by primates,” PLoS Biol., vol. 1, no. 2, p. e42, 2003

[6] J. Rosen, M. Brand, M. Fuchs, and M. Arcan, “A myosignal-based pow-ered exoskeleton system,” IEEE Trans. Syst. Man Cybern. Part A: Syst.Humans, vol. 31, no. 3, pp. 210–222, May 2001.

[7] E. Cavallaro, J. Rosen, J. Perry, and S. Burns, “Real-time myoprocessorsfor a neural controlled powered exoskeleton arm,” IEEE Trans. Biomed.Eng., vol. 53, no. 11, pp. 2387–2396, Nov. 2006.

[8] C. Kinnaird and D. Ferris, “Medial gastrocnemius myoelectric controlof a robotic ankle exoskeleton,” IEEE Trans. Neural Syst. Rehabil. Eng.,vol. 17, no. 1, pp. 31–37, Feb. 2009.

[9] C. Fleischer, C. Reinicke, and G. Hommel, “Predicting the intended mo-tion with emg signals for an exoskeleton orthosis controller,” in Proc.IEEE/RSJ Int. Conf. Intell. Robot. Syst., Aug. 2005, pp. 2029–2034.

[10] S. Lee and Y. Sankai, “Power assist control for walking aid with hal-3based on emg and impedance adjustment around knee joint,” in Proc.IEEE/RSJ Int. Conf. Intell. Robot. Syst., 2002, vol. 2, pp. 1499–1504.

[11] K. Kiguchi, K. Iwami, M. Yasuda, K. Watanabe, and T. Fukuda, “Anexoskeletal robot for human shoulder joint motion assist,” IEEE/ASMETrans. Mechatronics, vol. 8, no. 1, pp. 125–135, Mar. 2003.

[12] T. Petric, A. Gams, M. Tomsic, and L. Zlajpah, “Control of rhythmicrobotic movements through synchronization with human muscle activity,”in Proc. IEEE Int. Conf. Robot. Autom., May 2011, pp. 2172–2177.

[13] H. Kazerooni, J.-L. Racine, L. Huang, and R. Steger, “On the controlof the berkeley lower extremity exoskeleton (bleex),” in Proc. IEEE Int.Conf. Robot. Autom., Apr. 2005, pp. 4353–4360.

[14] S. Banala, S. Agrawal, A. Fattah, V. Krishnamoorthy, W.-L. Hsu, J. Scholz,and K. Rudolph, “Gravity-balancing leg orthosis and its performanceevaluation,” IEEE Trans. Robot., vol. 22, no. 6, pp. 1228–1239, Dec.2006.

[15] R. Ronsse, N. Vitiello, T. Lenzi, J. van den Kieboom, M. Carrozza, andA. Ijspeert, “Human-robot synchrony: Flexible assistance using adaptiveoscillators,” IEEE Trans. Biomed. Eng., vol. 58, no. 4, pp. 1001–1012,Apr. 2011.

[16] A. Gams, A. J. Ijspeert, S. Schaal, and J. Lenarcic, “On-line learning andmodulation of periodic movements with nonlinear dynamical systems,”Autonom. Robot., vol. 27, no. 1, pp. 3–23, 2009.

[17] L. Righetti, J. Buchli, and A. J. Ijspeert, “Dynamic hebbian learning inadaptive frequency oscillators,” Phys. D, vol. 216, no. 2, pp. 269–281,2006.

[18] T. Petric, A. Gams, A. J. Ijspeert, and L. Zlajpah, “On-line frequencyadaptation and movement imitation for rhythmic robotic tasks,” Int. J.Robot. Res., vol. 30, no. 14, pp. 1775–1788, 2011.

[19] R. Ronsse, N. Vitiello, T. Lenzi, J. van den Kieboom, M. Carrozza, andA. Ijspeert, “Adaptive oscillators with human-in-the-loop: Proof of con-cept for assistance and rehabilitation,” in Proc. IEEE 3rd RAS EMBS Int.Conf. Biomed. Robot. Biomechatron., Sep. 2010, pp. 668–674.

[20] Y. Muramatsu, H. Kobayashi, Y. Sato, H. Jiaou, T. Hashimoto, andH. Kobayashi, “Quantitative performance analysis of muscle suit—estimation by oxyhemoglobin and deoxyhemoglobin,” in Proc. IEEE Int.Conf. Robot. Biomimet., Dec. 2011, pp. 293–298.

