Split (1)

train · 118 rows

text stringlengths 710 6.67k
Delayed Output Feedback Control for Gait Assistance and Resistance Using a Robotic Exoskeleton Bokman Lim, Junwon Jang, Jusuk Lee, Ryungjune Choi, Younbaek Lee, and Youngbo Shim Abstract—In this study, we propose an interaction control framework for gait assistance and resistance using a robotic exoskeleton. We define a smoothed state variable that represents joint angle movements while walking. Furthermore, a self-feedback controller is designed with the delayed output state. By applying an appropriate time-delay and positive or negative feedback gain to the state variable, we can generate assistive or resistive torque stably without any gait phase or environment recognition. The time-delayed selffeedback controller reflects the movement of the wearer's joints at every moment of control, thereby stably coping with sudden task transitions (e.g., walk-stop-walk, forward-backward walking) as well as walking speed or environment changes. Case studies involved gait assistance with a knee exoskeleton and gait assistance and resistance with a hip exoskeleton. We performed various preliminary tests including metabolic energy measurements and a comparison of the positive or negative power of the generated torque profiles. The results show the flexibility and effectiveness of the proposed interaction control method for gait assistive or resistive training. Index Terms—Prosthetics and exoskeletons, rehabilitation robotics, motion control. I. INTRODUCTION AIT training (or exercises) using a robotic exoskeleton can be an appropriate solution for those who need to improve their walking performance due to aging or disease. Hip and ankle assistance, for example, is expected to lessen the problem of excessive use of the hip muscles or a bent posture to compensate for weakened distal muscle strength and balance [1], [2]. It is expected that knee assistance can effectively improve a gait rehabilitation and training program in patients who suffer from arthritis [3] or have joint replacements. To maximize the training effect using an exoskeleton device, interaction force should be applied naturally in accordance with the wearer's original walking pattern. Because the exoskeleton device's weight and motion constraints can distort the user's original gait pattern, reducing the device's weight and improving its wearability should take precedence. The device's usability, maintenance and manufacturing costs, and ease of wear are also key factors that cannot be overlooked. Manuscript received February 24, 2019; accepted June 22, 2019. Date of publication July 10, 2019; date of current version July 24, 2019. This letter was recommended for publication by Associate Editor T. Lenzi and Editor P. Valdastri upon evaluation of the reviewers' comments. (Corresponding author: Bokman Lim.) The authors are with the Samsung Advanced Institute of Technology, Suwon 16678, South Korea (e-mail: bokman.lim@samsung.com; jw526.jang@samsung.com; jusuk7.lee@samsung.com; bj81.choi@samsung.com; younbaek.lee@samsung.com; ddalbo.shim@samsung.com). Digital Object Identifier 10.1109/LRA.2019.2927937 In addition, it is necessary to design an interactive controller capable of responding robustly and stably to changes in the wearer's movement. This is because of the irregular walking patterns of those who require rehabilitation, such as neuropathic patients, and those who have suffered strokes, and because training programs are conducted in a wide variety of conditions [4] (e.g., step/stair walking, walking over ground that entails navigating obstacles, walking at a self-selected fast speed, walking using a rail to walk forwards and backwards, and walking while completing a cognitive task). Patients who underwent joint replacement surgery must also achieve the rehabilitation goal of increasing the wearer's leg joint range of motion [5], such that fine assist strength control is possible. Existing exoskeleton control methods based on walking phase/environment recognition with neural oscillators [6]–[9] or discrete gait events [10]–[12] cannot easily overcome the above problems because accuracy with irregular gait patterns in walking phase/environment recognition is more difficult to guarantee. To overcome this drawback (depending on the periodicity of the motion), Nagarajan et al. proposed an admittance control strategy based on modifying the dynamic response of a coupled human-exoskeleton system control [13]. Rehabilitation training programs are classified into two methods: applying assistive force and applying resistance. Recently, studies on the usefulness of walking resistance training have been reported [14], [15]. We expect that a variety of stimuli, including resistive interaction forces, will help balance training [16]. However, there are few examples of resistance torque control using exoskeletons [17] and it is difficult to find an interaction control method that can simultaneously cover assistive and resistive torque generation. Thus, we present a novel interaction control framework for gait assistance and resistance to overcome many of the limitations described above. The interaction controller is based on delayed output feedback control known for stabil
izing oscillatory systems under certain conditions [18]–[20]. By adding a timedelay buffer to the self-feedback control loop, we can generate assistive or resistive torque stably in the interaction between the user and exoskeleton. The proposed interaction controller can operate at various gait speeds and under environmental changes (e.g., stairs, up/down ramps) with only angular positions and without the need for gait phase or environment recognition. The proposed framework can appropriately handle non-uniform ground conditions, such as ramp–level–stair, sudden stopping, and forward–backward walking. 2377-3766 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. Fig. 1. Hardware prototypes for gait training with the Gait Enhancing and Motivating System (GEMS), and an interaction control framework for gait assistance and resistance. GEMS-K is our knee exoskeleton and GEMS-H is our hip exoskeleton. Previously, we proposed a hip assistance controller based on a time-delayed feedback control method [21]. In this study, we show that the proposed method can be applied not only to other types of exoskeletons, such as knee exoskeleton control, but also to resistive torque generation. The characteristics of the control parameters (viz., the time delay and feedback gain) are also analyzed and discussed. To my best knowledge, this is the first example of applying time delay control to walking resistance and knee assistance. This letter is organized as follows: Section II describes the interaction control framework based on delayed output feedback control for knee and hip exoskeletons. Section III provides the experimental results under various walking conditions. Finally, we conclude with a summary of the study and suggestions on how to extend our framework into other types of exoskeletons, such as single joint driven assistance (e.g., an ankle exoskeleton device worn only on one foot). II. FRAMEWORK AND ALGORITHM A. Interaction Control for Gait Assistance and Resistance Our interaction control framework for gait assistive or resistive training is shown in Fig. 1. The interaction control method is based on delayed output feedback control (DOFC). The DOFC-based controller design can be classified into two steps: Step 1: Define an output state representing the current leg's motion with joint angular positions. Step 2: Determine the smoothing rate, delay time, and feedback gain of the output state. Once the value or range of the control parameters is determined, it is configured as a self-excited feedback control loop. As shown in Fig. 1(c), the input of the interaction controller comprises the joint angular positions and the output is the interactive torque. The original state value $y_{raw}(t)$ is calculated as a function of the joint angles, $y_{raw}(t) = f(q(t))$ . Noisy sensing data $y_{raw}(t)$ is filtered through the state smoother. Delay due to state smoothing is not a problem because an additional time delay is used in the next procedure. The smoothed state y(t) is delayed for some time by passing through the state delayer, $y(t) \rightarrow y(t-\Delta t)$ . This process is easily implemented by using a constant time delay buffer. Here, reducing the time delay value leads to an early assist (or response), while increasing the time delay value results in a late assist (or response). Finally, interaction control torque is generated by multiplying the delayed state $y(t-\Delta t)$ by the feedback gain $\kappa$ . The magnitude of the gain is proportional to the magnitude of the generated interaction torque, i.e., as the gain increases, the generated torque becomes more assistive or resistive. At this time, amplifying the state value using positive gain generates assistive torque, and, conversely, negative gain generates resistive torque. B. Assistive or Resistive Torque Generation From Knee or Hip Joint Angles The proposed interaction control framework can be expressed in a more detailed and concrete form, as shown in Fig. 2. We first define an output state $y_{raw}(t)$ as that representing the projected leg motion: $$y_{raw}(t) = \sin q_r(t) - \sin q_l(t) \tag{1}$$ where $q_r$ and $q_l$ are the right and left joint angles, respectively. We use the sinusoidal projected joint angle difference for knee and hip exoskeleton control. The original noisy sensor data is smoothed by passing through a simple first-order low-pass filter: $$y^{i} = (1 - \alpha)y^{i-1} + \alpha y_{raw}^{i}, \ (0 < \alpha < 1)$$ (2) where i denotes
the current sample time and i-1 is the previous sample time. The current smoothed state $y^i$ is expressed as a weighted sum of the previous sample time state $y^{i-1}$ and the original state value of the current sample time $y^i_{raw}$ , and the smoothing rate can be adjusted by changing the smoothing factor $\alpha$ . (a) Assistance for both knees with DOFC framework (b) Left hip assistance/resistance with DOFC framework Fig. 2. Examples of DOFC framework-based assistance or resistance for knee or hip exoskeleton. $q_r$ and $q_l$ are the right and left joint angles, respectively, and $\kappa$ is the feedback gain. The assistive or resistive torque $\tau$ is then generated through a combination of appropriate time delays $\Delta t$ and positive or negative gains $\kappa$ : $$\tau(t) = \kappa y(t - \Delta t) = \begin{cases} \text{if } \kappa > 0, \text{ assist mode} \ \text{if } \kappa < 0, \text{ resist mode} \end{cases}$$ (3) 1) Gait Assistance With a Knee Exoskeleton: Fig. 2(a) shows an example of DOFC-based walking assistance with a knee exoskeleton. Assistive torque for the left knee $\tau_{l,assist}$ is generated with a positive gain, and that for the right knee can be obtained by reversing the left assistance torque $(\tau_{r,assist} = -\tau_{l,assist})$ . This assistance strategy has advantages in terms of motion synergy for assisting both legs. 2) Gait Assistance and Resistance With a Hip Exoskeleton: Fig. 2(b) shows an example of DOFC-based walking assistance and resistance with a hip exoskeleton. Assistive torque for the left hip $\tau_{assist}$ is generated with a positive gain, and, conversely, resistive torque $\tau_{resist}$ is generated with a negative gain. The extent of the assistance or resistance can be gradually increased or decreased by gradually increasing or decreasing the positive or negative gain. Assistive or resistive torque for the right hip can also be obtained easily with a similar strategy (see Fig. 2(a) for an example with assistance to both knees). III. CASE STUDIES: KNEE AND HIP EXOSKELETONS This section considers two case studies: (i) gait assistance with a knee exoskeleton, and (ii) gait assistance and resistance with a hip exoskeleton. Fig. 1(a) and 1(b) show our latest knee and hip exoskeleton prototypes with the Gait Enhancing and Motivating System for a Knee and Hip (GEMS-K and GEMS-H). The weight of GEMS-K (knee-only actuated) and GEMS-H (hip-only actuated) is about 3.6 kg and 2.1 kg, respectively. A. Gait Assistance With a Knee Exoskeleton The basic assistance strategy (Fig. 2(a)) can be extended for both right/left and extension/flexion knee torque generations by modifying the original torque equation in (3). Right knee flexion, left knee extension: $$\tau_{rk,des}(t) = -\tau(t)$$ $$\tau_{lk,des}(t) = \tau(t) \cdot \delta_K$$ (4) Left knee flexion, right knee extension: $$\tau_{lk,des}(t) = \tau(t)$$ $$\tau_{rk,des}(t) = -\tau(t) \cdot \delta_K$$ (5) where $\delta_K$ denotes the knee extension/flexion torque ratio (if $\delta_K = 1$ , the knee flexion and extension torque strength is the same). The applied smoothing factor $\alpha_K$ for the knee exoskeleton in (2) is 0.04. This smoothing factor $\alpha_K$ is manually set for smooth interaction and torque generation. Fig. 3 shows the joint angle and assistance torque values during walking task transitions from walking to stopping and stopping to walking (time-delay $\Delta t = 0.25$ s, feedback gain $\kappa = 10$ , extension/flexion ratio $\delta_K = 1$ ). The assistance controller and the output torque are directly affected by the knees range of motion, as shown in Fig. 3. 1) Effects on Metabolic Cost With Knee Assistance: Unlike ankle or hip assistance in walking, studies on the metabolic energy reduction in knee assists (or multiple-joint assists including the knee
) are difficult to find. Rather, it is known that metabolic energy greatly increases due to increases in weight and motion constraints due to a rigid frame and actuator structure covering the knee [22]–[24]. The increase in metabolic energy during the device's operation implies that it is a burden to the user even to wear the device. Therefore, the effect and usefulness of the gait rehabilitation/training is significantly reduced. First, we need to at least prevent the increase in metabolic energy due to knee exoskeleton assistance. Five male subjects participated in the experiment (age: $38 \pm 2.2$ ; weight: $68.4 \pm 6.3$ kg; height: $173 \pm 8.2$ cm; mean $\pm$ standard deviation). First the subjects stood for 5 minutes (as they did again at the end of the experiment) to obtain the average baseline from which to subtract the walking data Fig. 3. Knee joint angle, velocity, torque, and power trajectories for the selected task transitions from walking to stopping and stopping to walking. The knee joint sensing data are obtained from the knee exoskeleton. The walk-stop-walk (time-delay $\Delta t=0.25$ s, feedback gain $\kappa=10$ , extension/flexion ratio $\delta_K=1$ , smoothing factor $\alpha_K=0.04$ ) is performed with exoskeleton assistance. In the torque plot, the gray dashed line represents the estimated torques (from current sensing), while the blue and red solid lines denote the generated desired assistance torque. to obtain the net metabolic rate. Then they walked without the exoskeleton for six minutes (as they did again near the end of the experiment) to obtain the average metabolic rate under normal walking conditions. Then they wore the exoskeleton and walked on the treadmill with assistance. We took the median of the last three minutes of each condition to represent the metabolic rate expended under those conditions. The K5 "breath by breath" portable metabolic system (COSMED, Rome, Italy) was used to measure the metabolic energy expenditure. The treadmill speed was set to 4 km/h for all subjects. Knee control parameters were selected in advance as the most preferred values for the wearer. The applied assistance torque was mainly flexion rather than extension (extension/flexion ratio $\delta_K$ : 0.1 or 0.2; gain $\kappa$ : $10{\sim}13$ ; RMS torque: $3.4\pm0.6$ Nm; maximum flexion torque: $7.9 \pm 1.3$ ; maximum extension torque: $1.5 \pm 0.5$ Nm). The applied time-delay $\Delta t$ values were 0.2 or 0.25 s. Fig. 4 shows the knee exoskeleton sensing data and generated assistive torque and power applied to each subject. As shown in Table I, the metabolic energy expenditure in knee assistance was $3.55 \pm 0.52$ W/kg, which increased by an average of 3.4% compared to normal walking state (without the exoskeleton) of $3.45 \pm 0.56$ W/kg. However, the difference was not statistically significant ( $^\dagger p$ – value = 0.4 > 0.05 for the paired t-test no exo vs. exo). In other words, there was no significant change in metabolic energy expenditure due to wearing the knee exoskeleton and its assistance during walking. This means that the increase in metabolic energy due to weight and the movement restriction due to wearing the knee exoskeleton was offset by knee assistance during walking. Through this preliminary test, we showed that an increase in metabolic energy can be prevented through knee-only assistance for the five healthy subjects, and furthermore that it might Fig. 4. Knee exoskeleton joint sensing data for five subjects. The joint torque represents the generated input assistance torque. The joint power is calculated by multiplying the joint torque by the joint velocity. TABLE I REDUCED METABOLIC COST WITH THE KNEE EXOSKELETON \| Subject \| No Exo \| Exo Assist \| \| \|------------\|-------------\|-----------------\|-----------------------\| \| No. \| rNMR (W/kg) \| NMR (W/kg) \| rNMR (%) \| \| 1 \| 4.29 \| 4.43 \| -3.4 \| \| 2 \| 2.77 \| 3.13 \| -12.8 \| \| 3 \| 3.38 \| 3.62 \| -7.3 \| \| 4 \| 2.95 \|
2.94 \| 0.3 \| \| 5 \| 3.87 \| 3.63 \| 6.3 \| \| Mean (±SD) \| 3.45±0.56 \| $3.55 \pm 0.52$ \| -3.4±6.5 † \| NMR: net metabolic rate; rNMR: reduced net metabolic rate from free walking condition (No exo). $^\dagger p$ - value = 0.4 > 0.05 for paired t-test (No exo vs Exo). A negative value means that metabolic energy has increased compared to normal walking (with no exoskeleton). be possible to reduce metabolic energy by tuning the control parameters and optimizing the extension/flexion assistance ratio, magnitude, and timing selection. (Notice that three of the five subjects had increased metabolic energy expenditure, while two had decreased metabolic energy.) B. Gait Assistance and Resistance With a Hip Exoskeleton The basic assistance strategy (Fig. 2(b)) can be extended for both right/left hip torque generation $\tau_{r,des}$ , $\tau_{l,des}$ by modifying the original torque equation in (3). Right hip flexion, left hip extension: $$\tau_{rh,des}(t) = -\tau(t)$$ $$\tau_{lh,des}(t) = \tau(t) \cdot \delta_H$$ (6) Left hip flexion, right hip extension: $$\tau_{lh,des}(t) = \tau(t)$$ $$\tau_{rh,des}(t) = -\tau(t) \cdot \delta_H$$ (7) where $\delta_H$ denotes the hip extension/flexion torque ratio (if $\delta_H$ = 1, the hip extension and flexion torque strength is the same). In this study, we set the smoothing factor for the hip exoskeleton $\alpha_H$ to 0.05 in (2) and the hip extension/flexion ratio $\delta_H$ to 1. This smoothing factor $\alpha_H$ , like the knee exoskeleton case, was manually determined to produce smooth interaction torque. Fig. 5. Generated mean positive/negative power and the RMS torque changes due to time-delay changes ( $\Delta t = 0.05, 0.15, 0.25, 0.35 \text{ s}, \kappa = 8$ ). Positive: mean positive power, Negative: mean negative power. Fig. 6 shows the joint angle and assistance torque values during walking task transitions from forward-to-backward walking and backward-to-forward walking. The assistance controller and the output torque are directly affected by the hip motions, as seen in Fig. 6. For the forward-backward-forward walk transition task, the resulting mean positive and negative power values were $3.85~\mathrm{W}$ and $-0.13~\mathrm{W}$ , respectively. Lower negative power value compared to relatively large positive power supports the idea that the resistive torque generation was minimal. With six male subjects (age: $41\pm3.2$ ; weight: $71\pm5.0$ kg; height: $174\pm8.2$ cm), we previously showed that hip assistance can reduce metabolic energy from walking by an average of 20% compared to normal walking state (without the exoskeleton) [21]. We also showed with same six male subjects how the proposed assistance algorithm can be generalized in assisting in various conditions (speed and environment changes) by showing the differently adapted torque and power profiles with fixed control parameters ( $\Delta t = 0.25, \kappa = 8$ ) [21]. The generated torque and power under fixed control parameters showed a consistent trend of change in walking speed/environment (except for small variance due to individual hip pattern differences). For this reason, experiments were conducted on a single subject and on a variety of situations. 1) Relationship Between Time-Delay and Generated Power/ Torque: We can adjust the assistance response (or timing) by adjusting the time-delay $\Delta t$ . Figure 5 shows the generated MP (mean power) and RMS (root mean square) torque for four selected time-delay values $\Delta t = 0.05, 0.15, 0.25, 0.35$ s. As shown in Fig. 2(b), the generated torque has a sinusoidal form, so the RMS torque value difference denotes the generated torque difference. One male subject (age: 40; weight: 67 kg; height: 160 cm) wore the hip exoskeleton and walked with assistance. We used a fixed gain of $\kappa=8$ .
The treadmill speed increased from 1 km/h to 5 km/h in 1 km/h increments. The torque and power generated in the hip joint were calculated with a sensor attached to the exoskeleton. As shown in Fig. 5(b), the time-delay $\Delta t$ affects the generated torque magnitude, even though it is a control variable related to assistance timing. The generated torque amplitude increases as the time-delay increases. The positive and negative powers delivered to the subject is shown in Fig. 5(a). The magnitude of the positive power generated is likely to be highest at around a time-delay of Fig. 6. Hip joint angle, velocity, torque, and power trajectories for the selected task transitions from forward-to-backward walking and backward-to-forward walking. The hip joint sensing data are obtained from the hip exoskeleton. A forward–backward–forward walk ( $\Delta t=0.25~\rm s, \kappa=8, \delta_H=1, \alpha_H=0.05)$ is performed with exoskeleton assistance. In the torque plot, the gray dashed line represents the estimated torque (from current sensing), while the blue and red solid lines denote the generated desired assistance torque. $\Delta t=0.25$ s. If there is too much time delay, such as when $\Delta t=0.35$ s, this indicates that there is a problem with generating assistive (positive) power for high-speed walking. An increase in negative power generation means that the amount of work done by the wearer has increased due to a mismatch between the assist and the users movement. If the time-delay value is too small ( $\Delta t=0.05$ s), negative power generation does not increase even at high-speed gait, but there is also less positive power generation and assistive torque generation. It is therefore necessary to set an appropriate time-delay value. 2) Adjusting Transfer Power by Gain Change: We can adjust the assistance/resistance strength (strong or weak) by adjusting the feedback gain $\kappa$ . Same one male subject walked on the treadmill after wearing the device. The treadmill speed was fixed at 4 km/h. We set the time-delay $\Delta t$ to 0.25 s. From $\kappa=-5$ to $\kappa=10$ , we increased the gain value by 0.2 for every two steps (1 walking cycle). Fig. 7(a) shows the change in the torque and power profile when the gain gradually increased from -5 to 10. Fig. 7. Generated torque and power changes due to gain changes. The feedback gain κ was increased from −5 to 10 (Δt = 0.25 s, 4 km/h walking speed). Positive: mean positive power, Negative: mean negative power. Fig. 8. Raw metabolic rate data. The red line denotes the raw data. The black line denotes the filtered metabolic data shown for visual purposes. The horizontal black line denotes the median with respect to the last 3 minutes under each condition. The resistance torque generated at the negative gain κ = −5 was inverted as the gain increased little by little and smoothly changed to the assistive torque (κ > 0). The generated power also shows a gradual change in response to the gain change. Fig. 7(b) shows the gain and generated RMS torque and power relationships under given walking conditions (Δt = 0.25 s, 4 km/h walking speed). A linear proportional relationship between the generated torque/power and the gain κ is observed. In the plot on the right side of Fig. 7(b), when the gain is positive (κ > 0), mean negative power (MNP) generation is inhibited compared to mean positive power (MPP). On the other hand, when the gain is negative (κ < 0), MPP generation is effectively suppressed compared to MNP. 3) Effects on Metabolic Cost by Adjusting Gain: Fine control of power delivery can be extremely useful for gait rehabilitation or training. This is because customized rehabilitation is possible depending on the stage of the rehabilitation and the patients condition. Fig. 8 and Table. II show the resulting metabolic measurement for five selected gains κ = 8, 9.5, 11, 12.5, 14. TABLE II REDUCED METABOLIC COST WITH THE HIP EXOSKELETON \| Condition \| Exo-generated \| \| Human-reduced \| \| \|-------------------\|-------------------\|---------\|---------------\|----------\| \| (gain) \| $\tau_{RMS}$ (Nm) \| MPP (W) \| NMR (W/kg) \| rNMR (%) \| \| $\kappa_1 = 8.0$ \| 3.86 \| 7.48 \| 3.49 \| -3.1 \| \|
$\kappa_2 = 9.5$ \| 4.30 \| 8.29 \| 3.24 \| -10.1 \| \| $\kappa_3 = 11.0$ \| 4.78 \| 9.30 \| 3.07 \| -14.8 \| \| $\kappa_4 = 12.5$ \| 5.22 \| 10.36 \| 2.98 \| -17.1 \| \| $\kappa_5 = 14.0$ \| 5.70 \| 11.34 \| 2.85 \| -20.9 \| τRMS : RMS torque; MPP: mean positive power; NMR: net metabolic rate; rNMR: reduced net metabolic rate from free walking condition (No exoskeleton). Fig. 9. Generated torque and power changes due to cadence changes. The treadmill speed was increased from 2 to 6 km/h (Δt = 0.25 s, κ = 8). Positive: mean positive power, Negative: mean negative power. Same one male subject walked on the treadmill at 4 km/h speed with assistance. We used the same protocol as that used to measure metabolic cost with knee assistance except for the exoskeletons settings. The subject wore the exoskeleton and walked on the treadmill under five different assistance conditions (κ1 < κ2 < ... < κ5). The results from this preliminary test are shown in Fig. 8 and Table. II. The results demonstrate the possibility of stepwise intensity adjustment by controlling the gain value, and they are consistent with related studies which suggest that the assist power magnitude is proportional to the amount of reduced metabolic energy [25]. 4) Assistance Adaptation to Variable Cadence Using Fixed Control Parameters: Our proposed control algorithm is stable enough to cope with cadence (or speed) changes. The control parameters were fixed (Δt = 0.25 s, κ = 8) under variable cadence. For the cadence change experiment, the same one subject walked on the treadmill with assistance at different walking speeds. The treadmill speed increased from 2 km/h to 6 km/h in 0.1 km/h increments for every step. Fig. 9(a) shows the change in the torque and power profile when the cadence increased from 95 to 125 steps/m. The Fig. 10. Resistance adaptation to gait speed and environment with fixed gain and delay ( $\Delta t = 0.25 \text{ s}, \kappa = -4$ ). generated torque/power shows a gradual change in response to the cadence change. Negative power generation is very small compared to positive power generation. It means that the interference caused by mismatch during the cadence change is minimized. Fig. 9(b) shows the gain and generated RMS torque and power relationships under given walking conditions ( $\Delta t = 0.25 \text{ s}, \kappa = 8$ , variable cadence from 95 to 125 steps/m). A linear proportional relationship between the generated torque/power and the cadence (speed) is observed. 5) Generalizability of Resistance With Fixed Control Parameters: Same one male subject wore the exoskeleton and walked with resistance. The control parameters were fixed with the same values ( $\Delta t = 0.25 \text{ s}, \kappa = -4$ ) under various gait speeds and walking environments. For the speed change experiment, the subject walked on the treadmill at different walking speeds. The treadmill speed increased from 1 km/h to 5 km/h in 1 km/h increments. The subject walked for 1 minute at each speed. We took the last 30 s of each condition to calculate the generated mean positive/negative power. For the experiment regarding environmental changes, the subject walked at a self-selected speed on different ground levels (ramp ascent $\rightarrow$ ramp descent $\rightarrow$ level $\rightarrow$ stair ascent $\rightarrow$ stair descent). We took the section corresponding to each condition to calculate the generated mean positive/negative power. Through this experiment, we showed how the proposed interaction control algorithm can be generalized to resist under various conditions by showing the adapted torque and power profiles with fixed control parameters (time-delay $\Delta t$ and gain $\kappa$ ). Fig. 10 shows the generated joint torque and positive and negative power. Fig. 10(a) illustrates how the generated torque profile changed as the treadmill speed varied, all with the fixed control parameters ( $\Delta t = 0.25~\rm s$ , $\kappa = -4$ ). For the proposed controller, the interaction torque was determined by two factors: the control parameters (feedback gain $\kappa$ , time-delay $\Delta t$ ) and the gait motion state y. Since the control parameters were fixed, the hip
motion (or range of motion) determines the output torque profile. As shown in Fig. 10(c), the generated torque trajectory changed with changes to the gait environment. Without changing the control parameters ( $\Delta t = 0.25~\rm s$ , $\kappa = -4$ ), the generated torque and power values varied with notable differences, except at a relatively high-speed walking (4, 5 km/h). In Fig. 10(b) and Fig. 10(d), large negative power compared to small positive power indicates that the DOFC controller can operate under various conditions when resisting the user without assistance (pure interference). As shown in Fig. 7(a), maximum resistance torque is reached when the swing speed is high (when the legs cross) and the torque approaches zero when the swing speed is slow (when the swing direction of the leg changes). The above resistance strategy is advantageous because it may not be safe for a person who has difficulty walking to apply large resistance when changing the direction of the leg swing. (Normally at this time, even slight external force can be dangerous, because it passes through a zero point of human hip joint velocity.) C. Discussion of Limitations The proposed control method solely generates the assistance/resistance torque depending on the user's gait pattern. The interaction torque is limited to a sinusoidal shape as Fig. 7(a), and only the peak magnitude can be adjusted for flexion or extension motion, right or left joints. Therefore, it might be not suitable when you want to generate specific torque pattern (not sinusoidal) only in certain gait phase. However, the physical therapist (or medical doctor) may need to apply a well-defined torque only at specific gait phase. Another source of potential issues can derive from the fact that when you apply a significant amount of torque the angle joint profile is significantly affected by the interaction with the thigh/leg soft tissues. This problem could be reduced to some extent by smoothing the original angle sensor data, but there is still a possibility of instability when using large resistance torques. We have analyzed the stability of the DOFC controller under limited conditions in [21], but we need more rigorous theoretical/experimental studies on the stability of DOFC based assistance/resistance. Various parameter studies with a hip exoskeleton were performed on a single subject. This point should be considered as a limitation of this study. IV. CONCLUSION We presented an interaction control strategy for gait assistance and resistance. The DOFC-based controller stably handled gait pattern changes. By using the appropriate time-delay and feedback gain, we could generate assistive/resistive torque stably in the interaction between the user and exoskeleton. We showed the flexibility and effectiveness of the control method by measuring generated torque and power under various walking conditions and the metabolic energy expenditure from treadmill walking. We plan to measure metabolic costs and stabilize aspects of the controller under various conditions to show the controller's effectiveness across a wide array of conditions. We will extend the application of our interaction control to other types of exoskeletons, such as single joint driven assistance (e.g., an ankle exoskeleton device worn only on one foot). We expect to be able to apply a similar strategy, provided we can define state variables that characterize ankle joint motion and assistive torque. To guarantee consistent control under extreme conditions, such as running, we also have a plan to use online parameter modification.
MEDICAL ROBOTS Human-in-the-loop optimization of hip assistance with a soft exosuit during walking Ye Ding, 1,2* Myunghee Kim, 1,2* Scott Kuindersma, 1† Conor J. Walsh 1,2† Wearable robotic devices have been shown to substantially reduce the energy expenditure of human walking. However, response variance between participants for fixed control strategies can be high, leading to the hypothesis that individualized controllers could further improve walking economy. Recent studies on human-in-the-loop (HIL) control optimization have elucidated several practical challenges, such as long experimental protocols and low signal-to-noise ratios. Here, we used Bayesian optimization—an algorithm well suited to optimizing noisy performance signals with very limited data—to identify the peak and offset timing of hip extension assistance that minimizes the energy expenditure of walking with a textile-based wearable device. Optimal peak and offset timing were found over an average of $21.4 \pm 1.0$ min and reduced metabolic cost by $17.4 \pm 3.2\%$ compared with walking without the device (mean $\pm$ SEM), which represents an improvement of more than 60% on metabolic reduction compared with state-of-the-art devices that only assist hip extension. In addition, our results provide evidence for participant-specific metabolic distributions with respect to peak and offset timing and metabolic landscapes, lending support to the hypothesis that individualized control strategies can offer substantial benefits over fixed control strategies. These results also suggest that this method could have practical impact on improving the performance of wearable robotic devices. Copyright © 2018 The Authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original U.S. Government Works INTRODUCTION Wearable robotic devices have demonstrated the potential to enhance human economy and endurance (1-3). Recent breakthroughs in wearable robotics have substantially reduced energy expenditure in human walking using both passive (1) and active autonomous (2–4) or tethered (5-11) devices. In particular, advances in active devices provided flexibility to regulate assistance control parameters related to timing (5, 7, 10), magnitude (6, 12), or delivered power (12, 13). Studies have shown that control strategies can significantly affect performance (4-7, 10, 12), which raises questions about how to reliably and efficiently design optimal controllers. Assistive strategies have commonly been derived from simulations (14, 15) and biomechanical measurements (10, 16) or tuned manually based on average responses (11). Specifically, there is a growing interest in designing control strategies using musculoskeletal simulations (14, 17), and recently, this approach has shown promise in guiding assistive profiles for running (14, 18). However, physiological and neurological differences between individuals can cause divergent responses to an identical controller, that is, one participant's optimal control strategy may perform poorly on another (5-7, 10, 19). Thus, although generic musculoskeletal simulations may provide general guidelines on assistance, participant-specific models may be required when considering how to find optimal system parameters for individualized assistance. Conventionally, discrete step (1,4-7) and continuous sweep (20,21) protocols have been used to investigate a participant's performance and to explore the landscape of control parameter settings for wearable robotic devices. With these approaches, metabolic cost is measured by varying a control parameter in either a discrete or a continuous manner. A curve fitting process is then followed to identify the optimal parameter value that results in the maximum metabolic benefit (1,5-7,20,21). Unfortunately, both continuous and discrete protocols involve a lengthy evaluation process, and the time required increases exponentially with added control parameter dimensions. Long walking times in protocols may affect the accuracy of metabolic measurements due to high exertion or fatigue, which in turn leads to cardiopulmonary drift (22), especially for clinical populations who may not be able to sustain long walking bouts (23). Human-in-the-loop (HIL) optimization aims to address the aforementioned challenges in protocol length by adjusting control parameters based on real-time measurements of human physiological signals, such as metabolic cost. This optimization is inspired by an observation of humans continuously adjusting their coordination pattern to minimize the metabolic cost of walking (24) and expands the concept to wearable devices. Some promising efforts in this domain have recently demonstrated the ability to optimize both single and multiple control parameters using ankle exoskeletons (8, 9). In these cases, substantial metabolic reductions were achieved with the optimal parameter settings identified by either a one-dimensional (1D) gradient descent method using a pneumatically actuated ankle exoskeleton for a fixed 50 min (9) or a 4D Covariance Matrix Adaptation Evolution Strategy (CMA-ES) with an electromechanically actuated ankle exoskeleton for $83 \pm 14 \text{ min (mean} \pm \text{SEM)}$ (8
). Although these achievements are impressive, there remain opportunities to explore different wearable assistive hardware, control parameterizations, applications to other joints, and alternative optimization methods that could improve sample efficiency. We developed an experimental method to rapidly identify optimal control parameters in a 2D space that minimized the metabolic cost of walking (Fig. 1). This was achieved through the use of Bayesian optimization, an efficient global optimization strategy that is well suited to find the minima of objective functions that are noisy and expensive to evaluate (25–27). In a previous HIL study that optimized step frequency, we found that Bayesian optimization converged in half the time of a gradient descent method (28). For the current HIL scheme, a participant walked with hip extension assistance applied via a soft exosuit (Fig. 2A), a textile-based wearable device designed to apply forces across joints in parallel with human muscles (10, 29). The assistive profile was configured by multiple control parameters that were iteratively updated by the optimization and applied to the participant using a tethered actuation system with admittance force control (29). The optimization was initialized by <sup>1John A. Paulson School of Engineering and Applied Sciences, Harvard University, Cambridge, MA 02138, USA. 2Wyss Institute for Biologically Inspired Engineering, Harvard University, Cambridge, MA 02115, USA. *These authors contributed equally to this work. †Corresponding author. Email: walsh@seas.harvard.edu (CJ.W.); scottk@seas.harvard.edu (CJ.W.); Fig. 1. Experimental setup for HIL Bayesian optimization. Bayesian optimization was used to adjust the control parameters of an assistive device to minimize the metabolic cost of walking. The metabolic rate was estimated from respiratory measurements and used to compute a posterior distribution of metabolic rate with respect to the free control parameters. The posterior was initially generated by evaluating six prefixed control parameters. Given the posterior at the current iteration, the control parameters with maximum EI were chosen and applied to the wearable device. This process was repeated until convergence. During this process, the configured force profiles were delivered through a soft hip exosuit with a tethered actuation system. Fig. 2. Soft exosuit and assistive hip force profile. (A) The hip soft exosuit. A hip extension moment was generated by pulling the inner cable to create a tension between two anchor points. (B) Parameterization of hip force profile. The hip force profile was chosen to be a combination of two parameterized sinusoidal curves joined at the peak. Peak force was set to 30% of body weight, and onset timing was fixed to the time of maximum hip flexion. Peak and offset timing were actively adjusted by the optimization to determine the shape of the force profile as a function of gait percentage. Shaded purple and blue bars represent the range of peak and offset timing, respectively. (C) Examples of feasible hip force profiles. obtaining the metabolic cost of a prescribed number of assisted conditions with respect to pseudo-randomly selected control parameters from an evenly distributed parameter space. On the basis of this information, the optimization iteratively estimated the participant's metabolic cost distribution using a Gaussian process (fig. S1) (30) and selected control parameters for the next iteration by maximizing the expected improvement (EI) (25, 26). At each iteration, the metabolic cost was estimated by fitting a first-order dynamic model to 2 min of transient metabolic data (31). After a set number of iterations, the control parameters corresponding to the minimum value of the metabolic landscape (the mean of the metabolic cost distribution) represented the optimal values. The hip assistive profile was a combination of two halves of sinusoidal curves joined at their peaks. This profile was defined by two fixed parameters (peak force and onset timing) and two free parameters (peak timing and offset timing; Fig. 2B) that were adjusted by the optimization method. We fixed peak force to 30% body weight to ensure comfort during the long walking test while still maintaining assistance high enough to achieve sufficient metabolic reduction. Previous work demonstrated that higher assistance magnitude resulted in larger metabolic benefits for both hip (10, 11) and ankle (6), and the forces we evaluated here were approximately within the range of previous evaluations. We also fixed the onset timing at the maximum hip flexion event based on previous hip studies, which showed the largest metabolic reductions with onset timing set close to maximum hip flexion (7, 10). Further, we found less intraparticipant variability of metabolic cost and lower signal-to-noise ratio when varying onset timing in our pilot testing (table S1). For the purposes of this study, we defined the start of the gait cycle using the maximum hip flexion event. Peak and offset timing were bounded within 15 to 40% and 30 to 55%, respectively, in this newly defined gait cycle. It is worth noting that the maximum hip flexion event was on average at 86
.2% of the conventional gait cycle, defined with heel strike as 0% (table S2). The offset timing was constrained to occur at least 15% later than the peak timing. The range and constraint of peak and offset timing (Fig. 2C) were chosen by slightly extending the average range of the biological hip extension moment (32) while considering limitations on the ramp-up speed of assistance that our soft exosuit was capable of achieving. This configuration was able to shape our profiles similar to the hip assistive profiles used in previous assistive device studies (4, 7, 10, 11). We conducted a single-day experiment on eight participants (table S3), optimizing the assistance timings as they walked on a treadmill at 1.25 m s-1. For the optimization, 6 iterations (six pairs of peak and offset timing) were evaluated for the optimization initialization and 14 iterations were followed to adjust the tuning. These numbers of iterations were chosen based on our simulation results (fig. S2). After optimization, we performed a validation test confirming the optimal condition found during the optimization process, then compared both optimal condition and validation test with a no-suit condition. The primary analysis included (i) the net metabolic cost of walking, defined as the gross metabolic rate during walking minus the rate measured during quiet standing; (ii) the convergence time across participants; (iii) participant-specific optimal timings; (iv) participant-specific optimal assistive profiles; (v) participant-specific metabolic landscape, defined as the mean of the metabolic cost distribution with respect to the peak and offset timing along with participant-specific probability of improvement landscape, interpreted as the likelihood of exceeding the largest metabolic reduction. RESULTS Metabolic rate Participant-specific optimal assistance substantially improved energy economy for all participants by reducing the net metabolic cost of walking to $2.26 \pm 0.13 \text{ W kg}^{-1}$ and $2.27 \pm 0.18 \text{ W kg}^{-1}$ for the optimal and validation conditions, respectively, from $2.75 \pm 0.18 \text{ W kg}^{-1}$ for the nosuit condition (mean $\pm$ SEM). Net metabolic reduction of the validation condition ranged from 6.7 to 33.9%, with an average reduction of 17.4 $\pm$ 3.2% (mean $\pm$ SEM; paired t test, t = 0.003; Fig. 3A and table S4). Convergence time The optimization converged for all participants during the optimization process (fig. S3). The convergence time was on average $21.4 \pm 1.0$ min (mean $\pm$ SEM), ranging from 18 to 24 min. Optimal timing Participant-specific optimal peak and offset timings spread over about half of the feasible region of the control parameters (Fig. 3B). Most of the optimal timings were on the boundaries of the parameter ranges, with three participants having their optima at the latest peak and offset timing. Fig. 3. Experimental results. (A) The net metabolic rate for each condition. Optimal: Minimum mean value of the posterior distribution (metabolic landscape). Validation: Metabolic rate of 5-min walking with optimized assistance. No-suit: Metabolic rate of 5-min walking with a regular pair of pants. Bars are means, error bars are SEMs, and asterisks denote statistical significance. (B) Feasible parameter region and optimal timing values for all participants. Optimal timings were varied across participants, and three participants shared the same optimal timings at the latest peak and offset timing. (C) Optimal assistive force profiles for participants 3, 4, and 6. Dashed and solid lines are reference and measured forces normalized by body mass, averaged across 10 strides during the last minute of the validation condition. The maximum hip flexion event was used to initialize the gait cycle in this study. Optimal assistive force profile For the validation condition, the averaged delivered peak force was $215.6 \pm 10.1 \text{ N} (2.84 \pm 0.02 \text{ N kg}^{-1}, \text{ mean} \pm \text{SEM})$ . The average root mean square error of the optimal assistive force tracking of the validation condition was 4.1%. For a clear representation, only three representative optimal force tracking samples with the most different optimal timings from the validation condition are shown in Fig. 4C, whereas all optimal force profiles are shown in fig. S4. Metabolic landscape and probability of improvement landscape The representative participant-specific metabolic landscapes (Fig. 4, A to C) further illustrated the interparticipant variability with respect to timings. The participants' metabolic landscapes, represented as Gaussian process posteriors, showed substantial visual differences.
To quantitatively summarize the differences between the participants' metabolic landscapes, we computed the probability that each participant's optimal parameters would reduce the metabolic cost of other participants according to each participant's posterior landscape (Fig. 4, D to F). This analysis suggested that, in general, one participant's optimal peak and offset timing were likely to be suboptimal for another. DISCUSSION With the optimized hip extension assistance obtained from HIL Bayesian optimization, the average net metabolic reduction was 17.4% compared with walking without the device. Using a similar hip exosuit to assist loaded walking, our group previously showed an average reduction of 8.5% compared with an unpowered condition (10). Another study with a tethered hip exoskeleton using pneumatic actuators demonstrated average metabolic reductions of 10.3 and 9.7% when assisting either hip extension or hip flexion, respectively, compared with an unpowered condition (7). Last, a recent study assisting both hip extension and flexion simultaneously with an autonomous electromechanical hip exoskeleton reported an average metabolic reduction of 21.1% when compared with walking without an exoskeleton (4). The result of this study suggests that substantial metabolic reductions can be achieved by solely assisting hip extension with optimized assistance and indicates the potential improvement of assisting both flexion and extension with hip assistive devices. The average convergence time of our HIL Bayesian optimization was 21.4 min. Short convergence time could be important in some cases to mitigate widely observed inaccuracies stemming from cardiopulmonary drift and participant fatigue (22). This result also suggests that HIL Bayesian optimization could be applied to wearable devices designed for strenuous activities or clinical populations with limited physical strength—both cases where participant endurance is a limiting factor (23). The variability shown in the optimized assistance profiles demonstrates the importance of individualization. The participant-specific metabolic landscapes and the probability of improvement generated by the Bayesian optimization further illustrate the interparticipant variability with respect to timings. The nonparametric HIL Bayesian optimization was more effective than the model-based naïve grid search. We illustrated this problem by generating quadratic approximations with the first 10 iterations of the data from the optimization process, which was the average amount of data needed for the convergence of Bayesian optimization (table S5). The comparison showed that the model-based naïve grid search made unreasonable estimates of optimal parameter value in the high measurement noise environment. Fig. 4. Participant-specific metabolic landscape and probability of improvement landscape. (A to C) Metabolic landscapes (the mean of the metabolic cost posterior distribution with respect to peak and offset timing) for participants 3, 4, and 6. Diamonds indicate the locations of participant-specific optimal timings. (D to F) Probability of improvement landscapes (capturing the probability of reducing metabolic cost beyond the identified optimal) for participants 3, 4, and 6. The optimized assistance did not maximize the duration of force to maximize the positive mechanical power for the hip joint. This may be partially because assistance with a late offset timing may hinder hip flexion (32). However, most of the optimal timings for participants were on the boundaries of the parameter ranges, which may suggest that, with a larger parameter search area, further reductions in metabolic cost could be obtained. Currently, the selected parameter range was constrained by the limited ramp-up speed of assistance with our exosuit and a cautious approach to ensure that the assistance profile did not greatly exceed limits of the average range of biological hip extension moment (32). To reduce the likelihood that the optimal values are caught on the boundaries, future studies could expand the feasible parameter range by improving the exosuit stiffness to increase the ramp-up speed of the assistance and having participant-specific search areas based on training performance. Another limitation of the current optimization is the lengthy sampling time for each measurement, which could prevent straightforward extension to higher dimensional parameterizations. It may be beneficial to add additional flexibility to the optimization not only to choose the exploration points but also to adjust the length of sampling time (33). In addition to adaptive sampling time, it may be useful to use musculoskeletal models to provide an initial estimation of the metabolic landscape, which could reduce the number of samples required to find low-cost parameters. In addition, the smoothness and regularity assumptions imposed by the Gaussian process kernel function may not be valid for all metabolic landscapes and wearable devices, but in our experiments, these landscapes were well approximated using a squared exponential kernel and a single global noise parameter. Last, because Bayesian optimization uses all available data to compute the posterior metabolic distribution and acquisition function, additional methods such as "data forgetting" would have to be used to deal with human adaption effects (34). HIL optimization holds promise to improve the performance of wearable robotic devices for a wide range of tasks. The presented method shows a substantial metabolic reduction and suggests the possibility of optimizing wearable devices using low-dimensional control parameterization. The short convergence time would enable researchers to apply this method to individualize control parameters during strenuous tasks or for people with limited physical strength or endurance. Using
a noisy respiratory signal as the objective function of the optimization indicates that this method can be applied to other alternate physiological or biological signals, such as using kinematic symmetry to optimize wearable devices for poststroke patients or using balancerelated measurements to optimize prostheses. The participant-specific metabolic landscapes and probability of improvement landscapes demonstrate the significant variability between participants and suggest that participant-specific optimal timing provides the highest probability of achieving the largest metabolic reduction, further highlighting the benefit of individualization. MATERIALS AND METHODS Experimental design This was a single-day protocol without training sessions. To minimize the effects of adaptation, we recruited eight participants who had previous experience walking with the exosuit at least two times before. Participants walked without load on an instrumented treadmill (Bertec) at 1.25 m s−1 wearing a respiratory measurement device (COSMED; fig. S5). These conditions were chosen partially to lessen fatigue effects of the relatively long walking protocol, and the constant walking speed allowed the comparisons between studies (7, 10, 11). Each participant went through five conditions (fig. S6): (i) a 5-min quiet standing condition, (ii) a 5-min no-suit condition, (iii) a 40-min optimization condition intersected by two 3-min warm-up periods and 5-min rest periods, (iv) a 5-min validation condition with the optimal timing, and (v) a 5-min no-suit condition. Both warm-up periods were assisted walking with the same assistive profiles used in the follow-up iteration of the optimization condition. During the no-suit condition, participants walked with a regular pair of pants (mass, 715 g), which was chosen to assess the metabolic benefits from walking with active assistance to walking with normal clothes, similar to configuration in our previous hip assistance study (10). Resting breaks were given between all conditions besides the break during the optimization condition. Considering the relatively long walking time (61 min), two no-suit conditions were designed at both the beginning and the end of the protocol as a visual check of the possible fatigue reported by the participants. Participants Eight healthy male adults (n = 8; age, 30.3 ± 7.1 years; mass, 76.5 ± 8.9 kg; height, 1.77 ± 0.05 m; mean ± SD; table S3) participated in this study. Sample size was chosen based on the data from previous studies (10, 11). The study was approved by the Harvard Longwood Medical Area Institutional Review Board, and all methods were carried out in accordance with the approved study protocol. All participants provided written informed consent before their participation and after the nature and the possible consequences of the studies were explained. Soft exosuit The soft exosuit used in this study was designed to solely assist hip extension. The textile components of the hip exosuit consisted of a spandex base layer (mass, 181 g), a waist belt (mass, 275 g; fig. S7), two thigh braces (mass, 2 × 69 g; fig. S8), and two elastic straps (mass, 2 × 46 g) for mounting inertial measurement units (IMUs; mass, 2 × 13 g). Bowden cables and sensor wires including expandable braided cable sleeves for each leg (mass, 2 × 328 g) were tied together at the waist and connected to the actuation platform. The participant supported about half of the weight of the Bowden cable assembly. All textile components (size medium) and half of the weight of the Bowden cable assembly had a total mass of 0.859 kg. The stiffness evaluation of the soft exosuit used in this study is shown in (29). Actuation platform A tethered actuation system with two modular actuators was used to generate assistive forces. Each actuator consisted of one customized frameless brushless motor (Allied Motion), a customized spiroid gear set (ITW Heartland), a 90-mm-diameter pulley, and other supportive structures (29). Bowden cable was used to transmit the force from the actuator to the hip joint. On the actuator side, the Bowden cable sheath connected to the frame of the pulley cover and the inner cable attached to the pulley. On the exosuit side, the Bowden cable sheath connected to the anchor point on the bottom of the waist belt and the inner cable connected to the anchor point on the top of the thigh piece. When the actuator retracts, the distance between the two anchor points is shortened, generating a force to assist hip extension. Sensing and control Two IMUs (VN-100 Rugged IMU, VectorNav Technologies) attached to the front of each thigh detected the maximum thigh flexion angle to segment the stride (10, 35). Stride time was measured as the time between two consecutive maximum hip flexion events (35). By using the average stride time from
the previous two steps, the reference force profile was scaled for each stride. The actual force signal was measured by two load cells (LSB200, FUTEK Advanced Sensor Technology) placed in series with the Bowden cables on each leg. Combined with the actuator position signals measured by the encoders (AS5134, Ams) mounted on the back of the customized brushless motors, an admittance controller with feedforward models was implemented to track the force profiles with different peak and offset timings. The detailed controller design, frequency response, and force tracking evaluation with different ramping speeds are presented in (29). Instantaneous metabolic estimation The metabolic rate was estimated by fitting a first-order dynamic model to 2 min of transient metabolic data (21). The mathematical representation in the frequency domain takes the form $$Z(s) = H(s)R(s) \tag{1}$$ where Z(s) is the measured metabolic cost, R(s) is the instantaneous metabolic cost $f^{\text{inst}}$ in frequency domain, and H(s) is the first-order dynamic model $H(s) = 1/(\tau s + 1)$ with a time constant $\tau = 42$ s (31). In the discrete-time domain, Eq. 1 can be written as $$z(i+1) = \frac{(\tau - dt(i))}{\tau} z(i) + \frac{dt(i)}{\tau} f^{\text{inst}}$$ (2) where i is the number index of the measured breath and dt(i) is the time duration between the ith and (i+1)th breath. After measuring z and dt for 2 min, we obtained $f^{inst}$ by first calculating the change of the instantaneous metabolic rate from the last condition and then minimizing the error between the model estimation and measurements using least squares (21). Bayesian optimization Bayesian optimization is an efficient global optimization method that is particularly well suited to optimizing unknown objective functions that are expensive to evaluate (25–27, 36). It takes advantage of the information provided by the time history by computing a posterior distribution of cost as a function of the optimization variables and then using acquisition functions computed on this posterior to select the next points to evaluate. A prior belief over the objective function distribution is defined using mean and covariance functions. The posterior distribution of the objective function is iteratively computed in closed form when new data become available. Using this model, the algorithm balances exploitation with uncertainty reduction to guide exploration (37). In our study, we initialized the optimization by evaluating instantaneous metabolic cost $f^{\rm inst}$ for six iterations with different pairs of prefixed peak and offset timing, which were pseudo-randomly selected from evenly spaced timing intervals (fig. S9). This initialization is a common practice to avoid biased sampling that could lead to premature convergence (25). After initial evaluation, the optimization calculated the metabolic landscape, $f(\mathbf{x})$ , using Gaussian processes (30, 34), where the parameter $\mathbf{x} = [x_p, x_o]$ consisted of peak and offset timing. Given the calculated landscape, the next sampling timing was selected by maximizing EI, which naturally balances exploration and exploitation (25, 26). With the metabolic rate of the newly sampled timing added to the data set, the metabolic landscape was refined again for selecting the next sampling timing. This process was repeated for 14 iterations. In total, there were 20 iterations in the optimization process including 6 iterations of initialization, and fig. S10 shows one sample optimization process described above on iterations 6, 7, and 20. The metabolic landscape, $f(\mathbf{x})$ , was modeled using a Gaussian process. The prior of the Gaussian process is represented by mean, $\mu(\mathbf{x})$ , and covariance, $k(\mathbf{x}, \mathbf{x}')$ , functions. As is standard practice, we used zero mean and the anisotropic squared exponential kernel for the covariance function (25), $$k(\mathbf{x}, \mathbf{x}') = \sigma^2 \exp\left(-\frac{1}{2}(\mathbf{x} - \mathbf{x}')M(\mathbf{x} - \mathbf{x}')\right)$$ (3) where $\sigma^2$ is the metabolic rate (signal) variance and M is a diagonal matrix consisting of the length scale parameters of peak and offset timing, $l_1$ and $l_2$ . Intuitively, the signal variance captures the overall magnitude of the cost function variation, and the length scales capture the sensitivity of the metabolic rate with respect to changes in peak and offset timing. Assuming that metabolic cost has an additive, independent, and identically distributed noise, the samples can be expressed as $$f^{\text{inst}}(\mathbf{x}) = f(\mathbf{x
}) + \varepsilon, \varepsilon \sim N(0, \sigma_{\text{noise}}^2)$$ (4) where $\sigma_{\text{noise}}^2$ is the noise variance. Given the Gaussian process prior and data set D, the posterior metabolic cost distribution $f_^{\text{inst}}$ was calculated for a parameter $\mathbf{x}_$ as $f_^{\text{inst}}(\mathbf{x}_) \equiv f_^{\text{inst}} \sim N(E[f_^{\text{inst}}], s_^2)$ . The mean and variance are calculated as $$E[f_^{\text{inst}}] = \mathbf{k}_^T (K + \sigma_{\text{paise}}^2 I)^{-1} \mathbf{y}$$ (5) $$s_^2 = k(\mathbf{x}_, \mathbf{x}') - \mathbf{k}_^T (K + \sigma{\text{noise}}^2 I)^{-1} \mathbf{k}_$$ (6) where $\mathbf{k} = [k(\mathbf{x}1, \mathbf{x}), ..., k(\mathbf{x}_n, \mathbf{x})]'$ and K is the positive definite kernel matrix, $[K]_{ij} = k(\mathbf{x}_i, \mathbf{x}_j)$ . We optimized hyperparameters $(\theta = [\sigma \ l_1 \ l_2 \ \sigma_{\text{noise}}])$ at each iteration by maximizing log marginal likelihood of the data collected $(\mathbf{D} = {\mathbf{X}, \mathbf{y}}, \mathbf{X} = [\mathbf{x}_1, ..., \mathbf{x}_n]^T \in R^{N \times 2}, \mathbf{y} = [f_1^{\text{inst}}, ..., f_n^{\text{inst}}]^T \in R^N)$ using Matlab's fmincon function with 10 random initializations to avoid poor local minima The peak and offset timing, $x_{\rm p}$ , $x_{\rm o}$ , were selected by maximizing the expected reduction in the metabolic cost over the best timing previously assessed $\max(f_{\rm best}-f_^{\rm inst},0)$ (25). EI, which balanced between predictive minimum points and high uncertainty (25, 27), took the following form $$EI[\mathbf{x}] = (f_{best} - E[f_^{inst}])\phi(u) + s_\phi(u_*)$$ (7) where $f_{\text{best}} = \min_{i=1,\dots,N} \mathrm{E}[f^{\text{inst}}(\mathbf{x_i})], \ u_ = (f_{\text{best}} - \mathrm{E}[f^{\text{inst}}_])/s_, \text{ and } \phi(\cdot)$ and $\phi(\cdot)$ were the cumulative distribution function and probability density function of the normal distribution, respectively. The EI was set to zero when $s_$ was zero. At each iteration, the next sampling timing was selected by maximizing EI using Matlab's fmincon while enforcing the constraint that the offset timing be at least 15% later than the peak timing, $x_o - x_p \geq 15\%$ . We again used 10 random restarts to avoid poor local minima. We note that, as the dimensionality increases, the number of random restart points required to reliably maximize EI would likely need to increase. Metabolic measurement and analysis Respiratory data were collected throughout the protocol. Metabolic rates from the quiet standing, first no-suit, validation, and second nosuit conditions were calculated from the last 2 min of carbon dioxide and oxygen rates using a modified Brockway equation (38). For the optimization process, the instantaneous metabolic estimations for each 2-min measurement period were also collected. Net metabolic rate and net metabolic landscape were obtained by subtracting the quiet standing metabolic rate, then normalizing by each participant's body mass. The metabolic reduction of the validation condition was obtained by subtracting the net metabolic rate of the validation condition from the net metabolic rate of the second no-suit condition and then dividing the result by the net metabolic rate of the second no-suit condition. The second no-suit condition was chosen for the comparison of metabolic reduction because it is the closest no-suit condition to the validation condition. The metabolic reduction of the optimal condition was obtained with the same calculation by replacing the net metabolic rate of the validation condition with the minimum value from the net metabolic landscape generated by the optimization. One participant's data were not included in the metabolic analysis because of fatigue reported by the participant during the
protocol, where the net metabolic rate of the second no-suit condition increased by 32.4% compared with the first no-suit condition. Convergence time analysis The convergence time for each participant was calculated in a post hoc analysis (fig. S3 and table S4). We defined the convergence of the optimization with the following two conditions: (i) Two consecutive iteration-to-iteration changes of maximum metabolic reduction in percentage from the metabolic landscape fell below our preset convergence threshold ( $t_{\rm m}$ = 4%), and (ii) two consecutive iteration-to-iteration changes of hyperparameters from the Gaussian process fell below our preset convergence threshold ( $t_h = 3$ ). The convergence threshold for the changes of metabolic reduction $t_m$ was chosen based on the previous study (8), which has shown an average error of 4% on this instantaneous metabolic estimation. The convergence threshold $t_h$ was obtained from a separate simulation study. For this simulation, a generative model of metabolic landscape with added noise was first created. The noise was generated by Matlab's awgn function with a signal-to-noise ratio of 8.8 obtained from our pilot test (table S1). With this model, we ran Bayesian optimization for 50 iterations and calculated the iteration-to-iteration changes in the maximum metabolic reduction and the hyperparameters. The maximum changes for the metabolic reduction were set to 2% while evaluating the changes of all hyperparameters. We repeated the simulation 100 times and found that the metabolic reduction threshold was met when the threshold for all hyperparameters was set to 3. Ground reaction force Ground reaction forces (GRFs) were collected via the instrumented split-belt treadmill (Beltec) and synced with the actuation platform using the motion capture system (Qualisys AB). All the GRF force data were filtered with a zero-lag fourth-order low-pass Butterworth filtered with a 5- to 15-Hz optimal cutoff frequency that was selected using a custom residual analysis algorithm (32). A customized Matlab script was created using GRFs to segment the percentage of the gait cycle defined by the maximum hip flexion based on the detected heel strikes. Statistics Means and SEM of the net metabolic rate were calculated for each condition. According to the Jarque-Bera test (significance level $\alpha=0.05$ ; Matlab), the collected data followed the normal distribution (P>0.3). Therefore, we conducted a mixed-model, two-factor analysis of variance (ANOVA; random effect, participant; fixed effect, test condition to test the effect across different conditions including optimal, validation, and no-suit conditions (significance level $\alpha=0.05$ ; Matlab). For the outcome of the ANOVA test, it showed a significant difference of the net metabolic rate between conditions. We used paired t tests for the comparison between the conditions to identify which conditions exacted a significant change in the net metabolic rate (39). SUPPLEMENTARY MATERIALS robotics.sciencemag.org/cgi/content/full/3/15/eaar5438/DC1 Fig. S1. Illustration of 1D Gaussian process. Fig. S2. Simulation results on the number of iterations needed for the optimization. Fig. S3. Convergence analysis. Fig. S4. Optimized hip extension force profiles for all participants. Fig. S5. Experimental setup. Fig. S6. Experimental protocol. Fig. S7. Structure of the waist belt component. Fig. S8. Structure of the thigh brace. Fig. S9. Pseudo-randomly sampled timings for the initialization of Bayesian optimization. Fig. S10. Optimization process. Table S1. Signal-to-noise ratio and variations of metabolic cost of pilot tests. Table S2. Onset timing. Table S3. Participant characteristics. Table S4. Metabolic rates, optimal timing, and convergence timing for each participant. Table S5. Quadratic approximation of metabolic landscape.