[21] C. J. Walsh, K. Endo, and H. Herr, “A quasi-passive leg exoskeletonfor load-carrying augmentation,” Int. J. Humanoid Robot., vol. 4, no. 3,pp. 487–506, 2007.

[22] M. Bernhardt, M. Frey, G. Colombo, and R. Riener, “Hybrid force-positioncontrol yields cooperative behaviour of the rehabilitation robot lokomat,”in Proc. 9th Int. Conf. Rehabil. Robot., Jun./Jul. 2005, pp. 536–539.

[23] J. Veneman, R. Kruidhof, E. Hekman, R. Ekkelenkamp, E. Van Asseldonk,and H. van der Kooij, “Design and evaluation of the lopes exoskeletonrobot for interactive gait rehabilitation,” IEEE Trans. Neural Syst. Rehabil.Eng., vol. 15, no. 3, pp. 379–386, Sep. 2007.

[24] C. Krewer, F. Muller, B. Husemann, S. Heller, J. Quintern, and E. Koenig,“The influence of different lokomat walking conditions on the energyexpenditure of hemiparetic patients and healthy subjects,” Gait Posture,vol. 26, no. 3, pp. 372–377, 2007.

[25] G. Sawicki and D. Ferris, “Powered ankle exoskeletons reveal themetabolic cost of plantar flexor mechanical work during walking withlonger steps at constant step frequency,” J. Exper. Biol., vol. 212, no. 1,pp. 21–31, 2009.

[26] T. Lenzi, D. Zanotto, P. Stegall, M. Carrozza, and S. Agrawal, “Reducingmuscle effort in walking through powered exoskeletons,” in Proc. IEEEAnnu. Int. Conf. Eng. Med. Biol. Soc., Sep. 2012, pp. 3926–3929.

[27] D. Winter, Biomechanics and Motor Control of Human Movement, 4th ed.New York, USA: Wiley, 2009.

[28] P. Costigan, K. Deluzio, and U. Wyss, “Knee and hip kinetics duringnormal stair climbing,” Gait Posture, vol. 16, no. 1, pp. 31–37, 2002.

[29] P. Wretenberg and U. Arborelius, “Power and work produced in differ-ent leg muscle groups when rising from a chair,” Eur. J. Appl. Physiol.Occupat. Physiol., vol. 68, pp. 413–417, 1994.

[30] J. Pratt, B. Krupp, C. Morse, and S. Collins, “The roboknee: An exoskele-ton for enhancing strength and endurance during walking,” in Proc. IEEEInt. Conf. Robot. Autom., Apr./May 2004, vol. 3, pp. 2430–2435.

[31] R. Horst, “A bio-robotic leg orthosis for rehabilitation and mobility en-hancement,” in Proc. IEEE Annu. Int. Conf. Eng. Med. Biol. Soc., Sep.2009, pp. 5030–5033.

[32] J. J. Jeka and J. R. Lackner, “Fingertip contact influences human posturalcontrol,” Exper. Brain Res., vol. 79, no. 2, pp. 495–502, 1994.

[33] B. Ainsworth, W. Haskell, S. Herrmann, N. Meckes, D. Bassett Jr.,C. Tudor-Locke, J. Greer, J. Vezina, M. Whitt-Glover, and A. Leon, “2011compendium of physical activities: A second update of codes and met val-ues,” Med. Sci. Sports Exercise, vol. 43, no. 8, pp. 1575–1581, 2011.

[34] S. Almosnino, D. Kingston, and R. B. Graham, “Three-dimensional kneejoint moments during performance of the bodyweight squat: Effects ofstance width and foot rotation,” J. Appl. Biomech., 2012, To be published

[35] P. Rebersek, D. Novak, J. Podobnik, and M. Munih, “Intention detectionduring gait initiation using supervised learning,” in Proc. IEEE-RAS 11thInt. Conf. Humanoid Robot., Oct. 2011, pp. 34–39.

[36] K. N. Gregorczyk, L. Hasselquist, J. M. Schiffman, C. K. Bensel,J. P. Obusek, and D. J. Gutekunst, “Effects of a lower-body exoskele-ton device on metabolic cost and gait biomechanics during load carriage,”Ergonomics, vol. 53, no. 10, pp. 1263–1275, 2010.

Authors’ photographs and biographies not available at the time of publication.