The Effect of Hip Assistance Levels on Human Energetic Cost Using Robotic Hip Exoskeletons Inseung Kang ®, Hsiang Hsu ®, and Aaron Young, Member, IEEE Abstract—In order for the lower limb exoskeletons to realize their considerable potential, a greater understanding of optimal assistive performance is required. While others have shown positive results, the fundamental question of how the exoskeleton interacts with the human remains unknown. Understanding the optimal assistance magnitude is not simply relevant for control, it is a critical knowledge for exoskeleton designers. An accurate understanding of assistance levels will enable the designers to minimize exoskeleton mass and improve the performance by avoiding excessive actuators and drivetrains. We explored the relationship between the assistance magnitude and the energetic cost benefits by using a series elastic actuator driven powered hip exoskeleton. The exoskeleton controller mimics a human biological hip moment to provide the assistance during the gait cycle. Ten able-bodied subjects walked using the exoskeleton with different magnitudes of assistance in both hip flexion and extension. Generally, the resulting metabolic cost across different assistance conditions showed a U-shape trend which was consistent across all subjects (p < 0.01). The interpreted optimal assistance point through the quadratic fit resulted in a 6% metabolic cost reduction with respect to the noassistance condition. The study validated that simply increasing the assistance level did not yield higher energetic return. Index Terms—Wearable robots, human performance augmentation, robotic exoskeleton, energetic cost, hip orthosis. I. INTRODUCTION OWER limb exoskeleton technology has advanced greatly in the recent years and showed significant value in different applications [1]–[3]. Most of these technologies can be broken down into three main categories: industry settings, military purposes, and healthcare environments [1]. Several exoskeletons have been developed for industry usages where factory workers wear an exoskeleton suit to enhance physical strength to alleviate work load when lifting heavy weights [4], [5]. Another usage of exoskeleton technology is in the military settings where an exoskeleton can assist a soldier in safely and efficiently carry a Manuscript received July 18, 2018; accepted December 14, 2018. Date of publication January 4, 2019; date of current version January 16, 2019. This letter was recommended for publication by Associate Editor R. V. Patel and Editor A. Young upon evaluation of the reviewers' comments. This work was supported in part by the Georgia Tech Research Institute (GTRI) IRAD funding, in part by the Institute for Robotics and Intelligent Machines (IRIM) Seed Grant at Georgia Tech, and in part by the NSF NRI Award #1830215. (Corresponding author: Inseung Kang.) The authors are with the Woodruff School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA 30332 USA (e-mail: ikang7@gatech.edu; hhsu24@gatech.edu; aaron.young@me.gatech.edu). This letter has supplementary downloadable multimedia material available at http://ieeexplore.ieee.org, provided by the author. This includes a video file, which includes three parts about the powered hip exoskeleton device. This material is 41.1 MB in size. Digital Object Identifier 10.1109/LRA.2019.2890896 heavy load over long such as when a solider would carry heavy loads while walking over long distances [6], [7]. Lastly, exoskeletons are used in the healthcare environment as an assistive device for patients with disabilities such as stroke, spinal cord injury, and muscular dystrophy [8]–[10]. Patients using these technologies in healthcare settings can not only regain mobility options, but also potentially benefit from long-term rehabilitation strategies [11]. Some of these exoskeletons such as Indego [8], Rewalk [9], and Ekso [12] have been commercialized in the market for medical applications. These devices frequently feature actuators at every joint, which makes the devices heavy and challenging to control. While these devices may benefit patients with a complete lower limb paralysis, their benefit for less impaired subjects remains less clear [13]. To accommodate such limitations of full body exoskeletons, several research and industry groups started to develop single joint actuated exoskeletons that are more suitable to both ablebodied humans and patients with partial gait disability (i.e., stroke survivors) [14]–[17]. Often these exoskeletons have targeted the ankle due to the joint providing a high mechanical power during walking. Several ankle exoskeletons were able to show positive results in achieving higher metabolic benefits when walking with assistance at the ankle joint [14], [18]. Recent literature studies have shown that the hip joint also plays a leading role in providing high mechanical power, up to 45%, during walking [19]. The ankle joint efficiently utilizes the Achilles tendon unit in
storing the mechanical energy to perform positive joint power. However, due to different muscle characteristics and the lack of efficient elastic storage elements, the hip joint requires higher energetic cost for similar mechanical joint power [20]. Therefore, the hip joint represents an important area of exploration for engineers trying to increase human metabolic performance. Some of the hip exoskeletons that have been developed have shown positive outcomes [17], [21]–[25]. These exoskeletons are either autonomous with relatively low assistance levels or tethered to off board actuators to provide high magnitudes of assistance. Regardless, the majority of these devices have the same goal of achieving the best metabolic cost reduction. However, there is still a gap in understanding more in-depth relationships between energetic benefits and assistance The three key parameters that may contribute to the exoskeleton controller are onset timing, assistance duration, and assistance magnitude. Utilizing the assistance duration from the literature study [26], our previous work optimized the assistance onset timing for hip flexion and extension to achieve best metabolic cost reduction [27]. However, as the study provided a fixed magnitude of torque applied at the hip joint, it did not capture the effect of assistance magnitude in respect to the metabolic cost benefits. Thus, an extended study was required 2377-3766 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. Fig. 1. Powered bilateral hip exoskeleton design. (A) The orthotic thigh shells, pelvic band, and shoulder straps can be adjusted to accommodate different body sizes. (B) Exoskeleton device consists of multiple sensors and actuators. (C) The brushless DC motor drives a timing belt pulley to extend and retract the ball screw transmission (orange) where the motor housing (red) pivots around the hinge to create torque around the output hip joint (green). Deflection at the fiberglass spring provides corresponding measurement of the output torque. to understand the effect of magnitude changes when providing assistance during walking using a hip exoskeleton. This letter provides three core contributions to the field of assistive lower limb robotics. This letter presents a novel hip exoskeleton (Fig. 1A) utilizing a ball screw driven series elastic actuator (SEA) for power assistance at the hip joint in both hip flexion and extension along the sagittal plane. The device achieves a high fidelity closed loop torque control with a measurement of fiberglass spring deflection. This device presents an excellent test-bed for understanding assistance levels. Secondly, this letter presents a novel controller design that mimics a human biological hip torque profile during the gait cycle. The proposed controller is capable of applying different control parameters such as assistance onset timing, duration, and magnitude with the aid of different mechanical sensors on the device. Thirdly, this letter rigorously tests the design and controller through a ten human subject test to determine the effects of assistance level on energetic cost savings. These results highlight the experimental value of our test-bed system and illustrate that an optimal assistance magnitude indeed exists. The overall goal of our work is to determine the optimal assistance level applied at the hip joint using a powered exoskeleton during walking. The in depth understanding of relationship between the hip assistance levels and energetic cost savings using an exoskeleton will aid the field to move forward in developing an optimal exoskeleton device. We hypothesize that the increase of exoskeleton assistance provided at the hip joint will off load the work done by the hip flexor/extensor hence, reducing the metabolic cost of walking. II. POWERED HIP EXOSKELETON DESIGN We designed a bilateral hip exoskeleton device (Fig. 1B) to apply torque at the user's hip joint during walking. The overall mass of the device is around 7 kg where the main components of the device are actuators (1.5 kg each), main frame structure (2 kg), orthosis (1.5 kg for both waist band and thigh cuffs), and the on board electronics (0.5 kg). The device is capable of providing a peak torque of 60 Nm and a maximum continuous torque of 30 Nm with an angular velocity up to 180 °/sec. The desired exoskeleton peak and maximum continuous torque parameters were derived using Eq. 1 and Eq. 2 based on the biomechanic data [28]–[30] where $\tau(v)$ is the human biological hip moment over one gait cycle, v is the walking speed, m is the subject's body mass, and n is the number of moment data points over one gait cycle. $$\tau(v){peak} = \frac{m}{2} \left( \|\tau{\min}(v)\| + \tau_{\max}(v) \right) \tag{1}$$ $$\tau(v){cont} = m\sqrt{\frac{\sum{i=1}^{n} \tau(v)_i^2}{
n}}$$ (2) Using the both equations, our device can support approximately 90% of peak and maximum continuous torque for an average body mass of 70 kg subject walking at 1.2 m/s. The device allows 100° and 30° range of motion in the sagittal plane for maximum hip flexion and extension respectively. Additionally, a passive hinge joint allows for 15° of movement for both hip ab/adduction along the frontal plane. The entire device has three different attachment points to the user: orthotic thigh shell, pelvic band, and shoulder straps. Each attachment point can be adjusted to accommodate different body sizes. A. Mechatronic Design The powered bilateral hip exoskeleton consists of two ball screw driven series elastic actuators (SEAs) (Fig. 1C). The motor torque is initially speed reduced with a 2:1 timing belt transmission where the output torque is converted into a linear force with the ball nut travelling along the ball screw shaft with a carbon fiber tube attached to it. The ball screw transmission was chosen because it provides high efficiency and back drivability in a light weight package [31]. The entire actuator rotates along the pivot point located at the motor housing as the actuator extends or retracts. Lastly, the carbon fiber actuator is coupled in series with a fiberglass leaf spring where the deflection of the spring is measured with a set of strain gauges mounted in a form of full Wheatstone bridge. The use of a fiberglass spring enables reduced mass [32]. Each SEA is mounted to a main C-shaped frame made of carbon fiber, which ensures the orientation of the device is upright when the user is fitted. The user interface, which includes the thigh orthosis, the pelvic band, and the polycarbonate back plate, is attached to the main frame. The back plate holds the electronics as well as shoulder straps. The 200 W brushless DC motor rated at 36 V (EC 30, Maxon Motor) is controlled with a single board computer (myRIO, National Instrument). The myRIO is equipped with a FPGA chip for a closed loop torque control with a PD controller using the spring deflection reading measured with from the strain gauge (Omega Engineering). A servo driver (ESCON 50/5 Module, Maxon Motor) operates in current control mode and uses a Hall Fig. 2. Control system architecture of the exoskeleton device. On board computer interacts with different mechanical sensors on the device for high level control and commands the desired torque in currents to a servo driver for a low level closed loop torque control. effect sensor and an incremental motor shaft encoder for commutation. Both actuators are powered by two 18.5 V lithium polymer batteries (Venom Power) connected in series. A 14-bit absolute magnetic encoder (Orbis, Renishaw) is used to measure the hip joint angle. Three inertial measurement units (IMUs) (Micro USB, Yost Lab) are also mounted on the device. Two IMUs are placed bilaterally on each thigh and one is on the back unit to measure the limb and trunk orientation during walking. Lastly, force sensitive resistors (FSRs) are placed on each heel of the user to detect the heel contact during walking. An additional custom made printed circuit board is used for reading and filtering the analog sensor signals such as strain gauge amplifier for spring measurements. All of the on board mechanical sensor data are collected by myRIO for high and mid-level control such as estimating the gait phase and generating the torque assistance profile. B. Controller Design The exoskeleton device control layer is broken down into three tiers for its purpose: high, mid, and low-level layers (Fig. 2). The high-level layer implements an algorithm to estimate the user's state such as gait phase. This aspect is critical as it provides information regarding the timing for power assistance. While it is mainly used for estimating the gait phase in this study, this layer can be enhanced with additional control algorithms such as classifying the user's intent with machine learning Fig. 3. Example of a commanded torque profile for a biological torque controller. Control profile (shown in purple) emulates a percentage of human biological hip moment (shown in green) over a gait cycle. Hip flexion (red region) and extension (blue region) assistance onset timing, duration, and magnitude (shown with different type of arrows) can be tuned for a desired profile. techniques. The mid-level layer dictates the device dynamic performance. This layer can implement different dynamic controllers used in the literature such as admittance, position, and myoelectric control [22], [33], [34]. For this study, our exoskeleton incorporates a torque controller where the commanded torque profile is generated with given timing and magnitude parameters with additional user state information computed from the high-level layer. Lastly, the low-level layer ensures that the output torque meets the desired torque by applying a closed loop torque control. Our biological torque controller mimics the human biological hip moment profile (Fig.
3). Over a gait cycle, the controller can generate a torque assistance for both hip flexion and extension with predefined control parameters. The three key parameters that dictate the assistance profile are onset timing, assistance duration, and assistance magnitude. The onset timing parameter was used with values found from our previous work [27] which was 45% and 90% of the gait cycle for hip flexion and extension respectively (where 0% is defined as heel contact). Assistance duration was optimized with an initial pilot test. Lastly, assistance magnitude can be changed relative to percentage of peak human biological hip moment during the gait cycle [29]. The peak biological hip moment value was normalized to the user's bodyweight to provide appropriate assistance magnitude. The biological torque controller (Eq. 3) intakes a gait phase percentage, x, as an input and outputs a commanded torque, f(x), where of and oe are flexion and extension onset timing respectively and d is the assistance duration all relative to the gait phase. $$f(x) = \begin{cases} -g\left(x - o_f\right), & o_f \le x < \operatorname{mod}\left(\left(o_f + d\right), 100\right) \ g\left(x - o_e\right), & o_e \le x < \operatorname{mod}\left(\left(o_e + d\right), 100\right) \ 0, & \text{otherwise} \end{cases}$$ (3) g(z)represents a function generating a single trapezoidal profile (Eq. 4) with a desired assistance magnitude u, input phase z, assistance u starting and ending set point $s_1$ and $s_2$ respectively, all relative to the gait phase. The assistance magnitude u (Eq. 5) is computed using the desired assistance level a (%) and the peak human biological hip moment, $\tau(v)_{peak}$ , calculated from Eq. 1. $$g(z) = \begin{cases} \frac{uz}{s_1}, & 0 \le z < s_1 \ u, & s_1 \le z < s_2 \ -\frac{u(z-s_2)}{d-s_2} + u, & s_2 \le z < d \end{cases}$$ (4) $$u = a\tau(v)_{peak} \tag{5}$$ During the gait cycle where the assistance is not provided, the device is put into zero impedance mode which is simply generating the interaction torque that is measured between the user and the actuator output. Due to the gear reduction stage in the mechanical transmission in the SEA, it exerts non-negligible interaction torque to the user when walking which hinders natural movement. To minimize this resistance, the device outputs an equal amount of torque to allow the user to move the limb freely, hence called zero impedance mode. A video of a subject walking in both assistance and zero impedance mode is included in the supplemental material. As the biological torque controller is heavily dependent on the gait cycle for generating the correct assistance profile, it is critical to estimate the gait phase accurately in real time. We utilized the FSR sensor to estimate the user's gait phase during walking. While walking, the myRIO stores the timestamp when the heel contact occurs through the FSR sensor readings. From this, a stride time was calculated by taking the time difference between the heel strikes. Using the previous five stride times, we calculated the average stride duration. Finally, the time since the most recent heel contact was divided by this average stride duration in order to compute the current gait phase in percentage [35]. III. INITIAL HUMAN CHARACTERIZATION An initial study was conducted to characterize the device performance. During this process, several key factors contributing to the device performance were observed: the device impedance in terms of actuator interaction torque, controller performance in terms of zero impedance mode compensating the interaction torque, and the low-level controller performance regarding torque tracking of the closed loop torque control. A. Exoskeleton Human Performance Characterization We conducted a pilot test where three able-bodied subjects (body mass of 71.3 $\pm$ 5.5 kg) walked on a treadmill at 0.4 m/s with the exoskeleton to analyze the controller's capability in compensating the interaction torque during zero impedance mode. During this testing, we collected the user's metabolic cost of walking to observe the zero impedance mode performance in terms of human energetics. The low treadmill speed was mainly due to the limitations of walking with the device powered off. When the device is powered off, the actuator has a certain impedance (interaction torque), mainly from the gear transmission and friction, which impedes the user's hip movement. The treadmill speed was chosen carefully so that the user could still easily walk normally. Overall, the user walked in four different conditions: exoskeleton powered off, actuator off (exos
keleton without the SEA), exoskeleton in zero impedance mode, and no exoskeleton condition. We have added the actu- Fig. 4. Initial human characterization results of the exoskeleton device. (A) Actuator interaction torque compensation. When the actuator operates with zero impedance mode, the interaction torque (shown in purple) that is exerted due to mechanical impedance of the transmission in the SEA was greatly reduced compared to the unpowered mode. (B) Hip joint kinematics across different actuator conditions. Actuator off condition is where the spring is disconnected from the output hip joint to allow the user to freely move the hip joint without any impedance. ator off condition where the spring linkage was disconnected from the transmission so that the actuator output does not exert any interaction torque to the user. This condition can provide an additional baseline information (along with no exoskeleton condition) to compare the controller performance with the exoskeleton mass added to the user (but not the impedance). The average metabolic costs of walking, after subtracting out the resting metabolic cost, were measured to be 1.92 $\pm$ 0.25 W/kg, $1.33 \pm 0.13$ W/kg, $1.59 \pm 0.13$ W/kg, and $0.75 \pm 0.04$ W/kg for exoskeleton powered off, actuator off, exoskeleton in zero impedance mode, and no exoskeleton condition respectively. It was observed that the powered off condition can increase the net metabolic cost nearly twice compared to the no exoskeleton condition. According to the literature studies, theoretical increase of metabolic cost of walking can be evaluated with the exoskeleton mass added [36], [37]. Since the exoskeleton mass added to the user can be represented as the actuator off condition, metabolic cost result comparison between zero impedance mode and the actuator off condition could validate whether the zero impedance mode can be used as a baseline condition. While the zero impedance mode did not yield the same metabolic cost Fig. 5. Torque tracking of the assistance torque profile of the gait cycle. A 10 Nm torque was assisted to the user during walking with predefined control parameters (onset timing and assistance duration). The overall mean RMS error was calculated to be 2.15 Nm over a gait cycle. reduction as the actuator off condition (mainly due to the residual torque in zero impedance mode), it was capable of reducing the metabolic cost vastly and compensated the impedance caused from the actuator. While these metabolic cost measurements cannot be utilized directly in the main experiment mainly due to difference in walking speed, this validation process allowed to quantify the exoskeleton actuator performance in terms of metabolic cost. The metabolic cost compensation from the zero impedance mode aligned well with the kinetic compensation results (Fig. 4). The RMS interaction torque between the exoskeleton and the human user was reduced from 4.66 Nm to 0.94 Nm using the zero impedance controller during walking (Fig. 4A). During this pilot experiment, more in depth characterization of the actuator behavior was investigated by observing how the interaction torques affect the human gait biomechanics during walking, specifically the hip joint angle (Fig. 4B). The pilot results showed that the actuator output produces non-negligible interaction torque to the user. Moreover, this interaction torque showed that it impedes the user significantly and causes them to reduce the hip joint's range of motion. This was demonstrated in that the peak hip flexion and extension angles were both reduced by 5° each. However, with zero impedance mode, the user's kinematics were closely consistent with the actuator physically disconnected. This illustrates the importance of effective zero impedance mode. Another preliminary assessment of the device design was to quantify torque tracking of the desired torque with the actual interface torque (Fig. 5). This can be validated by observing the low-level control performance with torque tracking during assistance mode. During the pilot testing, a user was assisted with 10 Nm for both flexion and extension with prechosen controller parameters (onset timing and assistance duration) while walking on a treadmill at 0.4 m/s for 2 minutes. The assistance profile in terms of actual torque measured followed the desired torque correctly validating that the desired torque was assisted. Overall, the RMS error over the 2 minute trial compared to the desired torque profile was 2.15 Nm. A video displaying the low-level torque tracking with hip assistance is included in the supplemental material. Fig. 6. Experimental setup of the exoskeleton device testing. The user is fitted with a metabolic mask to measure the energy expenditure while wearing the hip exoskeleton. B. Assistance Duration Optimization Characterization A pilot test was conducted to determine the optimal duration of hip flexion and extension for the controller. The subject walked on the treadmill with 0.8 m/s walking speed for six minutes with 26% of the subject's bodyweight for assistance magnitude for both hip flexion and extension. The
assistance duration window length was swept from 20% to 35% of the gait cycle with a 5% increment. The metabolic cost reduction for each condition was 6.2% for 20%, 14.6% for 25%, 11.2% for 30%, and 11.0% for 35% all relative to metabolic cost of walking in zero impedance mode. As the 25% window length achieved the highest metabolic cost reduction, the value was used for both hip flexion and extension assistance duration for the main experiment. IV. HUMAN SUBJECT TESTING A. Experimental Design The study was approved by the Georgia Institute of Technology Institutional Review Board, and informed written consent was obtained for all subjects. Ten healthy subjects (seven males and three females) with an average age of 22.4 $\pm$ 2.0 years, body mass of $70.6 \pm 8.6$ kg, and height of $1.73 \pm 0.1$ m were recruited. The subjects were asked to walk on the treadmill (TuffTread) for six minutes at 0.8 m/s walking speed in four different assistance conditions where one of them was the zero impedance mode (Fig. 6). The other three assistance conditions are 13%, 26%, and 40% of the peak hip flexion/extension moment based on subject's bodyweight at 0.8 m/s walking speed [29]. The walking speed was set as 0.8 m/s due to device limitation in providing high torque at faster walking speeds. Other control parameters such as onset timing (45% and 90% of flexion and extension) and assistance duration (25% of the gait cycle) were fixed throughout the entire experiment for assistance conditions. During all walking conditions, metabolic cost of walking was measured using an indirect calorimetry system. Before subjects began walking, we measured their resting metabolic rate for 3 minutes while they stood still wearing the exoskeleton. The metabolic cost was calculated using the modified Brockway equation [38] for the last 3 minutes of each six-minute trial to determine the metabolic energy expenditure. Each walking measurement was subtracted with the metabolic cost of a \| AssistanceLevel (%) \| Joint Kinematic \| \| \| \| Joint Torque \| \| \| \| \|-------------------------\|---------------------\|-----------------------\|---------------------\|--------------------------\|----------------------------\|-------------------------\|---------------------------\|------------------------------------------\| \| \| Peakflexion (°) \| Peakextension (°) \| Range of motion (°) \| Peak flexion torque (Nm) \| Peak extension torque (Nm) \| RMS flexion torque (Nm) \| RMS extension torque (Nm) \| Net RMS torque overa gait cycle (Nm) \| \| 0 \| 32.07 \| -5.06 \| 37.13 \| 1.41 \| 0.95 \| 0.76 \| 0.61 \| 0.55 \| \| 13 \| 33.37 \| -5.43 \| 38.8 \| 4.82 \| 4.14 \| 3.41 \| 2.31 \| 2.87 \| \| 26 \| 35.37 \| -4.24 \| 39.62 \| 8.95 \| 8.73 \| 6.95 \| 6.23 \| 6.04 \| \| 40 \| 36.33 \| -4.16 \| 40.5 \| 12.81 \| 13.37 \| 10.24 \| 9.83 \| 8.89 \| TABLE I KINEMATIC AND KINETIC RESULTS ACROSS DIFFERENT EXOSKELETON ASSISTANCE LEVELS All of the joint kinematic and kinetic results are represented with an averaged value across 10 subjects. Flexion and extension joint torques were decoupled by computing the torque during flexion ( $45\% \sim 70\%$ ) and extension ( $90\% \sim 15\%$ ) region of the gait cycle. For zero impedance mode, flexion and extension torque were computed respect to 0 Nm to calculate the residual interaction torque. Fig. 7. Metabolic cost result with different assistance magnitude during walking. Generally, increase of assistance magnitude achieved decrease in net metabolic cost of walking except for the highest assistance condition (40% condition) where the metabolic cost increased. The quadratic fit of the metabolic result computed that the maximum metabolic cost can be benefited with 6% reduction from the zero impedance mode with 20% assistance magnitude. Error bar in the graph represents +/- 1 SEM. resting condition. Additional biomechanic data such as the hip joint torque and angle were
measured using the device's SEA and encoder during walking in each condition. All of the joint kinematic and kinetic results are represented with an averaged value across 10 subjects. Flexion and extension joint torques were decoupled by computing the torque during flexion ( $45\% \sim 70\%$ ) and extension ( $90\% \sim 15\%$ ) region of the gait cycle. For zero impedance mode, flexion and extension torque were computed respect to 0 Nm to calculate the residual interaction torque. A regression analysis was run to the overall metabolic cost of all subjects across different assist conditions to observe a general trend. Both linear and quadratic fits were run where each coefficient of determination $(R^2)$ was calculated to evaluate the goodness of fit. Furthermore, a pairwise t-test was conducted by calculating the goodness of fit for both linear and quadratic fit to observe if this trend is repeatable across all subjects by setting $\alpha$ to 0.05. V. RESULTS Across the different assistance conditions, general trend showed that there was a U-shaped trend where the global optima for the lowest metabolic cost were in between the 13% and 26% conditions (Fig. 7). For the average metabolic cost results, a quadratic and a linear fit had the $R^2$ value of 0.869 and $9.38 \times 10^{-5}$ respectively. A pairwise t-test result showed a Fig. 8. Average hip kinematics across subjects during different magnitude of assistance levels. Darker shades represent increase of assistance level. With higher assistance magnitude, peak hip flexion angle increased. statistical significance of the $R^2$ of a quadratic fit to linear fit (p < 0.01). Using the quadratic fit to the metabolic cost model, theoretical maximum reduction of metabolic cost was computed to be 6% reduction compared to zero impedance mode. The RMS torque that was applied to the user during the flexion and extension region linearly scaled with different assistance levels while maintaining relatively constant residual interaction torque and showed that similar amount of assistance were provided for both regions during the gait cycle (Table I). Additionally, a post hoc analysis was conducted to compute the continuous torque that was applied to the user over a gait cycle for each assistance levels (normalized by the walking speed) which was 3.59 Nm/m·s-1 for 13%, 7.55 Nm/m·s-1 for 26%, and 11.11 Nm/m·s-1 for 40% respectively. Using the computed torque values, normalized continuous torque over a gait cycle at 20% was found to be $5.62 \,\mathrm{Nm/m \cdot s^{-1}}$ , which is what the study supports as the optimal continuous torque level for exoskeleton design. Hip joint kinematics changed with varying levels of assistance magnitude (Fig. 8 and Table I). During the first part of stance phase (from 0% to 15% of the gait cycle), there was an excessive hip extension with the increase of assistance magnitude. The average RMS error of hip kinematic deviation in the early stance across subjects was 1.21° for 13%, 2.57° for 26%, and 3.96° for 40% compared to the zero impedance mode. For the remaining early stance (up to 30% of the gait cycle), the hip kinematics returned to the normal hip joint trajectory. In late stance (right before the push off), the hip kinematic started to deviate again where the increase of assistance resulted in higher hip flexion. Especially when transiting to the mid swing (around 80% of the gait cycle), the peak hip flexion angle increased significantly with higher assistance. The average peak hip flexion across subjects increased 4% for 13%, 10.7% for 26%, and 15% for 40% assistance compared to zero impedance mode where the peak flexion occurred at approximately 85% of the gait cycle. Moreover, the average hip joint range of motion across subjects increased 4.49% for 13%, 6.71% for 26%, and 18.93% for 40% assistance compared to zero impedance mode. The quadratic fit best represented the resulting hip range of motion with different assistance levels (R2 = 0.962). Additionally, the excessive hip flexion (Especially in the 40% assistance condition) caused the subject to walk in a marching gait pattern. This gait pattern increased the overall stride frequency linearly with the increase of assistance levels. The average stride frequency across subjects was 0.75 stride/sec for zero impedance, 0.78 stride/sec for 13%, 0.81 stride/sec for 26%, and 0.84 stride/sec for 40% assistance. The linear regression showed that the assistance level and stride frequency indeed have a linear
relationship (R2 = 0.999). The stride frequency in zero impedance mode correlated well with human biomechanics data with the same walking speed [28]. A video of a subject walking in a marching gait pattern with hip assistance at maximal levels is included in the supplemental material. VI. DISCUSSION Overall, our exoskeleton was able to show a positive result in achieving metabolic cost benefits with the increase of assistance level. Our study resulted in a similar metabolic cost reduction (6% when optimized) compared to other recent exoskeleton studies [21], [25], [39]. For example, Seo et al. utilized an adaptive oscillator controller that generated a torque profile similar to our biological torque controller and was able to achieve a substantial amount metabolic cost reduction (∼13%) compared to no exoskeleton condition using a lightweight hip exoskeleton [21]. Ding et al. was also able to achieve a considerable amount of metabolic cost reduction (between 5.7∼8.5%) with cable driven hip extension assistance compared to unpowered condition [25]. The range of metabolic cost benefit differences across the studies were mainly due to other confounding factors such as the device structure and controller architecture. While providing hip assistance resulted a metabolic cost reduction, our initial hypothesis was rejected in that increase of assistance level did not yield higher metabolic cost reduction. The resulting U-shaped trend of metabolic cost provides an important information that there is an optimum for exoskeleton assistance to attain the best metabolic cost benefits. Our study resulted in a similar quadratic trend as the relevant literature study using the ankle exoskeleton [40]. While the study may not directly correlate with our work (as the ankle joint has different musculotendon structure than the hip joint), the general results aligned that excessive assistance at the joint may penalize both mechanically and biologically. Similar to the literature studies on assistance timing study [26], [27], the assistance magnitude should be optimized for best energetic savings rather than just being set as high as physically possible. Moreover, while it was outside the scope of this study, a possible future work can observe the energetic effect of hip assistance when the magnitude ratio of extension to flexion varies to fully optimize the hip assistance levels. Mainly we observed that larger assistance magnitudes biased the hip kinematics in the flexion direction during swing phase (Fig. 8). This excessive hip flexion induced a marching gait which increased the user's stride frequencies. Generally, increased stride frequencies in marching gait will direct the user to take shorter steps. Considering the stance phase as an inverted pendulum motion, positive work at the center of mass is required to restore the energy lost during the collision occurring at the heel contact [41], [42]. As increase in the exoskeleton assistance decreases the user's step length, overall negative work done through the collision gets reduced hence, higher metabolic cost reduction. On the other hand, the increase of stride frequencies in higher assistance levels penalize the net metabolic cost. Along with stride frequencies, results showed that the hip joint range of motion exhibited quadratic growth with increase of assistance levels which will force the subject to have a faster leg swing. Literature studies have shown that increase in the stride frequencies correlates with faster swinging leg motion where muscle fibers are required to produce larger forces at short durations. Overall, this will result in a higher metabolic cost due to the low economy muscle generating the required force over short time [43], [44]. Two main factors relating to increase in assistance, stride frequencies and collision time, direct the metabolic cost in an opposite direction which explains the reason for achieving a general U-shaped trend for the resulting metabolic cost where ideally the cross over point of two factors is the optimal assistance magnitude that attains highest metabolic cost reduction. Our findings in regard to optimal assistance levels provide valuable information about the exoskeleton design. For example, computed continuous torque to achieve the best metabolic cost benefits can help future exoskeleton designers to optimize the actuator specification. Furthermore, the non-negligible energy loss through the user interface should be investigated more as it may have affected the result. The hip kinematic deviation occurring at the early stance phase occurred due to the soft tissue of the limb segment. This dampening effect in the limb limits the exoskeleton by delaying the assistance applied at the skeletal structure. Lastly, further exploration in the interaction between control parameters such as integrating the human-inthe-loop optimization may be useful [18], [24]. This approach can improve the exoskeleton controller by having the control parameter scale to the user's state such as the user's stride frequencies. Through this, the dynamic controller may be able to accommodate the assistance in a stride by stride basis. VII. CONCLUSION Our powered hip exoskeleton showed that the metabolic cost reduction does not correlate linearly with assistance magnitude. The underlying biomechanical effects in the user during exoskeleton assistance resulted in a U-shaped trend in the metabolic cost. The exhibited exoskeleton behavior illustrated the importance of understanding human robot interaction. Moreover, as our result showed that additional levels of support are undesirable, there are clear ramifications for system design
. Future exoskeleton designers can utilize our findings to further investigate to optimize the mechatronic design for a more robust and versatile exoskeleton. Lastly, implementation of an integrated controller capable of scaling parameters dynamically may aid the exoskeleton technology to be translated to more realistic settings such as outdoor environments. ACKNOWLEDGMENT The authors would like to thank Dr. G. Kogler for his insights in designing the orthotic interface, C. Kilpatrick and S. E. Lee for their help in fabricating the interface, and J. Li and R. Hong for data collection. The authors would also like to thank C. Bivens and M. Mayo at Georgia Tech Research Institute for their contributions.
RESEARCH ARTICLE Comparing optimized exoskeleton assistance of the hip, knee, and ankle in single and multi-joint configurations Patrick W. Franks\* , Gwendolyn M. Bryan, Russell M. Martin , Ricardo Reyes, Ava C. Lakmazaheri and Steven H. Collins Department of Mechanical Engineering, Stanford University, Stanford, California, USA *Author for correspondence: Patrick W. Franks, Department of Mechanical Engineering, Stanford University, Stanford, California, USA. Email: pwfranks23@gmail.com Received: 11 May 2021; Revised: 13 September 2021; Accepted: 20 October 2021 Key words: biomechanics; exoskeleton; human-in-the-loop optimization Abstract Exoskeletons that assist the hip, knee, and ankle joints have begun to improve human mobility, particularly by reducing the metabolic cost of walking. However, direct comparisons of optimal assistance of these joints, or their combinations, have not yet been possible. Assisting multiple joints may be more beneficial than the sum of individual effects, because muscles often span multiple joints, or less effective, because single-joint assistance can indirectly aid other joints. In this study, we used a hip–knee–ankle exoskeleton emulator paired with human-in-the-loop optimization to find single-joint, two-joint, and whole-leg assistance that maximally reduced the metabolic cost of walking. Hip-only and ankle-only assistance reduced the metabolic cost of walking by 26 and 30% relative to walking in the device unassisted, confirming that both joints are good targets for assistance (N = 3). Knee-only assistance reduced the metabolic cost of walking by 13%, demonstrating that effective knee assistance is possible (N = 3). Two-joint assistance reduced the metabolic cost of walking by between 33 and 42%, with the largest improvements coming from hip-ankle assistance (N = 3). Assisting all three joints reduced the metabolic cost of walking by 50%, showing that at least half of the metabolic energy expended during walking can be saved through exoskeleton assistance (N = 4). Changes in kinematics and muscle activity indicate that single-joint assistance indirectly assisted muscles at other joints, such that the improvement from whole-leg assistance was smaller than the sum of its single-joint parts. Exoskeletons can assist the entire limb for maximum effect, but a single well-chosen joint can be more efficient when considering additional factors such as weight and cost. Impact Statement Exoskeletons could make walking easier for people, from military personnel to older adults. They can reduce the energetic cost of walking, but we still do not know the best way to assist walking. Which leg joints should exoskeletons assist? What torques should they apply? What is the greatest improvement we could expect? To study this, we optimized hip–knee–ankle exoskeleton assistance for each joint individually, for two-joint combinations, and for the whole leg. We found that assisting the whole leg reduced the energy cost of walking by 50%, double the state-of-the-art. We also found that while assisting the whole-leg was most effective, assisting a single joint may be more efficient when considering device mass. These findings will help exoskeleton designers choose which joints to assist. © The Author(s), 2021. Published by Cambridge University Press. This is an Open Access article, distributed under the terms of the Creative Commons Attribution licence (http://creativecommons.org/licenses/by/4.0), which permits unrestricted re-use, distribution and reproduction, provided the original article is properly cited. Introduction Lower limb exoskeletons can assist human locomotion by reducing the metabolic cost of walking. These devices have the potential to restore ambulatory ability lost from age or disability, and increase maximum performance for high-activity users like first responders, military personnel, or athletes. One important way exoskeletons can help is by reducing the metabolic cost of walking, which is the amount of biochemical energy consumed to produce walking at a given speed (Das Gupta et al., 2019). Humans tend to move in ways that minimize metabolic cost (Zarrugh et al., 1974; Bertram and Ruina, 2001; Donelan et al., 2001; Sanchez et al., 2020), indicating its importance. Reducing the metabolic cost of walking is considered the gold standard for evaluating performance-augmenting exoskeletons (Young and Ferris, 2017; Sawicki et al., 2020). By reducing metabolic cost, exoskeletons could help achieve related mobility outcomes like increasing the user's walking speed or decreasing fatigue. Exoskeletons have improved walking by reducing metabolic cost, but larger improvements may be necessary for widespread adoption of exoskeleton products. Reductions in metabolic cost have been demonstrated from assisting one or two joints with both tethered and mobile exoskeletons (Malcolm et al., 2013; Mooney
et al., 2014; Collins et al., 2015; Seo et al., 2016; Quinlivan et al., 2017; Zhang et al., 2017; Ding et al., 2018; Lee et al., 2018; Malcolm et al., 2018; Lim et al., 2019; MacLean and Ferris, 2019; Cao et al., 2020; Sawicki et al., 2020). The largest metabolic cost reductions have been around 18% relative to walking in no exoskeleton using hip-only assistance on a mobile device (Ding et al., 2018; Lim et al., 2019) and 24% relative to walking in an exoskeleton with no torques applied using tethered bilateral ankle assistance (Zhang et al., 2017). Despite demonstrated improvements, exoskeletons are in a nascent stage of commercial development, and widespread adoption has not yet occurred (Young and Ferris, 2017). To promote adoption, devices may need larger improvements that offset the negative impacts of exoskeletons, such as worn mass, bulkiness, or cost. The added mass of wearing a device, which can vary with device design, imposes a metabolic penalty, which could eliminate small benefits. Users may not be able to sense if the exoskeleton is assisting them because the largest improvements in the field are similar to the just-noticeable difference for metabolic cost (around 20%, Medrano et al., 2020). Improving our scientific and engineering understanding of exoskeleton assistance could deliver larger benefits that may lead to widespread adoption of these devices. Whole-leg exoskeleton assistance could produce the largest improvements to walking performance. Assisting all of the lower-limb joints simultaneously seems likely to yield the largest energy savings because the hips, knees and ankles all significantly contribute to biological energy consumption during walking (Winter, 1991; Farris and Sawicki, 2012). Simulations of exoskeleton assistance also indicate that whole-leg devices should be most effective (Uchida et al., 2016; Dembia et al., 2017; Franks et al., 2020). Unfortunately, there has been limited testing of whole-leg assistance for able-bodied users (Zoss et al., 2006), and these devices have not yet reduced metabolic cost (Gregorczyk et al., 2010). With improvements to exoskeleton hardware and control, larger benefits might be realizable. While whole-body assistance may produce larger benefits, single-joint assistance may be more efficient. Assisting just one joint could lead to a smaller, lighter and more cost-effective device, which could result in better net improvements. While a variety of single-joint devices have been tested, there has not been a well-controlled comparison between them owing to differences in actuator capabilities and control. Each device has had different limits on torque and power (Bryan et al., 2020), many below previously identified optimal values (Zhang et al., 2017), and larger torque and power capacity are associated with larger reductions in metabolic cost (Quinlivan et al., 2017; Ding et al., 2018). Most devices have not used control that has been systematically optimized for the participant, which can improve performance by as much as a factor of five (Jackson and Collins, 2015; Zhang et al., 2017) and has led to the largest improvements in the metabolic cost of walking (Zhang et al., 2017; Ding et al., 2018). A direct comparison of optimized single-joint assistance in a high-torque, high-power exoskeleton would be useful to designers as they choose which joint, and how many joints, to assist. A comparison of single-joint and multi-joint assistance would also provide scientific insights into the biomechanics of walking. A large portion of human leg musculature is biarticular (van Ingen Schenau et al., 1987), comprising muscles that span two joints. Biarticular muscles might be more effectively assisted by multi-joint exoskeletons, leading to a total benefit beyond the sum of assisting each joint individually. Some experiments have suggested that multi-joint assistance might be more effective than single-joint assistance (Ding et al., 2017; Malcolm et al., 2018). Alternatively, adding assistance at other joints may have diminishing returns. Users might adapt their walking pattern to maximize the benefit from assistance at a single-joint, thereby indirectly benefiting muscles at other joints, which may make it relatively less effective to assist additional joints. Some experiments have shown indications of such indirect assistance (Lenzi et al., 2013; Mooney and Herr, 2016; Jackson et al., 2017). A well-controlled comparison allowing observations of how users respond to different types of assistance would help us to develop improved models of biomechanical and neural adaptation to exoskeletons. The purpose of
this study was to find the single-joint, two-joint, and whole-leg exoskeleton torques that minimized the metabolic cost of walking and to understand how effective each device architecture is at assisting walking. We used a tethered hip–knee–ankle exoskeleton emulator that can assist hip flexion and extension, knee flexion and extension, and ankle plantarflexion of both legs (Bryan et al., 2020). This labbound device uses off-board motors and power to apply large torques to the user while minimizing worn mass. This allows for comparison of different actuation schemes without the difficulty of designing product-like mobile devices, which enables us to identify optimal actuation parameters to inform design specifications of future untethered exoskeletons. We paired this tool with human-in-the-loop optimization, a process where the control of the exoskeleton is updated in real time based on biomechanical measurements of the user (Felt et al., 2015; Koller et al., 2016; Kim et al., 2017; Zhang et al., 2017). We performed experiments optimizing and comparing assistance at the hips, knees, and ankles, individually, in pairs, and with all three joints assisted simultaneously. We optimized each assistance pattern to reduce the measured metabolic cost of walking and compared it to walking without assistance and walking without the exoskeleton. We measured changes in kinematics and muscle activity to see how users adapted to assistance and to gain insights into the potential biomechanical mechanisms that brought about reductions in metabolic cost. By finding and comparing optimized assistance for different potential device architectures, we expect these results to inform models of human adaptation to exoskeletons and lead to the design of more effective exoskeletons. Methods Participants Four healthy participants were included in this study (P1: M, 26 years old, 90 kg, 187 cm; P2: F, 26 years old, 61 kg, 170 cm; P3: M, 19 years old, 82 kg, 176 cm; P4: M, 23 years old, 62 kg, 171 cm). We were limited to four participants because of the extensive time required to complete the protocol (Supplementary Material, Sections 11 and 12). Each participant completed at least 50 hr of experiments in total. These participants were also authors of the study (P.W.F., G.M.B., R.M.M., and R.R.) as these were the people who could spend such time as a participant for the study. All four participants completed the whole-leg optimization, but due to external factors related to the COVID-19 pandemic, three participants completed one-joint (P1, P2, and P3) and two-joint (P1, P3, and P4) optimization. With three participants, we have a statistical power of 0.8 to detect metabolic reductions greater than 24%, assuming metabolic reductions have a standard deviation of 7.4% (Zhang et al., 2017; Supplementary Material, Section 12). With four participants, the 50% reduction detected from whole-leg assistance has a statistical power of 0.999. All four participants were experienced with the device at the time of optimization. P1 and P2 had previous experience walking in the exoskeleton before this experiment. P3 and P4 completed a training protocol prior to optimization to get accustomed to wearing the exoskeleton and walking with torques. More details on the training protocol are included in Supplementary Material, Section 1. Experimental Protocol We optimized single-joint, two-joint, and whole-leg assistance using a hip–knee–ankle exoskeleton (Figure 1; Bryan et al., 2020). We used human-in-the-loop optimization (Zhang et al., 2017), a strategy where the control of the exoskeleton is updated in real-time based on measurements of the user. For this study, the cost function to be minimized was the measured metabolic cost of walking at 1.25 m/s. First, we optimized single-joint assistance for P1, P2, and P3 in the order of ankle-only, hip-only, and then kneeonly assistance. We then optimized whole-leg assistance for all four users, meaning we optimized assistance of the hip, knee and ankle simultaneously. Finally, for P1, P3, and P4, we optimized two-joint assistance, in the order of hip–ankle, knee–ankle, and hip–knee assistance. After optimization, we performed validation experiments to compare the optimized assistance to the control conditions of walking without the exoskeleton and walking in the exoskeleton with no torque applied. During these validations, we measured metabolic cost, applied torques, kinematics, and muscle activity. Exoskeleton Hardware Assistance was applied using a hip–knee–ankle exoskeleton emulator (Figure 1; Bryan et al
., 2020). This device can apply large torques using offboard motors and Bowden cable transmissions to actuate an end effector worn by the user, enabling laboratory tests of different assistance strategies without actuation limits (Caputo and Collins, 2014). The device has a worn mass of 13.5 kg. It has carbon fiber struts along the length of the legs that are designed to minimize restriction of the user by being stiff in actuated directions but compliant in out-of-plane bending. The exoskeleton was fit to each user by adding boots for their foot size, by adjusting the length of the shank, thigh, and torso segments of the exoskeleton, and by adjusting the width of the exoskeleton at the knees, thighs, and hips. Straps were adjusted to fit the user at the shanks, thighs, hips, and torso. Exoskeleton Control The exoskeleton is controlled by commanding a desired torque for each joint (Bryan et al., 2020). When the desired torque is zero, the exoskeleton tracks the user's joint angles and applies no torques. During walking, we define these desired torque profiles as a function of percent stride. We consider heel strike, measured by ground reaction forces on the treadmill, to be the start of a stride. We calculate percent stride Figure 1. Overview of exoskeleton emulator system. (Left) Overview of exoskeleton emulator. Ten powerful off-board motors actuate a lightweight end effector worn by a user who walks on a treadmill. Metabolic cost is measured using a respirometry system and muscle activity is measured using electromyography (EMG). (Center) Isometric photo of experimental setup. (Right) Side view of exoskeleton. The exoskeleton can apply torques in hip flexion and extension, knee flexion and extension, and ankle plantarflexion. as the time since heel strike divided by the average stride time over the past 20 strides. The hip profile starts at 84% of stride after heel strike because hip extension torque is active during heel strike. Resetting the hips' stride time at heel strike caused discrete jumps in hip extension torque during pilot testing. The desired torque profile for each joint is made up of a spline (piecewise cubic hermite interpolating polynomial) anchored by nodes. Each node can be set in advance by an operator, or it can be updated in real time by an algorithm. For the knees, torque was also commanded as a function of joint state. Along with a torque-time profile, knee torque had one spring-like phase during stance, and one damping-like phase during late swing. Our exoskeleton accurately applied desired torques using closed-loop proportional control with iterative learning and joint-velocity compensation (Zhang et al., 2015; Bryan et al., 2020). Root-meansquare (RMS) error for tracking desired torques was 0.6 Nm at the hips, 3.0 Nm at the knees, and 0.4 Nm at the ankles during whole-leg assistance (Supplementary Material, Section 13). Error was highest at the knees because the state-based control allowed for discontinuous jumps in desired torque that were not possible for our device to track, and the desired torque would change step to step making it harder to track. When zero torque was commanded it was realized effectively, with an RMS applied torque of less than 1 Nm. Controller Parameterization The optimization algorithm varied parameters that affected the desired torque control of the exoskeleton (Figure 2). These parameters are mostly related to the timing and torque magnitude of the nodes that define our splines. We chose these parameters by considering previously successful human-in-the-loop optimizations (Zhang et al., 2017; Ding et al., 2018), considering biological torques during walking (Winter, 1991), and by pilot Figure 2. Desired torque profiles defined by the controller for the hips, knees, and ankles. Hip-only assistance was defined by 8 parameters, knee-only assistance was defined by 10 parameters, and ankleonly assistance was defined by 4 parameters, meaning whole-leg assistance was optimized using 22 parameters. For the hips (left) and ankles (right), torque (black) was commanded as a function of time, defined as a spline fit to nodes (red) that were optimized during the experiment. For the knees (center), torque was commanded both as a function of time (black), joint angle, and joint velocity. During stance, the knee torque was a function of knee angle to mimic a spring (red), where the spring's stiffness was optimized. During late swing, torque was a function of knee joint velocity to mimic a damper (red). The red curves shown for these periods of state-based control are the average applied torque at the knees from whole-leg assistance, but the applied torque could vary based on the user's kinematics. The steep increases and decreases in applied torque during knee extension assistance were due to the
impedance controller being turned on and off as a function of percent stride defined by the nodes while the knee angle was nonzero, resulting in discrete jumps in desired torque. testing. Before this, one participant completed a 9-parameter whole-leg optimization pilot study (Supplementary Material, Section 14), which indicated the need for more degrees of freedom in our controller. For these optimizations, we included 8 parameters for the hips, 10 for the knees, and 4 for the ankles (Figure 2). For optimization of whole-leg assistance, the optimizer could adjust all 22 parameters. Each parameter had a minimum and maximum allowed value. The allowed parameter ranges were based on user testing to ensure all tested profiles would be sufficiently comfortable for the user to walk in. The optimization tended not to optimize to these limits. However, ankle torque was often as large and as late in stride as possible, which meant the fall time was minimized to prevent torque application during swing. During single-joint assistance, ankle torque magnitude was limited to 1 Nm/kg. After single-joint optimization concluded and multi-joint optimization began, participants noted that these torques were too large to walk in comfortably. Participants noted that the ankle assistance felt like it was extending the ankle too quickly, and it is possible that kinematic adaptations during ankle-only assistance that mitigated this discomfort were not possible when assistance was also present at the hip and knee. To prevent discomfort, the torque magnitude limit was reduced to 0.8 Nm/kg for ankle torque during multi-joint assistance. Tables with the parameter ranges, as well as their initial and optimized values, are available in Supplementary Material, Section 5. The hip profile was defined by eight parameters. Hip extension was defined by the rise time, peak time and peak magnitude while hip flexion was defined by the peak time, peak magnitude and fall time. There was a period of no torque in between the two peaks defined by the midpoint timing and its duration. The period of no torque dictated the hip extension fall time and the hip flexion rise time. Knee torque was defined by 10 parameters. Knee torque was commanded both as a function of percent stride and of joint state. The first phase of knee torque was knee extension defined as a virtual spring, with torque proportional to knee angle, which was zero when the knee was straight. Knee extension was defined by the virtual spring onset timing, stiffness, and offset timing. If the joint angle reached zero degrees before the offset time, the knee torque would stay at zero torque for the remainder of the stiffness period. During knee flexion around toe-off, torque was defined as a function of time similar to the hips. This torque was determined by the rise time, peak torque magnitude, peak time, and fall time. Late in swing, knee flexion torque was commanded as a virtual damper, so torque was proportional to a filtered measurement of knee joint velocity. The damping period was defined in a similar way to knee stiffness, with optimization of the onset timing, the damping coefficient, and the offset timing. As the knee joint angle and velocity were not necessarily zero at the start of the state-based controllers, desired torque could instantaneously change at the onset. Ankle torque was defined using four parameters, which were previously effective for optimization of ankle assistance (Zhang et al., 2017). Torque was defined by rise time, peak torque magnitude, peak time, and fall time. To ensure large torques were not applied too late in the stride, torque was set to be zero by 65% of stride, so if peak time optimized to its latest allowed value (55% of stride), the fall time would be set to the minimum allowed fall time (10% of stride). Human-in-the-Loop Optimization Protocol To optimize assistance, we used the covariance matrix adaptation evolutionary strategy (CMA-ES) (Hansen, 2006), which has previously been effective for human-in-the-loop optimization of exoskeletons (Zhang et al., 2017; Witte et al., 2020). CMA-ES samples a "generation" of conditions from a distribution defined by parameter means and a covariance matrix, ranks the performance of the samples, and uses those results to update the mean and covariance before sampling the next generation. The optimizer's goal was to minimize metabolic cost, which was estimated for each condition after 2 min of walking using a first-order dynamical model (Selinger and Donelan, 2014), similar to previous work (Zhang et al., 2017; Witte et al., 2020). More details about the optimization, including hyperparameters and numbers of conditions per generation, are included in Supplementary Material, Section 15. The initial parameter values for each optimization were carefully selected to try to reduce convergence time. For the single-joint optimizations for P1, initial parameter values were based on previously optimized assistance (Zhang et al., 2017; Ding et al., 2018), hand-tuning, and a 9-
parameter pilot study (Supplementary Material, Section 14). For whole-leg optimization for P1, initial values were based on the optimized values for single-joint assistance. For P2, P3, and P4, initial values for optimization were based on the optimized values for P1. Finally, for the two-joint assistance optimizations, initial parameters were based on each participant's previously optimized whole-leg assistance values. The initial values for all parameters are included in Supplementary Material, Section 5. The optimization time was intended to balance being long enough to ensure convergence while short enough to be experimentally feasible. P1 underwent a longer optimization to ensure convergence, to estimate expected reductions, and to inform our understanding of how the optimizer would perform (Supplementary Material, Sections 11 and 16). Ankle-only optimization was conducted for 12 generations over 3 days, and hip-only, knee-only and whole-leg were conducted for at least nine generations over 3 days. Each two-joint assistance optimization was conducted for six generations over 2 days. For all participants, single-joint and whole-leg optimization each occurred over at least 3 days, which seemed sufficiently long for P1 to reach metabolic reductions that were consistent in future sessions and that matched previous studies for previously assisted joints (Zhang et al., 2017) (Supplementary Material, Section 11). For whole-leg assistance for P3, the optimization was restarted because the user had an abnormally high stride frequency that had high metabolic cost, indicating a maladaptation to assistance similar to some users in a previous optimization study. The exact number of generations and days for each optimization is included in Supplementary Material, Section 11. Participants were permitted but not required to take breaks between generations. While walking, participants were allowed to listen to podcasts using wireless headphones. Validation Protocol We conducted validation experiments to evaluate the effectiveness of the optimized assistance. Metabolic reductions were validated for each assistance strategy after each optimization. After all the single-joint and whole-leg optimizations were completed, torques, kinematics, and muscle activity were compared between assistance strategies on the same day. Finally, the two-joint assistance strategies were validated after each optimization. After each optimization was completed, a validation experiment was used to accurately assess the metabolic cost of walking and calculate the percent reduction. This collection was on a separate day, before optimization of the next assistance strategy began. Users walked in longer bouts for the exoskeleton conditions to ensure accurate measurements of steady-state metabolics and to ensure users were acclimated to the device and assistance. We recorded the user standing quietly for 6 min, walking without the exoskeleton for 6 min, walking in the exoskeleton with no torque applied for 10 min, and walking with the optimized assistance torques for 20 min, in a double-reversed order (ABCDDCBA). This order was not randomized due to the time it takes to get in and out of the exoskeleton, as well as to maximize acclimation to the device by presenting progressively more novel conditions. Users rested for at least 3 min between walking conditions, and at least 5 min before a quiet standing condition, to ensure their metabolics returned to baseline. For the no exoskeleton condition, users wore the same brand and model of boots that are included in the exoskeleton (McRae 8189). After all single-joint and whole-leg optimizations were completed, we evaluated all these optimized strategies in one data collection to directly compare conditions. We measured applied torque, kinematics, muscle activity, power, and vertical ground reaction forces. For this validation, we recorded the user standing quietly for 6 min, walking without the exoskeleton for 6 min, walking in the exoskeleton with no torque applied for 10 min, and then walking for 10 min in each of the four optimized assistance conditions (hip-only, knee-only, ankle-only, and whole-leg) in a random order, and then the no torque, no exoskeleton, and quiet standing conditions a second time (ABCDEFGCBA). For the three users who completed optimization of two-joint assistance, a validation day was completed after each optimization similar to the protocol following the single-joint and whole-leg optimizations. We recorded the user standing quietly for 6 min, walking without the exoskeleton for 6 min, walking in the exoskeleton with no torque applied for 10 min, and walking with the optimized assistance torques for 20 min, in a double-reversed order (ABCDDCBA), so each condition was evaluated twice. For these validations, we measured metabolics, torques, kinematics, muscle activity, power, and vertical ground reaction forces. Measured Outcomes We collected biomechanical data of the user when walking in the different conditions during the validation. We calculated the average of these measurements over the last 3–5 min of walking of each condition to ensure the user's metabolics and gait had reached steady-state. Metabolic cost Metabolic cost
was calculated using indirect calorimetry. We measured volumetric carbon dioxide expulsion, oxygen consumption, and breath duration on a breath-by-breath basis (Quark CPET, COSMED). For each condition, we calculated metabolic rate using a modified Brockway equation (Brockway, 1987) similar to previous studies (Zhang et al., 2017; Witte et al., 2020). Average metabolic cost was calculated for each condition using the last 3 min for quiet standing and walking with no exoskeleton, and using the last 5 min for walking with no torque and walking with optimized assistance. Because each condition was evaluated twice in the double-reversed order, the average was calculated across both measurements of the condition. The cost of quiet standing was subtracted from the measured cost of all the other conditions to calculate the cost of walking. Users fasted for at least 2 hr before optimizations and at least 4 hr before validations to minimize the possible thermal effects of food on metabolic cost measurements. Two-tailed paired t-tests were used to evaluate if the metabolic cost of walking with exoskeleton assistance was significantly different from walking in the control conditions. For all two-joint optimizations except P3 hip-ankle, participants wore a cloth mask underneath the metabolics mask to comply with COVID-19 safety protocols. This mask did affect the metabolics measurements, by seeming to cause an underestimation of measured metabolic cost (Supplementary Material, Section 17). While this disrupted the accuracy of the absolute measurements, we expected the percent reductions in metabolic cost to be accurate, because we are comparing between conditions. We excluded these affected measures when calculating the average absolute measurements reported for standing quietly, walking with no exoskeleton, and walking with no torque. Torques and kinematics Applied torques were measured using load cells and strain gauges on the exoskeleton. Exoskeleton joint angles were recorded to estimate user kinematics, meaning that we could not calculate kinematics for walking without the exoskeleton. Stride frequency was calculated using vertical ground reaction forces measured by the instrumented treadmill (Bertec). Measurements were averaged over the last 3 min of the condition for walking with no exoskeleton, and averaged over the last 5 min for walking with no torque and walking with each assistance condition. Measurements were averaged across both legs. For conditions that were evaluated twice (walking with no exoskeleton and walking with no torque), results were averaged across the two conditions. Biological torques reported for reference (Figure 4) were from a separate study. Reference walking data is from three able-bodied male subjects (Arnold et al., 2013), different from the participants included in this study. Walking data was collected from three gait cycles of motion capture data during treadmill walking at 1.25 m/s including marker trajectories, ground reaction forces, and EMG measurements (Arnold et al., 2013). Biological joint torques during walking were calculated using the Inverse Dynamics tool in OpenSim (Seth et al., 2018), as described in Franks et al. (2020). Muscle activity Muscle activity was measured using surface EMG (Delsys Trigno). We applied a third-order bandpass filter of 40–450 Hz, rectified, then applied a third-order low pass filter of 10 Hz (De Luca et al., 2010). Muscle activity was averaged for a stride over the last 5 min of the device conditions and over the last 3 min for the no device conditions. We subtracted the baseline noise offset then normalized to the maximum of the no torque condition profile. The sensor locations are similar to the protocol of previous gait analysis experiments (Winter, 1991) with adjustments to avoid interfering with the device structure and straps. Results Metabolic Cost The metabolic cost of walking was reduced in all assistance conditions (Figure 3). To evaluate the metabolic cost of walking, we subtracted the metabolic cost of standing quietly (1.52 W/kg on average, Supplementary Material, Section 2) from the metabolic cost measured in each walking trial. The metabolic cost of walking without wearing the exoskeleton was 2.86 W/kg on average. The cost of walking in the exoskeleton with no torque applied was 3.92 W/kg on average, which was higher than the no-exoskeleton condition because of the added mass and impedance to the user. Percent reductions in metabolic cost were calculated in comparison to this no-torque condition to assess the effect of the designed torque assistance specifically without the effect of device mass, allowing for best comparisons between assistance conditions, which could require different device architectures as mobile systems. Single-joint assistance at each joint reduced metabolic cost, with the largest improvements coming from the ankles and hips. Hip-only assistance reduced the metabolic cost of walking by 26% relative to walking with no torque (N = 3, range of reductions: 24–30%, p = .005). Knee-only
assistance reduced metabolic cost for each participant, with an average reduction of 13% relative to walking with no torque, although this was not statistically significant (N = 3, range of reductions: 5–18%, p = .07). Ankle-only assistance performed best of the single-joint strategies, reducing metabolic cost by 30% relative to walking with no torque (N = 3, range of reductions: 28–31%, p = .004). When assisting a single joint, exoskeleton designers should consider the ankles or the hips. Two-joint assistance outperformed single-joint assistance. Two-joint assistance reduced metabolic cost of walking relative to no torque by 33% for hip–knee (N = 3, range of reductions 29–37%, p = .008), 37% for knee–ankle (N = 3, range of reductions 35–40%, p = .02), and 42% for hip–ankle assistance Figure 3. Metabolic cost of walking. Average metabolic cost (bar) of each condition reported as a percentage of walking in the exoskeleton with no torque. Individual participant values are shown with symbols (P1 X, P2 O, P3 Δ, and P4 þ). Metabolic cost of walking was calculated by subtracting out quiet standing. The percent reduction relative to walking with no torque is shown above each bar. For each participant, the cost of walking without the exoskeleton (No exo., gray) was averaged over all validations. Whole-leg assistance (blue, N = 4) provided the largest improvement to metabolic cost of walking, reducing it by 50% relative to walking in the exoskeleton without assistance. (N = 3, range of reductions: 36–49%, p = .03) (Supplementary Material, Section 3). Hip–ankle assistance provided the most benefit, mirroring single-joint reductions and aligning with expectations based on biological power. Whole-leg exoskeleton assistance led to the greatest reductions in metabolic cost of any condition. Whole-leg assistance (hips, knees, and ankles simultaneously) reduced the metabolic cost of walking by 50% relative to walking with no torque (N =4, range of reductions: 46–53%, p= .003), corresponding to a reduction of 33% relative to walking without wearing the exoskeleton (N = 4, range of reductions: 17– 41%, p = .016). This shows that about half of the metabolic energy expended during walking can be saved through exoskeleton assistance, and suggests that whole-leg assistance could provide large net benefits in untethered systems, even after accounting for the effects of added mass. Optimized Exoskeleton Torque Optimized torques differed from biological torques in both timing and magnitude (Figure 4, Supplementary Material, Sections 4–6). Optimized torque magnitudes were smaller than biological torques, with peak exoskeleton torques ranging from about 15 to 60% of biological peaks. Ankle torque magnitudes were largest and optimized to the comfort-limited parameter constraints in all but two cases. The timing of the optimized assistance only partially aligned with biological torque. For example, the peak of optimal hip flexion assistance occurred later than peak biological flexion torque. Sometimes, assistance torque opposed typical biological torques. For example, knee flexion assistance around 60% of stride opposed biological knee extension torque for normal walking. These optimized magnitudes indicate the design requirements for mobile devices and show that optimized assistance is not a scaled version of biological torques. The shape and timing of optimized assistance was consistent across conditions and participants, but optimal magnitudes differed. For example, the optimal timing of peak hip extension assistance was about 11% of stride for all joint combinations and participants, while optimal magnitudes ranged from 0.24 Nm/ kg (hip-only) to 0.5 Nm/kg (hip–knee). One exception to the consistency of optimal timing was ankle torque rise time, which was shorter during single-joint assistance, possibly due to adjustments for comfort Figure 4. Optimized exoskeleton torques. Optimized single-joint (green, purple, and pink), two-joint (orange, yellow, and red) and whole-leg (blue) exoskeleton assistance torques at the hips (left), knees (center), and ankles (right). Lines are the average of the measured applied torque profiles across both legs and all participants (N =3 for single-joint and two-joint, N =4 for whole-leg), with the range of optimized profiles shown with their respective clouds for each type of assistance. Biological joint torques for unassisted walking without an exoskeleton (black) are included from a different study with different participants (Arnold et al., 2013; Franks et al., 2020) for reference; gray clouds indicate standard deviation of biological torques. For the hips and
knees, whole-leg assistance optimized to smaller magnitudes than single-joint assistance. For the ankles, maximum torque had to be constrained to find comfortable profiles for walking. Ankle torques were limited to 1 Nm/kg for single-joint assistance, and 0.8 Nm/kg for two-joint and whole-leg assistance. in multi-joint conditions (Supplementary Material, Section 5). Optimized assistance torques were typically larger when acting alone at a joint, and smaller when acting in a multi-joint configuration. For example, for P1, applied knee flexion torque peaked at 0.25 Nm/kg for knee-only assistance and at 0.14 Nm/kg for whole-leg assistance. The consistency of optimal timing parameters suggests that optimization could occur in a lower-dimensional parameter space of torque magnitudes, and that a generalized assistance profile could be almost as effective as a customized one. Kinematics Kinematics varied between assistance conditions, indicating that the user's walking pattern is not fixed and adapts to best utilize assistance (Figure 5, Supplementary Material, Sections 7 and 8). These changes were beyond the deviation measured during walking with no torque (gray cloud, Figure 5). In some cases, assistance shifted joint angles in the direction of the applied torque. For example, peak ankle plantarflexion angle increased with whole-leg assistance, and increased even more during ankle-only assistance, which had larger ankle torques. However, some kinematic changes were not the direct result of applied torques. For example, the indirect effects of hip-only and ankle-only assistance on the knee during stance were larger than the direct effect of knee assistance. These kinematic adaptations indicate the user adjusts their walking strategy to maximize the benefit they get from the exoskeleton, and that these adaptations do not always match intuition. Muscle Activity Muscle activity decreased with assistance, but it was not completely eliminated (Figure 6; Supplementary Material, Section 9). Typically, reductions in activity were seen in muscles that crossed assisted joints and acted in the same direction as assistance. For example, soleus activity decreased during all conditions that applied ankle assistance. Sometimes activity increased during opposing assistance, such as in the vastus lateralis during periods of knee flexion torque. Some reductions in muscle activity occurred during assistance at other joints. For example, gluteus maximus activity decreased when the hip was assisted directly, but also decreased during ankle-only assistance. This is consistent with the observation that exoskeleton assistance at one joint can indirectly assist muscles that cross other joints. The indirect assistance could be from the complex dynamics of the leg during walking, or could be facilitated by the Figure 5. Average joint kinematics. Average joint angle as a percentage of stride at the hips (left), knees (center), and ankles (right) for each assistance condition (denoted by color). Shown here are the average for both legs across all participants (N = 3 for single-joint and two-joint, N = 4 for whole-leg). All singlejoint and whole-leg conditions for P1, P2, and P3 were tested on the same day to reduce changes in alignment between user and device. Two-joint and P4's three-joint conditions were each collected individually. For walking in the exoskeleton with no torque (black), the standard deviation of angles is shown (gray cloud) to contextualize the magnitude of changes between conditions. Figure 6. Muscle activity. Muscle activity measured during walking using surface EMG for each condition. Lines shown are the average across all participants (N = 3 for single-joint, hip–knee, and hip– ankle, N = 2 for knee–ankle due to a technical difficulty with P3's EMG collection, and N = 4 for wholeleg). The EMG signal was filtered, averaged, had baseline activity removed to eliminate noise, and normalized to the peak value of walking in the exoskeleton without assistance (black). Gluteus maximus activity (second row, third column) decreased for hip-only, hip–knee, hip–ankle, and whole-leg assistance as expected, and also decreased during ankle-only and knee–ankle assistance, indicating that the gluteus was indirectly assisted by ankle exoskeleton torque. This effect was less pronounced for the soleus (top row, first column), where hip-only and hip–knee assistance only slightly reduced muscle activity. kinematic adaptations to maximize the effectiveness of each type of assistance. Optimized assistance did not cause users to eliminate muscle activity, suggesting that either some amount of activity is still useful or further advancements to control architecture would be needed to reduce energy expenditure further. Discussion With capable devices, optimization, and training, exoskeletons can provide very large improvements in locomotor performance. Whole-leg assistance reduced the metabolic cost of walking by 50% relative to walking with no torque, a substantial improvement over the state-of-the-art (17–24%) (Zhang et al., 201
7; Ding et al., 2018; Lim et al., 2019). This corresponded to a 33% reduction relative to walking with no exoskeleton, much greater than the just-noticeable difference in metabolic cost (20%) (Medrano et al., 2020), indicating that participants could feel the reduction in effort compared to walking normally. Because whole-leg assistance produced the largest benefit of all assistance conditions, exoskeleton designers who want to maximize performance should consider assisting the whole leg. Among single-joint assistance strategies, ankle-only assistance was most effective, followed closely by the hip, with smaller reductions possible at the knee. Hip and knee assistance resulted in greater metabolic cost reductions than previous lower-torque exoskeletons assisting these joints (Ding et al., 2018; Lim et al., 2019; MacLean and Ferris, 2019), while ankle assistance resulted in similar improvements as found with high-torque exoskeletons and human-in-the-loop optimization (Zhang et al., 2017; Jackson and Collins, 2019). Knee-only assistance reduced the metabolic cost of walking, although the reduction was not statistically significant for our sample size and significance level. Knee-only assistance may be more effective for walking up inclines (MacLean and Ferris, 2019; Haufe et al., 2020), considering the increased positive power requirements from the knee (Montgomery and Grabowski, 2018). Devices designed to assist just one joint during level walking should target the ankle or hip, which showed reductions of 30 and 26%, respectively. With similar metabolic reductions, designers could compare between the ankle or hip based on other aspects of the design. For example, it may be easier to interface with the ankle using a simple device, while a hip exoskeleton places the mass more proximally on the body where it is easier to carry (Browning et al., 2007). The best-performing two-joint assistance strategy, hip-ankle assistance, reduced metabolic cost by 42%, nearly as much as whole-leg assistance. With 37% and 33% reductions for knee–ankle and hip–knee assistance, the small added benefit of knee-assistance may not be worth the added device complexity compared to ankle-only or hip-only assistance for level ground walking. The inclusion of knee assistance may be more effective for different walking conditions such as incline walking or during sit-to-stand, where the knee is expected to contribute more to movement. Assisting multiple joints results in larger net benefits, but smaller benefits per joint, possibly because of the way people adapt to exoskeleton assistance. Whole-leg assistance led to the largest metabolic cost reduction (50%), but it was smaller than the sum of the reductions of the single-joint assistance strategies (26% þ 13% þ 30% = 69%). During single-joint assistance, users may have been able to adapt the kinematics of their unassisted joints to take most advantage of the single-joint assistance. During wholeleg assistance, their subconscious walking strategy could have been to maximize metabolic benefit overall from assistance of the whole-leg, while not getting maximal benefit from any joint in particular. This adaptation might also explain why optimized torque magnitudes during single-joint assistance were larger than whole-leg torque magnitudes. It was previously hypothesized that multi-joint assistance may be more effective at assisting biarticular muscles, but the relative benefit of single-joint assistance indicates that indirectly assisting muscles at other joints from single-joint assistance leads to larger benefits on a perassisted-joint basis. The reduction in metabolic cost per joint assisted (Table 1) could be helpful for designers when considering potential device architectures. These results suggest ways of designing better exoskeleton products. Optimized torques did not mimic biological torques, with magnitudes smaller than biological for all joints and peak torques later than biological peaks for the hips and ankles (Figure 4). Optimized torque magnitudes were within the range of reported capabilities of some existing mobile devices for the lightest participant (61 kg) (Mooney et al., 2014; Shepherd and Rouse, 2017; Lee et al., 2018) but not for the heaviest participant (90 kg) (Bryan et al., 2020). Exoskeletons could be designed in different sizes (e.g., small, medium, and large) that meet the optimized torque magnitudes for different sized users while minimizing worn mass. The similar timing parameters across users and across assistance strategies suggest that these optimized profiles could translate well to existing devices with lower capabilities and could be generalizable to a wide range of users. Device designers could consider control strategies that allow for kinematic adaptations because they seem to be useful to maximize device effectiveness. We also recommend considering state-based control at the knees, which was more effective than pilot tests of strictly torque-time control. This strategy may have facilitated adaptation to knee extension assistance during stance, because the torques grew as the
user bent more into flexion, allowing for a "stabilizing" effect that prevented buckling of the knee. In this study, we used a tethered exoskeleton emulator to compare assistance, but future work should attempt to recreate this assistance on mobile exoskeletons. We reported our improvements primarily relative to walking in the exoskeleton unassisted, and we expect these relative improvements to be consistent for mobile systems, but future mobile device development will need to focus on keeping worn mass low to ensure a good benefit relative to no exoskeleton as well. These findings should translate well to a wholeleg mobile device because the worn mass (13.5 kg) is similar to the expected mass of a mobile device Table 1. Metabolic reduction per joint assisted, relative to walking in the exoskeleton with no torque \| Condition \| Ankle-only(%) \| Hip-only(%) \| Hip–ankle(%) \| Knee–ankle(%) \| Whole-leg(%) \| Hip–knee(%) \| Knee-only(%) \| \|--------------------------\|-------------------\|-----------------\|------------------\|-------------------\|------------------\|-----------------\|------------------\| \| Reduction per \| 30 \| 26 \| 21 \| 18 \| 17 \| 16 \| 13 \| \| jointTotal reduction \| 30 \| 26 \| 42 \| 37 \| 50 \| 33 \| 13 \| capable of the optimized torque magnitudes (10 kg) based on published torque densities (Pratt et al., 2004; Zoss et al., 2005; Meijneke et al., 2014; Mooney et al., 2014; Giovacchini et al., 2015; Seo et al., 2016; John et al., 2017; Shepherd and Rouse, 2017; Bryan et al., 2020; Supplementary Material, Section 10). These results can be used to improve our models of human coordination, especially when using assistive devices. The larger metabolic cost reductions we saw from hip and ankle assistance support the idea that the hips and ankles are primary energy consumers during walking (Farris and Sawicki, 2012). Unlike some simulations (Franks et al., 2020), muscle activity did not go to zero even when assisted without hitting torque limits, indicating the user is optimizing for more than just metabolic cost for an average steady-state stride. Simulations could capture that more complicated objective function, including control required for balance. These results show that kinematic adaptations to assistance are important and should be considered in simulations. Biomechanical models could also be used to study the kinematic adaptations and relative metabolic benefits seen in this study. Muscle-level simulations of single-joint and multi-joint assistance could better understand how assistance at one joint could be indirectly assisting muscles at another joint, facilitated by the user's kinematic adaptations. This work could also try to understand how single-joint and multi-joint assistance compare in assisting biarticular muscles specifically. This study could have been improved by testing more participants, providing additional training, or testing additional controller parameterizations. This was an extremely arduous experiment with long optimization times, with each participant having completed over 50 hr of experiments. It was then interrupted by difficult external conditions (the COVID-19 pandemic). As such, we were only able to complete three participants for the single-joint and two-joint optimizations, and four participants for the whole-leg optimizations. However, given the magnitude of the changes and the consistency of the responses across participants, this sample size is sufficient to identify the efficacy of the joint combinations tested. Given three participants and a desired statistical power of 0.8, and assuming metabolic reductions have a standard deviation of 7.3% (Zhang et al., 2017), we can confidently detect metabolic reductions of 24% and larger. Although we gave our users substantial training and optimization time, more time may have improved the outcomes. Longitudinal studies with mobile devices that can be worn daily could show greater improvements to walking as users adapt. During optimization, torque magnitudes reached the comfort-based limits at the ankles, which were set due to discomfort at the biological ankle joint, possibly from extending the ankles too quickly or too far during torque application at push-off. If we were able to apply larger torque magnitudes comfortably, we could unconstrain the optimization of assistance and allow the torques to converge on the magnitude that provides the greatest benefit, which could be larger than 1 Nm/kg for the ankles. Using a more sophisticated control approach to ensure user comfort while allowing the largest possible exoskeleton torques might also lead to larger benefits. Due to the COVID-19 pandemic interrupting our protocol, participants had to wear cloth masks during optimization and validation of two-joint assistance, which could have impacted the accuracy of the metabolic cost measurements. While we believe the percent reductions should still be accurate, future work
studying two-joint exoskeleton assistance could confirm the results found here. These results suggest that new cost functions, gait environments, and user populations could be exciting topics for future studies. Future work could optimize metabolic cost alongside other costs that are important for gait, such as walking speed, balance, or user satisfaction (Ingraham et al., 2020). In our study, we did not rigorously collect user feedback, but participants often provided similar feedback, such as noting that the exoskeleton felt like it was making them march if torques were too large, or that they noticed how beneficial assistance was right after it was turned off. Future work could collect user feedback in an organized way and incorporate it in the design of exoskeleton assistance, possibly through humanin-the-loop optimization. Our study did not penalize high torques or powers, but future work could try to maintain sufficient metabolic cost reductions while minimizing actuator requirements, which can be costly to mobile devices. While our study assisted walking at a fixed speed on level-ground, future work can explore optimized assistance for walking in different conditions such as at different speeds, on inclines, or with worn loads. Our study was restricted to a treadmill due to our tethered device, but this work could be extended to unstructured environments by translating the paradigm to mobile devices. Our findings for assisting young, able-bodied users could be a starting point to optimize assistance for older adults and people with disabilities, hopefully speeding the discovery of effective assistance strategies. Acknowledgments. We would like to thank K. Gregorczyk, G. Kanagaki, M. O'Donovan, and the NSRDEC for their input on experimental design, N. Bianco for assistance in controller development and reference biological torques, and all of the Stanford Biomechatronics Lab for their feedback and support. We would also like to thank the staff and administrators who were able to reopen the lab during the COVID-19 pandemic. Funding Statement. This work was supported by the U.S. Army Natick Soldier Research, Development and Engineering Center (Grant number W911QY18C0140), by the National Science Foundation Graduate Research Fellowship Program (Grant number DGE-1656518), and by the Stanford Vice Provost for Undergraduate Education STEM Fellowship. Competing Interests. The authors declare no competing interests exist. Authorship Contributions. P.W.F. and G.M.B. designed and constructed the exoskeleton and developed the controllers. P.W.F. developed the optimization strategy and the experimental protocol, conducted experiments, and drafted and edited the manuscript. G.M.B., R.M.M., R.R., and A.C.L. conducted experiments and edited the manuscript. S.H.C. conceived and managed the project, provided design, controls and testing support, and edited the manuscript. Data Availability Statement. The data that support the findings of this study are available on request from the corresponding author, P.W.F., and will be made available online at biomechatronics.stanford.edu following publication. Ethical Standards. All user experiments were approved by the Stanford University Institutional Review Board and the US Army Medical Research and Materiel Command (USAMRMC) Office of Research Protections. All participants provided written informed consent before their participation as required by the approved protocol.
Task-agnostic exoskeleton control via biological joint moment estimation https://doi.org/10.1038/s41586-024-08157-7 Received: 17 October 2023 Accepted: 4 October 2024 Published online: 13 November 2024 Dean D. Molinaro1,2,5,8 ⋈, Keaton L. Scherpereel1,2,6,8, Ethan B. Schonhaut1, Georgios Evangelopoulos3,7, Max K. Shepherd4 & Aaron J. Young1,2 Lower-limb exoskeletons have the potential to transform the way we move1-14, but current state-of-the-art controllers cannot accommodate the rich set of possible human behaviours that range from cyclic and predictable to transitory and unstructured. We introduce a task-agnostic controller that assists the user on the basis of instantaneous estimates of lower-limb biological joint moments from a deep neural network. By estimating both hip and knee moments in-the-loop, our approach provided multi-joint, coordinated assistance through our autonomous, clothing-integrated exoskeleton. When deployed during 28 activities, spanning cyclic locomotion to unstructured tasks (for example, passive meandering and high-speed lateral cutting), the network accurately estimated hip and knee moments with an average $R^2$ of 0.83 relative to ground truth. Further, our approach significantly outperformed a best-case task classifier-based method constructed from splines and impedance parameters. When tested on ten activities (including level walking, running, lifting a 25 lb (roughly 11 kg) weight and lunging), our controller significantly reduced user energetics (metabolic cost or lower-limb biological joint work depending on the task) relative to the zero torque condition, ranging from 5.3 to 19.7%, without any manual controller modifications among activities. Thus, this task-agnostic controller can enable exoskeletons to aid users across a broad spectrum of human activities, a necessity for real-world viability. Lower-limb exoskeletons promise to reinvent human mobility by augmenting our capability and increasing longevity15,16. However, within powered exoskeleton technology lies a critical limitation: the controllers—which in many cases were optimized through extensive experimentation—only work for a single task or small set of tasks, offering little adaptability beyond passive devices17,18. Switching between tasks typically requires a 'high-level' task classification (for example, level walking, incline walking and stair ascent) often toggled manually or in some cases by an autonomous classifier1,19-24. Within each class, a 'mid-level' controller computes desired exoskeleton assistance, which is often tuned on a user- and task-specific basis3,11,25. For highly repetitive cyclic tasks, assistance is often designed as a function of time or phase3,10,12,24. For some gravity-fighting non-cyclic tasks, such as squats or sit-to-stand, impedance controllers assist, removing any dependence on time26-28. Although this approach has worked well for many laboratory-based experiments, this highly constrained discretization contrasts with the fluidity of natural human movement; we shuffle and side-step as we navigate a busy kitchen, stop our jog to take in a scenic view and regather our balance to again try the door that was heavier than we had anticipated. Our median walking bouts are a mere four steps29. Unstructured, non-cyclic and transitory tasks make up a large portion of our movements and interactions with the environment, but current exoskeleton controllers are incapable of recognizing or assisting these tasks. In fact, the expansion of traditional classification-based high-level control architectures to encompass these unstructured movements is intractable owing to the sheer number of movements that must be defined. We have developed a task-agnostic exoskeleton controller that short-circuits the need for high-level task classification or gait phase estimation by basing assistance on a fundamental, continuous physiological state: the human's biological joint moment (Fig. 1a and Supplementary Video 1). Biological moment can be calculated using optical motion capture and high-fidelity force plates to measure interactions with the ground30 but cannot be measured or solved for analytically through available wearable sensors, owing to sensor noise and incomplete information (particularly shear forces with the ground). Instead, biological joint moments can be estimated from wearable sensor data, often by including optimization or learning methods to account for incomplete sensor information 13,14,31-36, but very few studies have begun to explore the implications of using this technology in the control loop13,14,35,36. In these previous works, however, using instantaneous biological joint moment estimates in the control loop has shown substantial promise. Gasparri et al.13 developed a joint moment-based ankle exoskeleton controller, which has shown large metabolic benefits across inclines and declines, stairs and mixed terrain in both unimpaired individuals and those with cerebral palsy $^{
13,37-39}$ . Additionally, George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA, USA. Institute for Robotics and Intelligent Machines, Georgia Institute of Technology, Atlanta, GA, USA. 3X, The Moonshot Factory, Mountain View, CA, USA. 4College of Engineering, Bouvé College of Health Sciences, and Institute for Experiential Robotics, Northeastern University, Boston, MA, USA. 5 Present address: Boston Dynamics Al Institute, Cambridge, MA, USA. 6 Present address: Skip Innovations, San Francisco, CA, USA. 7 Present address: Google Mountain View, CA, USA. 8These authors contributed equally: Dean D. Molinaro, Keaton L. Scherpereel. ™e-mail: molinarodean@gmail.com $\label{Fig.1} \textbf{Fig.1}\| \textbf{Task-agnostic exoskeleton control with a clothing-integrated} \ \textbf{exoskeleton. a}, \textbf{The proposed approach continuously and seamlessly provides} \ assistive torque to the hip and knee using an estimate of the user's biological joint moments from a deep neural network. By basing assistance off a continuous physiological variable, no task classification is required; the same control law can effectively assist across the full range of human movement. The time series shown illustrates the average performance of our control approach with representative participant-averaged curves on the tasks shown. ext., extension. \textbf{b}, \textbf{A} \textbf{n} \textbf{autonomous hip-knee exoskeleton system was constructed to} \$ capture a rich set of sensing modalities and then assist across a wide range of mobility tasks. $\mathbf{c}$ , The hybrid design consists of both soft textiles and semirigid structural components to efficiently transmit exoskeleton assistance to both the hip and knee joints. The human interface consisted of zero-stretch woven fabric to efficiently transmit forces, whereas low-stretch knitted fabrics covering joints helped avoid restrictions in the user's range of motion. $\mathbf{d}$ , Structural compliance and a passive translational degree of freedom between the hip and knee allowed hip ab- or adduction and rotation while maintaining actuator alignment across movements. energy shaping methods35,36 promise joint moment-based control agnostic to a specific lower-limb joint. Further, our previous work14 presented a deep learning-based approach that significantly reduced user metabolic cost during both level and incline walking using a hip exoskeleton; this approach demonstrated similar or even better outcomes than using previously optimized, spline-based assistance (that is, the previous standard of exoskeleton control) depending on the condition. Although these past studies demonstrate the efficacy of exoskeleton control using real-time biological joint moment estimates, they do not yet realize the key benefit of this approach: the potential for generalizability across the broad spectrum of human movement. In fact, these previous studies are almost entirely limited to the domain of level walking, ramps and stairs (domains that have been studied in exoskeleton control for decades15,16) aside from the sit–stand task investigated by Lin et al.35. Further, this approach could autonomously coordinate assistance across many lower-limb joints, a key component of generalizing exoskeleton technology across tasks that depend on different joints; however, significant augmentation of user energetics using a multi-joint exoskeleton controlled by joint moment estimates remains to be demonstrated. In this study, we introduce a task-agnostic controller enabled by a neural network-based joint moment estimator, which runs onboard an autonomous, hip-knee exoskeleton. By training the network on a diverse dataset of time-synced exoskeleton sensor data and ground-truthjoint moments, we found that it accurately estimated user joint moments during 28 cyclic and non-cyclic human activities when deployed online (corresponding dataset released with this study). Further, we found that the resulting controller significantly reduced metabolic cost (four activities tested) and lower-limb biological joint work (six activities tested) relative to a no-assistance condition in all tested activities without any manual user or experimenter intervention between activities. This work provides a path to generalizing assistance across human behaviour, a critical link for the adoption of exoskeleton technology in the real world. Clothing-integrated robotic exoskeleton We developed a new exoskeleton with the capacity to assist an extensive range of movements, with a particular focus on under-appreciated functional movements that are critical to independence, but can be difficult for many populations, such as older adults. The clothing-integrated research exoskeleton presented here (Fig. 1b–d), was designed at X, The Moonshot Factory and combines the advantages of rigid exoskeletons with the comfort of soft textiles for the human–exoskeleton interface16,40 (Fig. 1c). Compact quasi-direct drive actuators (AK80-9 T-Motor
, Nanchang) mounted coaxially with the hip and knee provided up to 15 N m of assistance at each joint. The semirigid structure consisted of carbon fibre and 3D printed nylon orthotics on which the actuators and sensors were mounted. Six inertial measurement units (IMUs), joint encoders on the hips and knees, and a pair of wireless force-sensitive insoles provided real-time human movement data for the joint moment estimator (Fig. 1b) with the IMUs being the most important for joint moment estimation (Extended Data Fig. 1). Sagittal-plane actuation is provided at the hip and knee whereas passive degrees of freedom at the hip (translation and rotation) provide flexibility (Fig. 1d). This new exoskeleton architecture gave the user the flexibility and range of motion needed to perform a diverse range of structured and unstructured activities and represents a vital step forward in designing exoskeleton interfaces that are compliant, comfortable and adjustable. Lower-limb joint moment estimation To train the joint moment estimator, we collected exoskeleton sensor data time-synced with motion capture and ground reaction forces (GRFs) while users performed a wide range of tasks. Standard Open-Sim inverse dynamics (detailed in the Supplementary Information) were used to calculate hip and knee moments41,42, providing the ground-truth labels (Fig. 2a). To achieve both task generalizability and user-independence, our extensive dataset consisted of 15 healthy participants performing 28 different activities consisting of 66 total conditions (Extended Data Fig. 2). We categorized the 28 activities as cyclic (Supplementary Video 2), impedance-like (Supplementary Video 3) or unstructured (Supplementary Video 4) on the basis of normative joint biomechanics43 (Fig. 2b). Using this dataset, we trained a temporal convolutional network (TCN) with optimized hyperparameters (see Extended Data Table 1) to estimate hip and knee moments from 19 of the 28 tasks, with the 19 tasks chosen by a forward selection algorithm to promote task generalization within the model (Fig. 2c and Extended Data Fig. 3a). The most helpful data for model generalization (aside from the seed task of level-ground walking whose importance cannot be assessed) was a series of static standing poses, allowing the model to learn the static characteristics of the human body (for example, standing upright requires near zero moment), which is critical for generalization (Extended Data Fig. 3b). Other critical tasks, such as jump and cut, are extremely high-effort tasks that probably helped establish the bounds of the system dynamics and thus are also important for generalization. These previously understudied activities in the exoskeleton domain, many of which are not suitable for gait phase or impedance control, were the most critical for training a model to infer joint moments across real-world tasks. The joint moment estimates were mapped to applied exoskeleton torque by a continuous transformation (Fig. 2a). Hip and knee moments were scaled to 20 and 15% of the total estimated biological moments, respectively. These scaling factors were established in pilot experiments and provided comfortable assistance while preventing substantial saturation and overheating of the motors during high-torque movements. Hip moment estimates were delayed by 100 ms to maximize positive work done by the exoskeleton44 and to potentially minimize user metabolic cost14. Furthermore, the delay between knee moment estimates and the resulting assistance was set to the minimum achievable by the system (a delay of 50 ms), on the basis of single-blinded pairwise preference tuning45 during pilot testing. Finally, the delayed joint moment estimates were lowpass filtered to better match the frequency content of human movement46 and increase user comfort. The model was validated online with ten participants to assess its ability to accurately estimate human joint moments while providing assistance. No user-specific data were included in training to keep the tests user-independent. Furthermore, we developed a best-case baseline method based on current state-of-the-art exoskeleton control to compare against our joint moment estimator (details in the Supplementary Information); for cyclic activities, the baseline method estimated the user- and stride-averaged hip and knee moments from each activity (for example, for level walking, the baseline used the average level walking curve) and for 'impedance-like' activities (for example, jumping in place), the baseline method estimated the hip and knee moments by estimating zero moment when in swing or flight and by using a linear spring-damper model fit to each activity during stance (Fig. 3a). Unstructured activities were omitted from the baseline because of their lack of phase or impedance-like behaviour, which highlights the limitations of current exoskeleton control. We implemented the baseline method post hoc with perfect gait phase estimates and task classification (that is, a perfectly accurate classifier of 28 classes), thus representing the theoretical best possible performance achievable by this type of control architecture. Our deep neural network estimated hip and
knee moments significantly better than the baseline method for both cyclic (hip R2 0.79, knee R2 0.86) and impedance-like activities (hip R2 0.81, knee R2 0.87) without any participant-specific calibration (Fig. 3b,c). Representative time series are shown in Fig. 3d–f. Comparing within each activity, our estimator significantly outperformed the baseline method for 12 of the 19 total comparisons of R2 at the hip and 13 of the comparisons at the knee (Extended Data Fig. 4a,b), with similar results in root mean-square error (r.m.s.e.) (cyclic hip and knee r.m.s.e. 0.15 and 0.13 N m kg−1, impedance-like hip and knee r.m.s.e. 0.21 and 0.16 N m kg−1) and in normalized mean absolute error (MAE) (cyclic hip and knee normalized MAEs 7.3 and 5.5%; impedance-like hip and knee normalized MAEs 7.1 and 6.0%) (Extended Data Figs. 4c–f and 5a–d). The baseline method did not significantly outperform our approach on any individual activity in R2 , r.m.s.e. or normalized MAE. In reality, the high-level state estimators required for the baseline method (that is, a task classifier, gait phase estimator and pose estimator) also have non-zero error12,20–23,25, further detracting from this approach and highlighting the benefits of our regression-based method. We also found that during unstructured tasks that were not well-defined as cyclic or impedance-like, our approach maintained performance with an average hip R2 of 0.80 and knee R2 of 0.82 (Fig. 3b). Thus, our task-agnostic controller mimicked the natural behaviours of human movement, seamlessly modulating assistance throughout the transient motions common in daily life29. Given the black box nature of our approach it is possible that the neural network could generate large, erroneous joint moments leading to undesirable exoskeleton assistance. To analyse model under- and overestimation, we computed the normalized hip and knee estimate error at each time instance as the difference between the absolute value of the joint moment estimate and the absolute value of the ground-truth label, normalized by the peak-to-peak range of the ground-truth label. Extended Data Fig. 5e depicts the distribution of the normalized hip and knee error from the online validation trials (representing roughly 10 million instances total), in which negative and positive values correspond to under- and overestimates, respectively. As shown in the figure, large under- and overestimates were uncommon with means close to zero (hip mean −2.7%; knee mean −1.3%) and standard deviations of 7.8 and 6.6% for the hip and knee, respectively. Furthermore, time series examples of the most severe instances of under- and overestimation from the joint moment estimator are shown in Extended Data Fig. 5f. Of the 28 evaluated activities, nine were withheld from the training set (details in Methods). The average R2 of our estimator on these held-out tasks was 0.83 and 0.85 for the hip and knee, respectively, demonstrating the ability of the network to generalize to the hold-out tasks. To further investigate estimator generalization, three users also completed Fig. 2 \| Deep neural network training and deployment for joint moment estimation. a, Lower-limb joint moment labels were calculated in OpenSim using optical motion capture, force plate data and user-specific musculoskeletal models; a TCN was trained to predict these joint moment labels from timesynchronized exoskeleton sensor data. During deployment, to improve power delivery and user comfort, the estimates were transformed into commanded exoskeleton torque through a continuous function consisting of a scale, delay and a lowpass filter. est., estimate. b, Users wore the exoskeleton while performing a wide range of cyclic, impedance-like and unstructured tasks. c, Training activities for the moment estimator were selected using a forward selection algorithm to maximize the relative improvement in model generalization across tasks. Validation r.m.s.e. decreased as the training set grew, with the first 19 tasks reducing r.m.s.e. to 0.133 N m kg−1, which was within 5% of the best model accuracy with all the tasks included. This task set was used to train the real-time models used in the rest of this study. For reference, peak-topeak hip and knee moments ranged from 2 to 4 N m kg−1 for most activities in the dataset. Results were computed from leave-one-participant-out cross-fold validation using a 12-participant dataset (error bars omitted for clarity). eight completely new tasks,
described in Extended Data Table 2, that had not been previously tested or analysed (Supplementary Fig. 1 and Supplementary Video 5). These tasks were intentionally designed to be highly unique from the original dataset to push the limits of our approach, including burpees, mimicking a basketball layup and walking on a split belt treadmill with differing belt speeds. Our approach generalized well to the tasks reflective of typical human movement, and when pushed to extremely dynamic behaviours outside of the training set, our approach provided directionally correct assistance, but the magnitude and shape lost accuracy (R2 ranged from 0.24 to 0.92 at the hip and from 0.32 to 0.91 at the knee for the eight new tasks; Supplementary Fig. 1c,d). These results demonstrate the ability of the estimator to generalize to never-before-seen activities but highlights that task-specific training data is beneficial for activities with highly different dynamics (for example, when offloading bodyweight through the hands on the ground). Extra details and discussion comparing our joint moment estimator relative to previous methods and regarding its performance during new tasks are provided in the Supplementary Information. Augmenting user energetics across tasks To quantify the impact of our task-agnostic controller on the user, we measured user metabolic cost during four activities under three assistance conditions: wearing the exoskeleton with our task-agnostic controller (exo on), without wearing the exoskeleton (no exo) and wearing Fig. 3 \| Online joint moment estimation performance. a, We compared our neural network-based joint moment estimator (deployed online) to a best-case baseline method (computed offline) that relied on perfect task classification and gait phase. b, Our approach significantly improved $R^2$ at the hip by $0.13 \pm 0.04$ (19 $\pm$ 6%, P = 2 $\times$ 10 $^{-7}$ ) and at the knee by $0.25 \pm 0.04$ (40 $\pm$ 7%, P < 10 $^{-16}$ ) compared to the baseline method during cyclic activities. For impedance-like tasks, our approach improved $R^2$ by $0.31 \pm 0.05$ (63 $\pm 10\%$ , $P = 4 \times 10^{-16}$ ) at the hip and by $0.32 \pm 0.02$ (57 $\pm 3\%$ , $P < 10^{-16}$ ) at the knee compared to the baseline method. Black squares depict inter-participant mean, coloured boxes depict interquartile range, horizontal lines within boxes depict inter-participant median and error bars depict inter-participant minimum and maximum (n=10). c, Estimator $R^2$ is shown per task for our approach and the baseline method. Each marker corresponds to the inter-participant average per single task (n = 10, except for the run condition where n = 9). d, Representative strides from various cyclic tasks are shown. The baseline method required a different task classification for each depicted ambulation mode, whereas our approach did not require any discrete switching. e, Representative trials are shown when squatting to the left, right and symmetrically. The impedance control-based approach failed to capture changes in joint moments by relying solely on kinematics. Instead, our approach accurately modified joint moments with the change in weight distribution across the user's legs. $\mathbf{f}$ , A representative trial during leftward cutting is shown, depicting the ability of our approach to seamlessly modulate assistance during highly unstructured behaviours. As it is unclear how to extend the baseline method to these types of activity, it was omitted. Estimator $R^2$ relative to ground truth is shown for our approach (ours) and the baseline method (base) above each representative trial. the exoskeleton without assistance (zero torque). The task-agnostic controller significantly reduced metabolic cost for all four tasks compared to zero torque (P < 0.05) with relative reductions ranging from 8.0% during the lift weight task to 19.7% during 5° inclined walking (Fig. 4a). Relative to no exo, our approach significantly reduced user metabolic cost during the weight lifting task and during running (P < 0.05); however, our approach increased metabolic cost during level-ground walking (P < 0.05). Given the similarity in level walking estimator accuracy and metabolic cost reduction relative to zero torque in our previous work, which did reduce metabolic cost relative to no exo using a lighter weight hip-only exoskeleton14, it is likely that the increase in metabolic cost in this study was due to the added mass penalty of the hip-knee exoskeleton, not the controller itself47. Nevertheless
, these results demonstrate the ability of our approach to autonomously modulate assistance across tasks in a beneficial manner, a critical hurdle in developing task-agnostic exoskeleton controllers. To further quantify the effect of the device during transient tasks, we measured metabolic cost for three participants performing a varying speed and incline circuit that ranged from walking to running with inclines ranging from 0° to 15°. During this highly transient trial, our approach reduced user metabolic cost by 12.9% relative to zero torque and by 1.6% relative to no exo (Fig. 4b). Thus, our approach seamlessly accommodated these transient behaviours so common to daily life29 without any extra tuning or calibration. Furthermore, positive lower-limb biological joint work of the user was also evaluated during six extra tasks described in Extended Data Table 3, which provided insight into the joint-level effects of our controller14,48-50 and was less taxing on the participants than the metabolic trials. Total positive lower-limb biological joint work was computed as the sum of the components computed at the hip, knee and ankle. Positive biological joint work was computed by integrating the positive biological power at each joint, which involved subtracting the Fig. 4 \| Human outcome performance. a, Participant metabolic cost was measured during four activities while wearing the exoskeleton using the task-agnostic controller (exo on), without wearing the exoskeleton (no exo) and while wearing the exoskeleton without assistance (zero torque). Tasks other than running were conducted using 6 min trials and a counter-balanced design (ABCCBA). Owing to the strenuous nature of the running trials, conditions were only completed once (ABC) and each trial only lasted 3.5 min. Below each activity is the cycle-averaged commanded torque during exo on as a function of movement percentage. The shaded region around each curve depicts ±1 standard deviation about the mean. b, Three participants returned and performed a varying speed and incline protocol while metabolic cost was measured. Individual traces are provided for each participant as well as the average. The detailed variations in speed and incline are shown below the activity. c, The average positive biological joint work per movement cycle summed across each participant's hip, knee and ankle are shown. Participants completed the six activities under the same three assistance conditions as the metabolic trials. Again, the cycle-average commanded torque during exo on for each activity is shown as a function of movement percentage whereas the shaded region around each curve depicts ±1 standard deviation about the mean. Each black square depicts the inter-participant mean, each coloured box depicts the interquartile range, each horizontal line within the boxes depicts the interparticipant median and each error bar depicts the inter-participant minimum and maximum. Asterisks indicate statistical significance (P < 0.05; exact P values are provided in Supplementary Data 1). exoskeleton torque from the ground-truth total joint moment from inverse dynamics (further details are provided in the Methods section). Our controller significantly reduced positive lower-limb biological joint work of the user during all six tasks compared to zero torque (P < 0.05) with decreases ranging from 5.3 to 15.7% (Fig. 4c). During four of the six tasks, our controller also reduced positive joint work compared to no exo with significant decreases ranging from 8.5 to 22.3%. The other two tasks, stair ascent and step up, showed no significant difference. Furthermore, participant lower-limb kinematics showed little variation across exoskeleton conditions (Supplementary Fig. 2), suggesting that reductions in biological joint work were achieved primarily through reductions in biological joint moments as opposed to modified kinematics (details in the section 'Exoskeleton effects on user kinematics' in the Supplementary Information). With further improvements to the exoskeleton, such as reducing exoskeleton mass, relocating knee actuators more proximal to the body47, and increasing the maximum magnitude of assistance, we anticipate substantially greater capacity for our approach to enhance human performance across activities compared to this first evaluation. Furthermore, it is likely that our mid-level controller was not optimal for all tasks and similarly may not be optimal for populations outside of young, able-bodied individuals. Further optimization of the mid-level controller3,8,11, consideration of how this approach could generalize to extra populations and investigation of the physiological mechanisms that drive the relationship between exoskeleton assistance and user outcomes could result in further improvements in user outcomes and expand the scope of this approach. Nevertheless, these comparisons demonstrate the efficacy of our task-agnostic controller to dynamically and beneficially modulate assistance with changing user behaviour without user- or task-specific tuning, which is a critical component for exoskeleton controllers deployed in the real world. Conclusion By relying on internal physiological state estimates rather than humanengineered gait parameterizations, control is given back to the user and the exoskeleton can respond to that user's specific, real-time joint moments without
any user-specific calibration or hand-engineered state machine criteria. Whereas recent work has demonstrated how to leverage instantaneous joint moment estimates as a promising alternative for exoskeleton control during walking13,14,35,36, here we present the missing piece: task generalization. Specifically, our deep neural network approach to exoskeleton control provides a task-agnostic framework, capable of seamlessly augmenting human effort by coordinating exoskeleton assistance across joints during both structured and unstructured tasks. Building from these advances, we believe this technology will become paramount in many industries, allowing researchers to test new hypotheses during natural, less-structured behaviours, relieving worker exhaustion in manual labour industries and enabling disaster relief teams to act with extra endurance in time-sensitive missions. Online content Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41586-024-08157-7. 1. Kim, J. et al. Reducing the metabolic rate of walking and running with a versatile, portable exosuit. Science 365, 668–672 (2019). Witte, K. A., Fiers, P., Sheets-Singer, A. L. & Collins, S. H. Improving the energy economy of human running with powered and unpowered ankle exoskeleton assistance. Sci. Robot. 5, eaay9108 (2020). 3. Slade, P., Kochenderfer, M. J., Delp, S. L. & Collins, S. H. Personalizing exoskeleton assistance while walking in the real world. Nature 610, 277–282 (2022). Awad, L. N. et al. A soft robotic exosuit improves walking in patients after stroke. Sci. Transl. Med. 9, eaai9084 (2017). Malcolm, P., Derave, W., Galle, S. & Clercq, D. D. A simple exoskeleton that assists plantarflexion can reduce the metabolic cost of human walking. PLoS ONE 8, e56137 (2013). Mooney, L. M., Rouse, E. J. & Herr, H. M. Autonomous exoskeleton reduces metabolic cost of human walking during load carriage. J. NeuroEngineering Rehabil. 11, 80 (2014). Ishmael, M. K., Archangeli, D. & Lenzi, T. Powered hip exoskeleton improves walking economy in individuals with above-knee amputation. Nat. Med. 27, 1783–1788 (2021). 8. Franks, P. W. et al. Comparing optimized exoskeleton assistance of the hip, knee, and ankle in single and multi-joint configurations. Wearable Technol. 2, e16 (2021). Baltrusch, S. J. et al. The effect of a passive trunk exoskeleton on metabolic costs during lifting and walking. Ergonomics 62, 903–916 (2019). 10. Lim, B. et al. Delayed output feedback control for gait assistance with a robotic hip exoskeleton. IEEE Trans. Robot. 35, 1055–1062 (2019). 11. Zhang, J. et al. Human-in-the-loop optimization of exoskeleton assistance during walking. Science 356, 1280–1284 (2017). 12. Shepherd, M. K., Molinaro, D. D., Sawicki, G. S. & Young, A. J. Deep learning enables exoboot control to augment variable-speed walking. IEEE Robot. Autom. Lett. 7, 3571–3577 13. Gasparri, G. M., Luque, J. & Lerner, Z. F. Proportional joint-moment control for instantaneously adaptive ankle exoskeleton assistance. IEEE Trans. Neural Syst. Rehabil. Eng. 27, 751–759 (2019). 14. Molinaro, D. D., Kang, I. & Young, A. J. Estimating human joint moments unifies exoskeleton control, reducing user effort. Sci. Robot. 9, eadi8852 (2024). 15. Siviy, C. et al. Opportunities and challenges in the development of exoskeletons for locomotor assistance. Nat. Biomed. Eng. 7, 456–472 (2023). 16. Sawicki, G. S., Beck, O. N., Kang, I. & Young, A. J. The exoskeleton expansion: improving walking and running economy. J. NeuroEngineering Rehabil. 17, 25 (2020). 17. Collins, S
. H., Wiggin, M. B. & Sawicki, G. S. Reducing the energy cost of human walking using an unpowered exoskeleton. Nature 522, 212–215 (2015). 18. Yang, J., Park, J., Kim, J., Park, S. & Lee, G. Reducing the energy cost of running using a lightweight, low-profile elastic exosuit. J. NeuroEngineering Rehabil. 18, 129 (2021). 19. Li, Y. D. & Hsiao-Wecksler, E. T. Gait mode recognition and control for a portable-powered ankle-foot orthosis. In Proc. 2013 IEEE 13th International Conference on Rehabilitation Robotics (ICORR) 1–8 (IEEE, 2013). 20. Laschowski, B., McNally, W., Wong, A. & McPhee, J. Environment classification for robotic leg prostheses and exoskeletons using deep convolutional neural networks. Front. Neurorobotics 15, 730965 (2022). Kang, I., Molinaro, D. D., Choi, G., Camargo, J. & Young, A. J. Subject-independent continuous locomotion mode classification for robotic hip exoskeleton applications. IEEE Trans. Biomed. Eng. 69, 3234–3242 (2022). Camargo, J., Flanagan, W., Csomay-Shanklin, N., Kanwar, B. & Young, A. A machine learning strategy for locomotion classification and parameter estimation using fusion of wearable sensors. IEEE Trans. Biomed. Eng. 68, 1569–1578 (2021). 23. Qian, Y. et al. Predictive locomotion mode recognition and accurate gait phase estimation for hip exoskeleton on various terrains. IEEE Robot. Autom. Lett. 7, 6439–6446 (2022). 24. Medrano, R. L., Thomas, G. C., Keais, C. G., Rouse, E. J. & Gregg, R. D. Real-time gait phase and task estimation for controlling a powered ankle exoskeleton on extremely uneven terrain. IEEE Trans. Robot. 39, 2170–2182 (2023). 25. Kang, I. et al. Real-time gait phase estimation for robotic hip exoskeleton control during multimodal locomotion. IEEE Robot. Autom. Lett. 6, 3491–3497 (2021). 26. Huo, W. et al. Impedance modulation control of a lower-limb exoskeleton to assist sit-to-stand movements. IEEE Trans. Robot. 38, 1230–1249 (2022). Yang, X. et al. Spine-inspired continuum soft exoskeleton for stoop lifting assistance. IEEE Robot. Autom. Lett. 4, 4547–4554 (2019). 28. Shepherd, M. K. & Rouse, E. J. Design and validation of a torque-controllable knee exoskeleton for sit-to-stand assistance. IEEE ASME Trans. Mechatron. 22, 1695–1704 (2017). 29. Orendurff, M. S., Schoen, J. A., Bernatz, G. C., Segal, A. D. & Klute, G. K. How humans walk: bout duration, steps per bout, and rest duration. J. Rehabil. Res. Dev. 45, 1077–1089 (2008). 30. Winter, D. in Biomechanics and Motor Control of Human Movement Ch. 5, 107–138 (John Wiley & Sons, Ltd, 2009). 31. Dorschky, E. et al. CNN-based estimation of sagittal plane walking and running biomechanics from measured and simulated inertial sensor data. Front. Bioeng. Biotechnol. 8, 604 (2020). 32. Molinaro, D. D., Kang, I., Camargo, J., Gombolay, M. C. & Young, A. J. Subject-independent, biological hip moment estimation during multimodal overground ambulation using deep learning. IEEE Trans. Med. Robot. Bionics 4, 219–229 (2022). Camargo, J., Molinaro, D. & Young, A. Predicting biological joint moment during multiple ambulation tasks. J. Biomech. 134, 111020 (2022). 34. Hossain, M. S. B., Guo, Z. & Choi, H. Estimation of lower extremity joint moments and 3D ground reaction forces using IMU sensors in multiple walking conditions: a deep learning approach. IEEE J. Biomed. Health Inform. https://
doi.org/10.1109/JBHI.2023.3262164 (2023). 35. Lin, J., Divekar, N. V., Thomas, G. C. & Gregg, R. D. Optimally biomimetic passivity-based control of a lower-limb exoskeleton over the primary activities of daily life. IEEE Open J. Control Syst. 1, 15–28 (2022). 36. Zhang, J., Lin, J., Peddinti, V. & Gregg, R. D. Optimal energy shaping control for a backdrivable hip exoskeleton. In Proc. 2023 American Control Conference (ACC) 2065–2070 (IEEE, 2023). 37. Fang, Y., Orekhov, G. & Lerner, Z. F. Improving the energy cost of incline walking and stair ascent with ankle exoskeleton assistance in cerebral palsy. IEEE Trans. Biomed. Eng. 69, 2143–2152 (2022). Bishe, S. S. P. A., Nguyen, T., Fang, Y. & Lerner, Z. F. Adaptive ankle exoskeleton control: validation across diverse walking conditions. IEEE Trans. Med. Robot. Bionics 3, 801–812 (2021). 39. Tagoe, E. A., Fang, Y., Williams, J. R. & Lerner, Z. F. Walking on real-world terrain with an ankle exoskeleton in cerebral palsy. IEEE Trans. Med. Robot. Bionics 6, 202–212 (2024). 40. Young, A. J. & Ferris, D. P. State of the art and future directions for lower limb robotic exoskeletons. IEEE Trans. Neural Syst. Rehabil. Eng. 25, 171–182 (2017). 41. Delp, S. L. et al. OpenSim: open-source software to create and analyze dynamic simulations of movement. IEEE Trans. Biomed. Eng. 54, 1940–1950 (2007). 42. Seth, A. et al. OpenSim: simulating musculoskeletal dynamics and neuromuscular control to study human and animal movement. PLoS Comput. Biol. 14, e1006223 (2018). 43. Scherpereel, K., Molinaro, D., Inan, O., Shepherd, M. & Young, A. A human lower-limb biomechanics and wearable sensors dataset during cyclic and non-cyclic activities. Sci. Data 10, 924 (2023). 44. Ding, Y. et al. Effect of timing of hip extension assistance during loaded walking with a soft exosuit. J. NeuroEngineering Rehabil. 13, 87 (2016). 45. Ingraham, K. A., Tucker, M., Ames, A. D., Rouse, E. J. & Shepherd, M. K. Leveraging user preference in the design and evaluation of lower-limb exoskeletons and prostheses. Curr. Opin. Biomed. Eng. 28, 100487 (2023). 46. Winter, D. A. Biomechanical motor patterns in normal walking. J. Mot. Behav. 15, 302–330 (1983). 47. Browning, R. C., Modica, J. R., Kram, R. & Goswami, A. The effects of adding mass to the legs on the energetics and biomechanics of walking. Med. Sci. Sports Exerc. 39, 515 (2007). 48. Farris, D. J. & Sawicki, G. S. The mechanics and energetics of human walking and running: a joint level perspective. J. R. Soc. Interface 9, 110–118 (2012). Farris, D. J., Hampton, A., Lewek, M. D. & Sawicki, G. S. Revisiting the mechanics and energetics of walking in individuals with chronic hemiparesis following stroke: from individual limbs to lower limb joints. J. NeuroEngineering Rehabil. 12, 24 (2015). 50. Farris, D. J. & Sawicki, G. S. Linking the mechanics and energetics of hopping with elastic ankle exoskeletons. J. Appl. Physiol. 113, 1862–1872 (2012). Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
© The Author(s), under exclusive licence to Springer Nature Limited 2024 Methods Autonomous robotic hip–knee exoskeleton In this study, we used the clothing-integrated robotic exoskeleton developed at X, which was designed to enhance human mobility by providing powered sagittal-plane assistance to the hips and knees (design rationale provided in the supplementary section 'Why the hip and knee and not the ankle?'). A Raspberry Pi 4B (RPi) (Raspberry Pi) served as the primary onboard computer that ran the exoskeleton control loop at 55 Hz. The RPi managed Controller Area Network (CAN) Bus and Bluetooth communication with peripherals, saved experimental data locally and provided all functions other than joint moment estimates. Joint moment estimates were generated on a machine learning coprocessor (NVIDIA Jetson Nano) also mounted onboard the device; thus, all computation was fully onboard the exoskeleton. The Jetson Nano provided a low power consumption (5 V, 2 A) coprocessor, easily integrated into the exoskeleton by means of ethernet connection and commonly available portable charging banks. Actuated hip flexion and extension and knee flexion and extension was provided by quasi-direct drive actuators (T-Motor AK80-9s, Nanchang), with a peak intermittent torque constrained to 15 N m. Open-loop torque commands were sent to the actuators from the RPi over CAN and encoder measurements were returned to the RPi. Encoder velocity was lowpass filtered using a second-order Butterworth filter with a 10 Hz cut-off frequency. Six-axis IMUs (OpenIMUA) were mounted to the shank and thigh struts and communicated with the RPi by means of CAN. Pressure-sensitive insoles (Moticon) measured vertical GRF and centre of pressure (COP) and had an embedded six-axis IMU. They communicated with the RPi by Bluetooth and were powered by coin-cell batteries. The RPi and actuators were powered by two 20 V, 3 Ah drill batteries (DeWalt) connected in parallel, providing power for roughly 2 h of continuous walking at a minimum. The RPi interfaced through WiFi with a laptop for data visualization using a custom user interface. The human–robot interface was designed to be as compliant, comfortable and adjustable as possible, while maintaining the minimal structure required for effective transfer of actuator torques to the body. The semirigid structure included thin, waterjet-cut carbon fibre plates in the thigh struts that were compliant in ab- or adduction and internal or external rotation of the hip, but supported flexion and extension torques. A passive translational degree of freedom allowed the shortening of the thigh strut required for hip ab- or adduction. The 3D printed nylon shin struts, shin cuffs and pelvic orthosis were designed to apply the assistance to comfortable sections on the shin and pelvis. The custom trousers provided tight integration of the semirigid structure with the body, preventing buckling of the thin thigh struts. Quick release snaps, adjustable with Velcro, connected the pelvic orthosis and thigh or shank struts to the trousers. Woven zero-stretch fabric on the thigh and shank allowed slop-free transfer of the applied exoskeleton torque to the leg through a 'hammocking' effect. Knit fabric, with some stretch and elasticity, around the knee and hip allowed free range of motion and was used on the lateral sides of the trousers to allow variation in sizing between users. Zippers on the shank allowed tight integration of the fabric with the lower shank while allowing don and doff over the heel. Overall, the entire exoskeleton added roughly 7 kg to the user depending on the exoskeleton size worn during the experiment. Furthermore, although comfort was a key consideration of the exoskeleton design, some participants voiced discomfort regarding the shape of the backplate and the load-bearing seams of the soft textile trousers. Real-time joint moment estimation To deploy the joint moment estimator within the exoskeleton controller, we integrated the machine learning coprocessor (that is, the Jetson Nano) into the device using an asynchronous TCP/IP connection over wired ethernet with the RPi. With each control loop, sensor data were measured from the actuators, IMUs and pressure insoles. The sensor data were sent from the RPi to the coprocessor, which returned estimates of the total hip and knee flexion or extension moments from the neural network. Biological joint moment estimates were then computed by subtracting the measured actuator torques from the previous loop from each corresponding joint moment estimate. Desired torque assistance was computed from the resulting biological joint moment estimates using three steps. (1) The biological joint moments at the hip and knee were scaled by 20 and 15%, respectively, to maximize assistance while maintaining safe operating regions for the device hardware (15 N m at each joint). (2) The scaled hip and knee moments
were then delayed by 100 and 50 ms, respectively. This delay was chosen at the knee because 50 ms was the minimum possible delay to guarantee a consistent relationship between biological joint moment estimates and exoskeleton assistance owing to limitations in loop rate reliability of the exoskeleton. Hip assistance was further delayed by an extra 50 ms because this approach can maximize the positive work done by the exoskeleton during walking, which can lead to further benefits for the user14,44. (3) The exoskeleton assistance was lowpass filtered using a second-order 10 Hz Butterworth filter to preserve the frequency content of human motion43,51–53 while removing jitter from the estimator14. This filter added an extra delay of 25 ms, resulting in a total hip delay of 125 ms and knee delay of 75 ms. In the case in which the resulting assistance was larger than the peak exoskeleton torque, commanded torque was clamped to the peak exoskeleton torque. Because the exoskeleton controller intentionally delayed the hip moment estimates by an extra 50 ms relative to the minimum achievable system delay with our system, we chose to train the network to estimate hip moments delayed by 50 ms relative to the input sequence. Our previous work found that delaying joint moment estimates relative to the input sequence can further improve model accuracy54, and we found that this approach resulted in an extra 5% improvement in hip moment validation MAE for this study. Neural network architecture optimization In our previous work, we achieved state-of-the-art accuracy in estimating user joint moments with a TCN32,54. In this study, we implemented the TCN as originally introduced by Bai et al.55 and modified for joint moment estimation in our previous work14,32,54. The TCN input consisted of a sequence of unilateral hip and knee encoder data, thigh, shank and foot IMU data, and pressure insole data (vertical GRF and COP). Owing to the exoskeleton loop rate of 55 Hz, the data were upsampled in real-time to 200 Hz to match the sampling frequency previously used for the TCN. Each of the model inputs were also normalized using their corresponding mean and standard deviation computed from the training set. The TCN was designed with two output heads for the instantaneous estimates of the total hip and knee flexion or extension moments (that is, the sum of exoskeleton torque and human biological moment). Joint moment labels in the training set were scaled by participant body mass during training, such that the model was trained to estimate joint moments in units of N m kg−1 (refs. 32,54). Furthermore, we trained the TCN to estimate the total joint moments to maintain the relationship between exoskeleton sensor data and TCN joint moment outputs, regardless of the specific parameters of the exoskeleton controller, such as assistance magnitude. We then computed biological joint moments later in the control framework by subtracting the exoskeleton torque from the total estimated moment. In our previous work, we conducted a thorough hyperparameter optimization of the TCN for estimating sagittal-plane hip moments32, however, this approach did not consider model generalizability (that is, during the hyperparameter optimization the model training set and validation set consisted of the same ambulation modes). Furthermore, this optimization was conducted under different conditions (only sagittal-plane hip moments, cyclic ambulatory activities and kinematic sensors). Thus, we conducted a rigorous hyperparameter optimization using a multi-stage approach, specifically targeting model generalizability under conditions consistent with this study (Extended Data Table 1). In stage 1 of the optimization, we used our previous dataset of human lower-limb biomechanics during cyclic and non-cyclic activities43 to optimize a large, 11-dimensional hyperparameter space using Bayesian optimization implemented in Vizier56 (training and testing more than 10,000 models). In stage 2, we finetuned the network hyperparameters using actual exoskeleton sensor data from phases 1 and 2 of the experimental protocol (details below) over a smaller, six-dimensional search space that could be rigorously optimized using grid search (training and testing 1,440 models). The six-dimensional space was constructed to finetune the most sensitive network hyperparameters determined from the marginal and conditional results of the stage 1 optimization. The resulting network hyperparameters (shown in Extended Data Table 1) resulted in an 8% improvement in MAE using leave-one-participant-out cross-fold validation compared to using the original hyperparameters from Molinaro et al.32. Task optimization for generalizability Because the collection of actuated, motion capture-labelled data is difficult and costly, we first sought to discover a subset of tasks that could allow a user-independent lower-limb joint moment estimator to generalize to the rest of human activities. To determine the subset of training activities that best promoted generalization, we used the same dataset used for stage 1 of
the hyperparameter optimization to conduct a forward activity selection optimization (Extended Data Fig. 3a). During each optimization step, the TCN was trained and tested using leave-one-participant-out validation to compute the expected model performance when evaluated on a new participant. First, model performance was computed by training the model using only the level-ground walking data but tested on all activities. The TCN was then iteratively trained from random initialization, including one activity from the candidate task set into the training set at a time. On the ith optimization step, the relative improvement in generalizability $s_{\sigma}[i]$ associated with including a candidate activity (g) into the training set (that is, the relative improvement in model performance on all activities beside g) was computed as $$s_{g}[i] = \frac{1}{n-1} \sum_{i=0}^{n-1} e_{g^{}j}[i-1] - e_{g,j}[i], \quad j \neq g,$$ (1) $$g^[i] = \underset{g}{\operatorname{argmax}} s_g[i] \ \forall g \in G[i],$$ (2) where n is the total number of tasks and $e_{g,j}[i]$ is the MAE of estimating joint moments during the jth task when trained using the updated training set over the first i-1 optimization steps and the extra candidate task g. Thus, $s_g[i]$ evaluated the overall improvement in model performance across all activities, excluding the relative improvement of the task at hand. Furthermore, $g^*[i]$ was the activity associated with the largest improvement score, which was then added to the training set for all further optimization steps and removed from the set of candidate tasks G to be selected in the next optimization step. This process was repeated until all tasks were selected. We defined the activity set that saturated model generalization as the minimum set of selected activities that contributed more than 95% of the sum total of relative improvement in generalizability across the complete optimization. After the first seven tasks, the model had reached this threshold, indicating that further tasks failed to substantially improve the model's ability to estimate joint moments on other tasks (Extended Data Fig. 3b). Aside from level-ground walking, which was used to seed the optimizer, none of the selected tasks was cyclic and four were unstructured. Although model generalization saturated rapidly, task-specific data continued to improve user-independent, task-specific validation error (Fig. 2c). The overall model performance required 19 selected activities before validation MAE fell within 5% of the validation MAE when trained on all activities. This demonstrates that there was added benefit to be gained from training on task-specific data even after generalization saturated. For all further analyses, the TCN was trained using the data from these 19 selected tasks unless otherwise stated. Sensor contribution to moment estimation Previous joint moment estimation studies often use data from IMUs and joint encoders as model inputs 14,32,34,54,57,58; however, the relative importance of each of these sensors (and others) on model performance is less explored. Furthermore, the sensitivity of model performance relative to sensor dropout (for example, from sensor disconnection) is also a critical real-world consideration. To investigate these two topics, we trained several extra models under two different conditions: first, using different subsets of available sensors and second, simulating sensor disconnection during model deployment. We tested the performance of these models offline on the data from the ten participants used to test our approach online. For the first condition, two of the sensor sets were inspired by common exoskeleton design choices: (1) removing the GRF and COP contributed by the insoles (that is, -insoles) and (2) removing all foot mounted sensing (that is, -insoles, -foot IMU). The other three sensor sets were chosen to demonstrate the contribution from each unique sensor modality: (1) only using the thigh and shank IMUs (IMU only), (2) only using the hip and knee encoders (encoder only) and (3) only using the GRF and COP from the foot insole (insole only). For the second condition, a single sensor was effectively dropped out during model testing, by zeroing that respective sensor's inputs to simulate a sensor losing connection during device deployment. These comparisons are presented in Extended Data Fig. 1. In the first condition, removing the GRF and COP resulted in a 0.03 and a 0.07 reduction in $\mathbb{R}^2$ at the hip and knee, respectively, demonstrating the moderate benefit of adding kinetic based sensing, with an additional penalty when removing the foot IMU. The 'IMU only' condition resulted in a further drop in performance; the benefit gained from the enc
oders indicated that our six-axis IMUs did not fully capture the relevant kinematic input information. Overall, the IMUs contributed the most to the accuracy of the model, followed by the encoders and, last, the insole. Previous work has shown that kinematic sensors can be effectively used to estimate GRFs, indicating a potential reason the IMUs contributed the most to the model $^{34,59}$ . Likewise, because the insoles initially measure pressure, from which the GRF is calculated, the amount of information provided to the model may be less than other more accurate sensors, such as in-ground force plates. In the second condition, dropping out different sensors showed the reliance of our network on each sensor. In general, loss of a sensor resulted in a significant drop in accuracy, indicating that the model generally used all available sensing inputs in the neural network weights. The loss of the knee encoder, however, was much worse for estimating knee joint moments and, similarly, the loss of the hip encoder was much worse for estimating hip joint moments. The losses of the thigh IMU and foot insole were likewise much more consequential to model accuracy when compared to the loss of the shank or foot IMU. These results indicate that our trained network learned to rely on each sensor for generating joint moment estimates, underscoring the importance of the need for high-quality hardware and sensor integration; however, further analyses could explore training the network with synthesized sensor dropout to potentially improve robustness. Experimental data collection The data used in this study were collected over three phases to facilitate model training and online testing of the joint moment estimator. Over the three phases, a total of 22 able-bodied participants participated in the study protocol. Participants provided written informed consent to participate in the study under Georgia Institute of Technology Institutional Review Board protocol H21184. Consent to publish participant images was also obtained. In each phase, participants completed a set of 28 cyclic and non-cyclic activities, consisting of 66 total conditions as detailed in Scherpereel et al.43 and outlined in Extended Data Fig. 2. During these activities, retroreflective markers on both the body and exoskeleton were tracked using a motion capture system at 200 Hz (Vicon Motion Systems). Furthermore, overground force plates and an instrumented treadmill (Bertec) were used to measure GRFs at 1,000 Hz. Owing to software limitations early in this study, motion capture data collected for two participants during phase 1 were collected at 120 Hz and subsequently upsampled to 200 Hz. To sync the exoskeleton data with the participants' ground-truth joint biomechanics, the exoskeleton data were first upsampled to 200 Hz to match the frequency of the motion capture data. At the start of each exoskeleton trial when motion capture and GRF data were collected, participants kicked three times with their right leg while exoskeleton actuation was off. On the basis of this movement, the exoskeleton data were time shifted to maximize the R2 between the right knee encoder and the resulting right knee joint kinematics computed from the biomechanical model. Phase 1: initial data collection In phase 1, ten participants (six males, four females, age of 23.7 ± 2.0 years, height of 176.3 ± 8.7 cm and body mass of 76.5 ± 13.3 kg) participated in a single day protocol collecting exoskeleton data and ground-truth human joint moments during the 28 cyclic and non-cyclic activities. As we could not deploy the task-agnostic controller until initial training data were collected, participants completed all activities with exoskeleton assistance turned off while sensor data were collected. To further increase the richness of our training data, we also hand-designed activity-specific controllers to collect actuated data for 46 out of the 66 total experimental conditions. Spline-based assistance was implemented as a function of gait phase when possible using eight- and six-node piecewise cubic Hermite interpolating polynomials at the hip and knee, respectively. Gait phase was estimated in real-time using the duration of the previous two strides measured from the pressure insole data12. The splines were shaped on the basis of the biological hip and knee moments reported in our previous work43. Furthermore, an impedance-based stance-swing state machine was used during several non-cyclic, impedance-like activities to control the device. Stance and swing phase were determined using the vertical GRF data measured from the pressure insoles. During leg swing, assistance was set to zero. During stance, the exoskeleton commanded torques using a spring-mode with a stiffness of 5 N m rad−1 and an equilibrium angle of 0°, which allowed all tasks to be completed without exceeding the actuator capabilities. These controllers were not intended to provide optimal assistance, but instead were used to collect a rich dataset for model training that included
actuated exoskeleton data as well as the unactuated data. Phase 2: training data with pilot model In phase 2, five new participants (four males, one female, age of 23.4 ± 4.9 years, height of 171.4 ± 7.4 cm and body mass of 68.0 ± 12.7 kg) participated in a single day protocol similar to phase 1. However, during phase 2 we deployed the task-agnostic controller on the exoskeleton using a preliminary joint moment estimator trained using the phase 1 dataset. Each participant completed all 28 tasks while the exoskeleton provided assistance, which generated a dataset closely representing the controller to be deployed in the final phase. Phase 3: online accuracy and user outcomes In phase 3, ten participants (seven males, three females, age of 21.8 ± 2.5 years, height of 174.8 ± 8.5 cm and body mass of 71.7 ± 10.2 kg) participated in a multi-day protocol consisting of three sessions each centred on a specific outcome metric: model accuracy, metabolic cost and lower-limb biological joint work. Model accuracy was assessed with R2 , r.m.s.e. and normalized MAE with respect to ground-truth total hip and knee moments31,32,60. Normalized MAE was normalized to the peak-to-peak range of the corresponding ground-truth joint moments per task. Metabolic cost and lower-limb biological joint work provided an indication of the exoskeleton's impact on user effort3,16,61–63, with metabolic cost more directly related to user energetics but only possible to measure and fairly compare during long bouts of repetitive tasks with consistent mechanical work requirements. During each session, the exoskeleton was controlled using the task-agnostic controller informed by the model trained on phase 1 and 2 data. Three of the participants were already enrolled in previous phases of the protocol. To ensure that the controller was evaluated on a participant-independent basis, we retrained separate models from random initialization for each of these participants while withholding the participant-specific data from the training set, ensuring that each participant was truly new to the network. The network accuracy session was performed using the same protocol as phase 2 in which the exoskeleton was powered on for all 28 activities. During each condition, motion capture and GRF data were collected to compute ground-truth joint moments, which were then used to evaluate model performance. Each participant completed this session first, which served as a training session for the participant to adapt to the exoskeleton assistance, a critical component to evaluating human–exoskeleton interactions64. Owing to an error when exporting the segmented trials, data for the run task from one participant was not included in this study. Additionally, three participants returned to perform eight truly new tasks while collecting the same data to compare estimates to ground-truth moments (Extended Data Table 2). During the lower-limb biological joint work session, participants completed six tasks detailed in Extended Data Table 3. These tasks were selected for our joint work analysis as measuring their resulting metabolic cost was infeasible owing to participant fatigue and time constraints. Each activity was repeated under three conditions: (1) wearing the exoskeleton with the task-agnostic controller providing assistance (exo on), (2) without wearing the exoskeleton (no exo) and (3) wearing the exoskeleton without actuation (zero torque). The order of these conditions was randomized; however, to minimize the time taken to complete the experimental protocol, the no exo condition was always placed either at the beginning or at the end, and all of the no exo activities were completed in succession, whereas the zero torque and exo on were alternated between each activity. Lower-limb positive biological joint work of the participant was computed by summing the average positive biological joint work from the hip, knee and ankle joints. Positive biological joint work for each joint was computed as the integral of the positive biological joint power for each joint, which was calculated as the product of the biological joint velocity and the biological joint moment (that is, the joint moment computed from ground-truth inverse dynamics after subtracting exoskeleton assistance torque). Results are presented as the average right leg positive joint work per repetition or stride. Owing to a bug in the exoskeleton data logger during the experimental protocol, stair ascent results for one participant and lunging results for one participant were not included in this study. Further, although this analysis quantified changes in the energetics of the biological joints, it did not account for muscle-level changes, such as muscle cocontraction. At the beginning of the metabolic cost session, each participant completed a habituation protocol to reacclimate to the device. The habituation protocol consisted of level-ground walking on the treadmill at 1.25 m s−1
while wearing the exoskeleton controlled with the task-agnostic controller. Exoskeleton assistance was sequentially ramped up in four evenly spaced increments every 2 min until the participant reached full torque assistance. The participant then walked at full assistance (20 and 15% of the estimated biological hip and knee moments, respectively) for 5 min to complete the habituation protocol. During each metabolic trial, oxygen intake V̇ ( O )2 and carbon dioxide exhaust V̇ ( CO )2 from each breath were measured using a metabolic measurement system (TrueOne 2400, ParvoMedics). User metabolic cost was computed from the $\dot{V}O_2$ and $\dot{V}CO_2$ measurements using the modified Brockway equation 11,65. These measurements were taken during four activities: level-ground walking at 1.25 m s-1, lifting a 25 lb weight, 5° incline walking at 1.25 m s-1 and running on level ground at 2.5 m s-1. For the lift weight trials, a metronome played a tone at 10 bpm and participants were instructed on each tone to use both hands to lift a 11 kg (25 lb) kettle bell off a shelf at waist height, touch the weight to the ground between their feet and then replace the weight on the shelf before returning to neutral standing and waiting for the next tone. For each activity, we tested the same three conditions as those of the joint work protocol: no exo, exo on and zero torque. Furthermore, the basal metabolic rate for each participant was measured from a 6 min standing trial while not wearing the exoskeleton. Owing to experimental time constraints, motion capture data were not collected during the metabolic experiments. The level-ground walking, lift weight and incline walking trials were each completed using a within-participant counter-balanced design (ABCCBA)12. Each condition lasted 6 min, and steady-state metabolic cost for each trial was computed as the average of the last 3 min of data11.66. Owing to the strenuous nature of running, especially during zero torque, the running conditions were only completed once (ABC) and lasted 3.5 min. Steady-state metabolic cost was then computed from the output of a first-order model fit to the running data11.66. The basal metabolic rate measured for each participant was subtracted from the steady-state metabolic cost of each condition to compute the user's net metabolic rate required to complete each activity. Within each activity, the order of conditions was pseudorandomized, with the no exo condition either in the A or C position to minimize don and doff time. Owing to improper calibration of the metabolic system, the metabolic data for one participant were omitted from this study. To further understand the effect of our approach on user metabolic cost during transient activities, we developed an extra protocol in which we recorded metabolic cost during a varying speed and incline treadmill circuit for three of the participants. Each 6 min trial involved walking at speeds varying from 0.6 to 1.5 m s-1 and running speeds of 2.0, 2.25 and 2.5 m s-1. The speed of the treadmill was changed according to the profile shown in Fig. 4b. Similarly, the treadmill incline varied between 0° and 15° throughout the trial (Fig. 4b). The speed and incline profile of the treadmill was designed to decrease walking speeds at higher inclines, to keep the participant within an acceptable aerobic respiration range for valid metabolic cost measurements. Similar to the outcomes testing described above, the participants completed a randomized ladder protocol consisting of no exo, exo on and zero torque conditions (ABCCBA12 as above) while wearing a metabolic measurement system (TrueOne 2400, ParvoMedics). Owing to an exoskeleton malfunction, one trial of data from one participant had to be removed; however, this trial was part of the C condition within the protocol, which should mitigate any adverse ordering effects from the removed datapoint. As described above, metabolic cost was computed using the Brockway equation and averaged across the last three minutes of the treadmill trial. This represents a consistent snapshot of metabolic cost across varying conditions. Whereas qualitative feedback from participants about exoskeleton assistance was not formally collected, it is important to note that some participants expressed that the knee extension assistance did not feel helpful during level-ground walking and occasionally felt uncomfortable. Otherwise, the overall response from participants about the exoskeleton assistance was quite positive across the trials used to evaluate metabolic cost and lower-limb biological joint work. Data availability All of the exoskeleton sensor data and time-synced ground-truth biomechanics used for training and evaluating our joint moment estimator have been released with this publication. The data and details are available at https://doi.org/10.35090/gatech/75759. The
training data includes 15 users performing all 66 conditions from the 28 task groups (Extended Data Fig. 2). The first 10 users include both unactuated data for all tasks and actuated data for those that could be mimicked with a heuristic controller. The following five users include actuated data for all tasks using a preliminary joint moment estimation model. The validation data includes ten users performing the same 66 conditions from the 28 task groups. The exoskeleton was actuated for all activities using the final model. Overall, this dataset includes more than 22 million labels of ground-truth moments across the left and right legs per lower-limb joint. Furthermore, the P values and corresponding participant count for all statistical tests are available in the Supplementary Data and a detailed description of our statistical tests is provided in the Supplementary Information. Code availability A Python package containing the code to deploy and test the models trained in this study using our corresponding dataset is available at https://doi.org/10.24433/CO.7641031.v2. Camargo, J., Ramanathan, A., Flanagan, W. & Young, A. A comprehensive, open-source dataset of lower limb biomechanics in multiple conditions of stairs, ramps, and level-ground ambulation and transitions. J. Biomech. 119, 110320 (2021). Reznick, E. et al. Lower-limb kinematics and kinetics during continuously varying human locomotion. Sci. Data 8, 282 (2021). Winter, D. A., Sidwall, H. G. & Hobson, D. A. Measurement and reduction of noise in kinematics of locomotion. J. Biomech. 7, 157–159 (1974). Molinaro, D. D., Park, E. O. & Young, A. J. Anticipation and delayed estimation of sagittal plane human hip moments using deep learning and a robotic hip exoskeleton. In Proc. 2023 IEEE International Conference on Robotics and Automation (ICRA) 12679–12685 (IEEE. 2023). Bai, S., Kolter, J. Z. & Koltun, V. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. Preprint at https://doi.org/10.48550/ arXiv.1803.01271 (2018). Golovin, D. et al. Google Vizier: a service for black-box optimization. In Proc. 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 1487–1495 (Association for Computing Machinery, 2017). Lim, H., Kim, B. & Park, S. Prediction of lower limb kinetics and kinematics during walking by a single IMU on the lower back using machine learning. Sensors 20, 130 (2020). Mundt, M. et al. A comparison of three neural network approaches for estimating joint angles and moments from inertial measurement units. Sensors 21, 4535 (2021). Ancillao, A., Tedesco, S., Barton, J. & O'Flynn, B. Indirect measurement of ground reaction forces and moments by means of wearable inertial sensors: a systematic review. Sensors 18, 2564 (2018). Forner-Cordero, A., Koopman, H. J. F. M. & van der Helm, F. C. T. Inverse dynamics calculations during gait with restricted ground reaction force information from pressure insoles. Gait Posture 23, 189–199 (2006). Nuckols, R. W. et al. Mechanics of walking and running up and downhill: a joint-level perspective to guide design of lower-limb exoskeletons. PLoS ONE 15, e0231996 (2020). Alexander, N., Strutzenberger, G., Ameshofer, L. M. & Schwameder, H. Lower limb joint work and joint work contribution during downhill and uphill walking at different inclinations. J. Biomech. 61, 75–80 (2017). Lenton, G. K. et al. Lower-limb joint work and power are modulated during load carriage based on load configuration and walking speed. J. Biomech. 83, 174–180 (2019). 64. Poggensee, K. L. & Collins, S. H. How adaptation, training, and customization contribute to benefits from exoskeleton assistance. Sci. Robot. 6, eabf1078 (2021). Brockway, J. M. Derivation of formulae used to calculate energy expenditure in man. Hum. Nutr. Clin. Nutr. 41, 463–471 (1987). Selinger, J. C. & Donelan, J. M. Estimating instantaneous energetic cost during non-steady-state
gait. J. Appl. Physiol. 117, 1406–1415 (2014). Acknowledgements We acknowledge E. Rouse and K. Zealand for their contribution to the overarching technical direction; and A. Azocar, A. Memo and R. Jackson for their contribution to experimental design and analysis. In addition, E. Lamers, T. Malko, A. Metzger, N. Hite, C. Muntz, K. Chen, B. Piercy, P. Franks, X. Qin, D. Tachibana and J. Cogan contributed to developing and maintaining the exoskeleton hardware, soft goods and software used in this study. This research was also supported in part through research cyberinfrastructure resources and services provided by the Partnership for an Advanced Computing Environment (PACE) at the Georgia Institute of Technology, Atlanta, GA, USA. This project was funded in part by X, The Moonshot Factory (A.J.Y.), in part by the Nation Science Foundation NRI grant no. 1830215 (A.J.Y.), in part by the National Science Foundation FRR grant no. 2233164 (A.J.Y.), and in part by the National Science Foundation Graduate Research Fellowship Program grant no. DGE-2039655 (D.D.M. and K.L.S.). Author contributions D.D.M. and K.L.S. contributed equally to the conceptualization, data acquisition, analysis, interpretation and writing of this work. E.B.S. contributed to the data acquisition, analysis and writing of this work. G.E. contributed to the conceptualization, analysis and writing of this work. M.K.S. contributed to the conceptualization, data acquisition, interpretation and writing of this work. A.J.Y. contributed to the conceptualization, interpretation and writing of this work. Competing interests D.D.M. and A.J.Y. are inventors on a patent filed with the US Patent Office (US 18/340,981) by the Georgia Institute of Technology that covers some of the methods for state estimation described in this work. X, The Moonshot Factory, funded this work, contributed to its conceptualization and methodological design, and contributed much of the exoskeleton hardware and software used in this study. Additional information Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41586-024-08157-7. Correspondence and requests for materials should be addressed to Dean D. Molinaro. Peer review information Nature thanks the anonymous reviewers for their contribution to the peer review of this work. Reprints and permissions information is available athttp://www.nature.com/reprints. Extended Data Fig. 1 \| Sensor ablations and dropouts for joint moment estimation. R2 for (a) hip and (b) knee joint moment estimation was compared across five conditions where different combinations of sensors were removed. R2 for (c) hip and (d) knee joint moment estimation was compared across six conditions where different individual sensor signals were set to zero to simulate sensor dropout. A one-way ANOVA was used to test for a main effect at the hip and knee, and for each joint, a post-hoc multiple comparisons test with a Bonferroni correction was conducted to evaluate comparisons between the All condition and each ablation or dropout condition. The bar height depicts the average performance across all 28 tasks and 10 participants. Error bars depict ±1 standard deviation across participants. The relative decrease in performance is depicted relative to the model trained on all sensors and without artificial sensor dropout. Asterisks indicate statistical significance (P < 0.05). Exact P values are provided in Supplementary Data 1. Extended Data Fig. 2 \| Task breakdown of our exoskeleton dataset labeled with ground-truth biomechanics. The 28 tasks were binned into three categories: cyclic, impedance-like, and unstructured based on their normative biomechanics. Additionally, many tasks contained multiple experimental conditions (66 conditions total). Extended Data Fig. 3 \| Forward task selection algorithm and results. (a) To determine the set of activities most important for model generalizability, tasks were iteratively added into the model training set, scored, and subsequently selected based on their contribution to overall model generalizability on the validation data. (b) The relative improvement in generalization for each selected task is shown. With each new task, the model was retrained and the relative improvement in its ability to predict biological hip and knee moments across all other (27) tasks was computed (average improvement shown). Saturated generalization was reached when the sum of the relative improvement reached 95% of the total sum from adding in all of the activities. Results were computed from leave-one-subject-out cross-fold validation using a 12-participant dataset; however, error bars were omitted for clarity. Extended Data Fig. 4 \| Detailed online joint moment estimation performance. The resulting (a) hip R2 , (b) knee
R2 , (c) hip RMSE, (d) knee RMSE, (e) hip normalized MAE, and (f) knee normalized MAE of our joint moment estimator is shown for each activity and is compared to the baseline method. The MAE results are normalized to their corresponding peak-to-peak range of the ground-truth joint moments. The bars depict the inter-subject mean across 10 subjects except for the Run condition where n = 9, and the error bars depict ±1 standard deviation. Asterisks indicate statistical significance (P < 0.05; exact P values are provided in Supplementary Data 1). Extended Data Fig. 5 \| Online joint moment estimation RMSE, normalized MAE, and analysis of under- and overestimation. (a) Our approach significantly reduced RMSE at the hip by 0.04 ± 0.02 Nm/kg (22 ± 10%, P = 0.03) and at the knee by 0.07 ± 0.02 Nm/kg (35 ± 8%, P = 3 × 10−8) compared to the baseline method during cyclic activities. For impedance-like tasks, our approach reduced RMSE by 0.11 ± 0.03 Nm/kg (35 ± 9%, P = 8 × 10−7) at the hip and by 0.14 ± 0.03 Nm/kg (47 ± 9%, P = 3 × 10−15) at the knee compared to the baseline method. (b) Estimator RMSE is shown per task for our approach and the baseline method. (c) Our approach significantly reduced normalized MAE at the hip by 22.2% ± 11.8% (P = 6 × 10−5) and by 34.1% ± 8.9% (P = 6 × 10−7) at the knee compared to the baseline method during cyclic activities. For impedance-like tasks, our approach reduced normalized MAE by 34.4% ± 8.0% (P = 3 × 10−9) at the hip and by 50.2% ± 14.2% (P = 8 × 10−14) at the knee compared to the baseline method. (d) Estimator normalized MAE is shown per task for our approach and the baseline method. (e) Histograms depicting the distribution of under- and overestimation of the hip and knee moments relative to groundtruth are shown. (f) Time series depicting the most severe example of underestimation and overestimation across all validation trials are shown (both of which occurred at the hip). In the box plots, black squares depict inter-subject mean; colored boxes depict the interquartile range; horizontal lines within boxes depict the inter-subject median; error bars depict the inter-subject minimum and maximum (n = 10). Asterisks indicate statistical significance (P < 0.05; exact P values are provided in Supplementary Data 1). In the scatter plots, each marker corresponds to the inter-subject average result for a single task (n = 10 expect for the Run condition where n = 9). Extended Data Table 1 \| TCN hyperparameter optimization \| Hyperparameter \| Stage 1 Search Space \| Stage 2 Search Space \| Selected Value \| \|------------------------------------------------------------------\|----------------------------------------------------------------------\|----------------------------------------------\|-------------------------\| \| # of Filters perLayer \| 8, 16, 24, 32, 48, 56, 64,72, 80 \| 56, 64, 72, 80 \| 80 \| \| # of ResidualBlocks (2ConvolutionalLayers per Block) \| 2, 3, 4, 5, 6 \| 5, 6 \| 5 \| \| Kernel Size \| 2, 3, 4, 5, 6, 7, 8, 9, 10 \| 2, 4, 5 \| 5 \| \| Activation Function \| ELU, GELU, ReLU ,Swish \| ELU, ReLU , Swish \| ReLU \| \| Learning Rate \| [1e-6, 1e-3] \| 5e-6, 1e-5, 5e-5 \| 5e-5 \| \| BlockNormalization \| Batch Normalization,Layer Normalization,Weight Normalization \| Batch Normalization,Weight Normalization \| WeightNormalization \| \| Dropout Type \| Element-Wise, Spatial \| Spatial \| Spatial \| \| Dropout Probability \| [0, 0.3] \| 0.15
\| 0.15 \| \| L1 KernelRegularization \| [1e-5, 1e-3] \| 0 \| 0 \| \| L2 KernelRegularization \| [1e-5, 1e-3] \| 0 \| 0 \| \| L2 Bias Regularization \| [1e-5, 1e-3] \| 0 \| 0 \| Values shown without brackets were optimized categorically. Arrays shown with closed brackets were optimized on the closed interval. Hyperparameter combinations resulting in an input sequence greater than 250 were omitted due to limitations in the simulated data at the bounds of each trial. In Stage 2, kernel and bias regularization were turned off since Stage 1 optimization generally minimized these values across multiple optimization instances. Extended Data Table 2 \| Activities for testing exoskeleton performance on entirely novel tasks \| Task \| Description \| \| \|----------------------------------\|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------\|--\| \| Burpees \| Participants were instructed to place their hands on the ground and step back into a push up posture. Then they stepped forward to return to standing. This was performed 3 times leading with the right leg and 3 with the left. \| \| \| Cart Walk \| Participants were instructed to push a stationary rod (simulating a heavy cart) while walking on the treadmill at 0.8 m/s for 20 seconds. They then performed the same task while pulling on the rod and walking backward. \| \| \| Crouch Walk \| Participants were instructed to walk on the treadmill at 0.8 m/s while in a crouched posture for 20 seconds. \| \| \| Dizzy Walk \| Participants spun around for 10 seconds and then walked back and forth across the force plates until no longer showing signs of dizziness. \| \| \| Steep Incline/Decline Walk (15°) \| Participants were instructed to walk up a 15° incline at 1.2 m/s for 20 seconds. They then repeated this while walking down the same decline. \| \| \| Layup \| Participants were instructed to run up to the force plates, plant a single foot, and jump to maximum vertical height while landing with both feet on the force plates. This was performed 3 times for both right and left legs. \| \| \| Mountain Climbers \| Participants were instructed to assume a push-up posture and then alternate bringing each knee up toward their chest as quickly as possible. This was performed for 20 seconds. \| \| \| Split Walk \| Participants walked on a treadmill with the right leg belt at 1.6 m/s and the left leg belt at 0.8 m/s for 20 seconds. They then repeated this task with the left belt at 1.6 m/s and the right at 0.8 m/s. \| \| Extended Data Table 3 \| Activities for testing exoskeleton effects on lower-limb biological joint work \| Task \| Speed \| Description \| \|---------------------\|----------------------------\|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------\| \| Lift & Place Weight \| 12 lifts per min \| Participants were instructed to lift a 25lb weighted bag with both hands. On the first tone, participants started from standing and lifted the bag to waist height. They then paused for the second tone, at which they set the bag back on the floor and then returned to standing. \| \| Squat \| 15 squats per minute \| Participants were instructed to squat with the 25lb weight until they barely touched a pillow resting on a low stool. Then they returned to standing. \| \| Lunge \| 15 lunges per minute \| Participants were instructed to lunge with their right leg forward until their left knee barely touched a pillow laid on the ground. Then they returned to standing. \| \| Sit & Stand \| 12 sit & stands per minute \| Participants were instructed toalternate sitting down andstanding up on the tone withoutusing their arms for assistance. \| \| Right Leg Step Up \| 12 step ups per min \| Participants were instructed to step up onto a 46 cm box with their right leg. On the first tone, participants stepped up with their right leg and paused standing on one foot. On the second tone they stepped back down to have both feet on the ground. \| \| Stair Ascent \| Self-selected speed \| Participants were instructed to complete four bouts of walking up a six-step staircase at their natural walking pace. \|
Human-in-the-Loop Optimization of Hip Exoskeleton Assistance During Stair Climbing Dongho Park , Jimin An, Dawit Lee , Inseung Kang , Member, IEEE, and Aaron J. Young , Senior Member, IEEE Abstract—Objective: This study applies human-in-theloop optimization to identify optimal hip exoskeleton assistance patterns for stair climbing. Methods: Ten participants underwent optimization to individualize hip flexion and extension assistance, followed by a validation comparing optimized assistance (OPT) to biological hip momentbased assistance (BIO), no assistance (No-Assist), and no exoskeleton (No-Exo) conditions. Results: OPT reduced metabolic cost by 4.5% compared to No-Exo, 11.44% compared to No-Assist, and 5.02% compared to BIO, demonstrating the effectiveness of the optimization approach. Statistical analysis revealed distinct characteristics in optimal assistance timing and magnitude that deviated systematically from biological hip moment patterns. Compared to BIO, OPT exhibited later peak flexion timing (76.4 $\pm$ 3.7% vs 65.0%), shorter flexion duration (29.2 $\pm$ 3.6% vs 40.0%), later peak extension timing (26.7 $\pm$ 3.8% vs 20.0% of gait cycle), and higher peak flexion magnitude (11.1 $\pm$ 1.5 Nm vs 10.0 Nm). While individual optimal assistance profiles varied across participants, comparison between individually optimized parameters and the best subjectindependent parameters identified through post-hoc analysis showed consistency. On average, metabolic rate convergence was achieved after 18 iterations, while most exoskeleton control parameters did not reach our convergence criteria within 20 iterations. Conclusion: These findings demonstrate that human-in-the-loop optimization can effectively identify task-specific assistance patterns for stair climbing, while the consistency between individual and subject-independent parameters suggests the potential for developing generalized assistance strategies. The systematic differences between optimized and biological Received 30 June 2024; revised 13 November 2024 and 9 January 2025; accepted 17 January 2025. Date of publication 30 January 2025; date of current version 27 June 2025. This work was supported in part by the NSF FRR under Award no. 2328051, in part by NSF NRI under Grant 1830215, and in part by the NIH Director's New Innovator under Award no. DP2-HD111709. (Corresponding author: Dongho Park.) Dongho Park is with the Woodruff School of Mechanical Engineering and the Institute for Robotics and Intelligent Machines, Georgia Institute of Technology, Atlanta, GA 30332 USA (e-mail: dpark@gatech.edu). Jimin An is with the Department of Mechanical Engineering, Carnegie Mellon University, USA. Dawit Lee was with the Woodruff School of Mechanical Engineering, Georgia Institute of Technology, USA. He is now with the Department of Bioengineering, Stanford University, USA. Inseung Kang is with the Department of Mechanical Engineering, Carnegie Mellon University, USA. Aaron J. Young is with the Woodruff School of Mechanical Engineering and the Institute for Robotics and Intelligent Machines, Georgia Institute of Technology, USA. Digital Object Identifier 10.1109/TBME.2025.3536516 moment-based assistance underscore the fundamental distinctions between biological torque-based control and optimal control for exoskeleton assistance. Index Terms—Bayesian optimization, hip exoskeleton, human-in-the-loop optimization, metabolic cost, personalized assistance, stair climbing. I. INTRODUCTION TAIR climbing is a biomechanically demanding task that requires higher joint moment and power compared to level walking [1], [2]. This increased demand poses significant challenges for older adults and individuals with mobility impairments, leading to increased energy expenditure, fatigue, higher fall risk, and reduced independence [3], [4]. Powered exoskeletons have emerged as a promising system to assist stair climbing by reducing biological joint works and metabolic costs [5], [6]. However, a critical knowledge gap exists in understanding optimal assistance strategies specific to stair ascent biomechanics. Previous studies on stair-climbing exoskeletons have focused primarily on hardware development, biomechanical evaluation, and energetic effects, but the systematic optimizaiton of control parameters remains largely unexplored [7], [8], [9]. Recent advances in human-in-the-loop optimization (HILO) have demonstrated promising results in identifying optimal exoskeleton assistance for level walking, running, and incline walking, but have never been extended to stair ambulation [10], [11], [12]. By directly incorporating real-time human feedback, HILO enables the development of personalized assistance without relying on explicit models, enabling adaptive solutions tailored to individual users and specific locomotor tasks. These approaches underscore the importance of empirical optimization in overcoming the inherent complexities of human-robot interaction [13]. One of its key advantages is the ability to handle multiple interacting
control parameters efficiently. In contrast to traditional parameter tuning approaches, such as grid search or sequential optimization, which becomes impractical as the number of parameters increases, HILO leverages statistical models and optimization algorithms to search the high-dimensional parameter space efficiently [14]. For instance, Bayesian optimization, a popular HILO approach, constructs a probabilistic surrogate model of the objective function such as metabolic cost, and uses an acquisition function to balance the exploration and exploitation of the parameter space [14]. This approach allows 0018-9294 © 2025 IEEE. All rights reserved, including rights for text and data mining, and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. for the efficient identification of optimal control parameters with a limited number of experimental trials, making it well-suited for human-in-the-loop applications [15]. For example, Zhang et al. used HILO to optimize the assistance profile of a soft exosuit during level walking, resulting in a 17.4% reduction in metabolic cost compared to walking without the exosuit [11]. Similarly, Ding et al. applied HILO to optimize the hip extension assistance of a soft exosuit, achieving a 22.83% reduction in metabolic cost compared to walking with the exosuit unpowered [10]. However, the effectiveness of assistance strategies optimized for level walking may not translate directly to stair climbing. Previous work in level walking has shown that human response to exoskeleton assistance can be counterintuitive, with bio-inspired approaches not always yielding optimal results [16]. This complexity arises from humans' adaptation to novel mechanical interventions, where subtle differences in assistance can lead to substantially different responses [16]. Indeed, recent studies have confirmed that humans are incredibly complex and respond to robotic devices in ways that cannot be modeled or predicted with sufficient accuracy [13]. This unpredictability in human response becomes particularly relevant for stair climbing, where biomechanical demands and muscle activation patterns differ substantially from level walking [1], [2], [17]. The effectiveness of assistance patterns optimized for level walking may not translate directly to stair climbing, and biological joint moments during stair ascent may not represent the optimal assistance strategy. This highlights the importance of task-specific optimization to identify truly effective assistance patterns. This study aims to bridge this gap by applying HILO to optimize hip exoskeleton assistance for stair climbing. The primary objectives of this study are to determine the optimal assistance patterns for stair climbing biomechanics and to quantify their effects on metabolic cost. We hypothesize that the optimized assistance profiles will exhibit distinct timing and magnitude characteristics that deviate from both level walking assistance patterns and biological joint moments during stair climbing. A systematic investigation of these assistance parameters will provide crucial insights into how optimal exoskeleton assistance should be structured for the specific demands of stair climbing. We further hypothesize that the optimized patterns will reduce metabolic costs compared to no assistance and no exoskeleton conditions. This reduction in metabolic cost is anticipated to enhance the efficacy of the exoskeleton during stair ascent, highlighting the potential of personalized assistive strategies in improving locomotion efficiency during stair climbing. Through comparison of optimized assistance profiles with biological hip moment patterns, this study will characterize the fundamental differences between biological torque-based control and optimal control for exoskeleton assistance during stair climbing, establishing design principles for future assistance strategies. II. POWERED HIP EXOSKELETON DESIGN A. Mechatronic Design The custom hip exoskeleton used in this study is an updated version of our previous design [18]. The bilateral exoskeleton is designed to assist both hip flexion and extension during locomotion (Fig. 1). Each hip joint is actuated by a 9:1 planetary gearhead integrated electromechanical actuator (AK80-9, T-Motor, China), which provides a rated torque of 9 Nm and a peak torque of 18 Nm. The actuator demonstrated tracking errors within ± 0.4 Nm during static testing and less than 2 Nm of dynamic backdrive torque during human-like motions [19]. To minimize the relative movement between the user and the exoskeleton interfaces at high torques, the peak torque is software-limited to 15 Nm, which was determined to be the optimal range for maintaining a secure fit during operation. The upgraded exoskeleton features a more lightweight design compared to its predecessor. By optimizing carbon fiber pelvic interfaces and backplate, and removing unnecessary material, the total weight of the exoskeleton has been reduced from 4.8 kg to 3.6 kg. This weight reduction includes the 24 V LiPo battery (Venom Fly, Venom Power, USA) powering the device and the microprocessor (MyRio, National Instruments, USA) executing the device operation and the controller. The exoskeleton covers a range of motion of 100 degrees in flexion and 40 degrees in extension, supporting a wide range of natural movements without hindering the user's mobility. The actu
ators are connected to lightweight, L-shaped carbon fiber pelvic interfaces, ensuring a secure fit while maintaining alignment between the actuator axes and the user's hip joint centers during use. The pelvis interfaces are designed to encompass the lateral sides and back of the user's pelvis. The back of the interfaces connects to the redesigned carbon fiber backplate, creating a structure that securely encircles the entire circumference of the pelvis. The backplate features multiple holes for pelvis width adjustment, allowing easy manual adjustment of the overall pelvis width using bolts and butterfly nuts to alter the connection points between the pelvic interfaces and the backplate. At the front of the pelvis interfaces, ratchet-equipped belts have been added to allow fine-tuning of the fit to accommodate various pelvic sizes and ensure a secure connection. To improve user comfort and wearability, the tensioning belts that secure the backplate and pelvis interface to the user's torso have been upgraded. The new belts feature velcro fastenings and dual pull tabs, making it easy for users to achieve an effective level of compression. Torque transmission is achieved through carbon fiber thigh interfaces that wrap around the user's thighs, with additional flexible belts securing the thigh and interface together. The electronic systems of the upgraded exoskeleton remain largely unchanged from the previous design [18]. The overall exoskeleton controller operates at 200 Hz, which is housed in a backpack designed for running to maximize user comfort and system reliability. The actuators communicate via a CAN bus network, employing a closed-loop current control strategy to accurately apply torque. A bespoke circuit, specifically engineered to connect SPI and CAN communications, is integrated into a custom-developed PCB. The exoskeleton also features Force Sensitive Resistors (FSR) located at the heel and toe positions of each foot to accurately detect foot contact, which are used to segment gait cycles. Fig. 1. Hip exoskeleton design. (a) CAD rendering of the exoskeleton, showcasing its lightweight and adjustable carbon fiber and aluminum frame, brushless DC motor actuators with a 9:1 gear ratio integrated into the frame, L-shaped pelvic interfaces for effective load transfer, and thigh cuffs for secure attachment and efficient torque transmission. (b) A user wearing the hip exoskeleton during a stair climbing trial, demonstrating the device's compact and unobtrusive design. The control unit and battery are housed in a running backpack to optimize weight distribution and minimize interference with natural movement patterns. B. Controller Design The high-level controller performs real-time gait phase estimation for each leg based on foot contact events detected by the FSRs, attached to the bottom of each foot. The current gait phase, in terms of percentage, is estimated by dividing the time passed from the most recent foot contact of the ipsilateral leg by the estimated stride duration, which is computed by averaging the stride durations for the most recent five strides of the ipsilateral leg [20]. The estimated gait phase serves as an input to the mid-level controller, which is responsible for generating the desired assistive torque profiles. In this study, we compared two different assistance profiles: the optimized assistance profile (OPT) and the biological hip moment-based assistance profile (BIO). The OPT profile was individualized for each participant using the human-in-the-loop optimization, while the BIO profile was designed to mimic the biological hip joint moment during stair climbing based on previous studies [1], [2], [17]. Following these biomechanical data, we implemented the BIO profile as hip extension assistance with onset at 0%, peak timing at 20%, and duration of 40% of the gait cycle, and hip flexion assistance with onset at 45%, peak timing at 65%, and duration of 40% of the gait cycle. For the OPT profile, the assistance was based on five parameters with symmetric profiles to balance optimization complexity and experimental feasibility (see Fig. 2). The control parameters are defined as follows: tf lexp , tf lexd : peak timing and duration of flexion assistance, textp , textd : peak timing and duration of extension assistance, uf lex: peak flexion assistance magnitude (Nm), uext: peak extension assistance magnitude (Nm), set to the maximum value of 15 Nm. The same assistance profile was applied for both legs and was controlled for each leg independently. Based on the biomechanical analysis of stair climbing in previous studies, the hip extension moment during stair ascent is significantly higher compared to level walking [21]. To provide adequate assistance during the demanding extension phase, the peak extension torque, uext, is set to the maximum value of 15 Nm, which is the hardware limit of the exoskeleton actuator. In previous studies on human-in-the-loop optimization, the optimal extension torques for tasks with lower biomechanical demands than stair climbing were found to be significantly higher than the maximum torque capacity of the exoskeleton used in this study [10]. For instance, the optimal hip extension torque during ramp ascent walking was reported to be approximately 1.2 Nm/kg, which translates to about 90 Nm for a 75 kg adult [10]. Based on these
findings, it can be inferred that the optimal extension torque for a more demanding task like stair climbing would likely exceed the maximum torque capacity of 15 Nm of the exoskeleton used in this study. Therefore, to ensure that the exoskeleton can meet the requirements of stair climbing and maximize the reduction in user effort and metabolic cost, we set the peak extension torque to the hardware limit of 15 Nm. Fig. 2. Experimental protocol of the human-in-the-loop optimization during stair climbing. The study consisted of two sessions: an optimization session and a validation session. In the optimization session, participants completed 20 iterations of 2.6-minute stair climbing trials with the exoskeleton, while a Bayesian optimization algorithm personalized the assistance profile based on the measured metabolic cost. The optimized assistance profile was then validated against three other conditions (BIO, No-Assist, No-Exo) in the validation session. The assistance profile was characterized by four timing parameters ( $t_{flex_p}$ , $t_{flex_d}$ , $t_{ext_p}$ , $t_{ext_d}$ ) and one peak flexion torque magnitude parameter ( $u_{flex}$ ), with peak extension torque ( $u_{ext}$ ) set to a maximum of 15 Nm based on hardware limitations and biomechanical considerations. The desired torque profile, T(x), is computed as a function of the gait phase percentage, x, using a piecewise cubic Hermite interpolating polynomial (PCHIP) function [22]: $$T(x) = PCHIP(t, u, x) \tag{1}$$ Where $t = [t_{flex_p} - t_{flex_d} \cdot 2^{-1}, t_{flex_p}, t_{flex_p} + t_{flex_d} \cdot 2^{-1}, t_{ext_p} - t_{ext_d} \cdot 2^{-1}, t_{ext_p} + t_{ext_d} \cdot 2^{-1}]$ and $u = [0, u_{flex}, 0, 0, u_{ext}, 0]$ represent the vectors of timing and magnitude parameters, respectively. By integrating gait phase estimation, and symmetric torque profile generation, the proposed control strategy enables the optimization of assistance profiles based on individual user feedback while ensuring consistent and accurate torque delivery. III. EXPERIMENTAL PROTOCOL Ten able-bodied young adults participated in this study after providing written informed consent. The study protocol was approved by the Institutional Review Board of the Georgia Institute of Technology (H22051, 04/27/2022). The study consisted of two sessions: an optimization session and a validation session. These sessions were separated by at least 24 hours to avoid fatigue effects. In the optimization session, participants were introduced to the hip exoskeleton, and the device was fitted to each participant's body shape. Participants then underwent a brief training session to become accustomed to walking on the stair treadmill (StepMill3, Stairmaster, USA). The staircase consisted of a speed-adjustable flight with a step height of 15 cm, a step depth of 23 cm, and a step width of 43 cm. To further acclimate participants to the stair treadmill walking, a 5-minute warmup session without the exoskeleton was initially conducted at the speed of 74.8 steps/minute, which was determined from preliminary experiments as the average natural stair-walking speed for three able-bodied adults. Following the warmup, participants were given a 5-minute rest period before commencing the main optimization session involving stair walking with the exoskeleton. The optimization protocol was specifically adapted for the demanding nature of stair climbing. Through pilot experiment with three subjects, we determined that 2.6 minutes of data collection per iteration could maintain estimation accuracy within 2% while allowing necessary rest periods between iterations (Fig. 3). Based on pilot testing feedback, where subjects consistently reported fatigue with more than 20 stair ascent iterations, we carefully balanced measurement requirements with participant fatigue management. The optimization session consisted of 20 iterations, each lasting 2.6 minutes. This iteration time was determined based on the metabolic data collected from three able-bodied participants walking on the stair treadmill, which indicated that 2.6 minutes of metabolic data collection per iteration would yield accurate estimations within 2% error (Fig. 3). Participants walked on the stair treadmill in each iteration while wearing the hip exoskeleton and a respiratory measurement device (TrueOne 2400, ParvoMedics, USA). Due to the high physical demand of stair climbing, a 3-minute rest period was provided between iterations to ensure continuous performance across multiple iterations. Fig. 3. Metabolic cost estimation error as a function of the duration of metabolic data used for estimation. A preliminary experiment was conducted with three participants walking on the stair treadmill for 6 minutes. The average of the last 2 minutes of metabolic data was considered as the ground
truth. The metabolic cost was then estimated using data from the beginning of the trial with varying durations. The results show that using 2.6 minutes of data yields an estimation error of less than 2%, providing a balance between accuracy and experimental feasibility for the optimization protocol. On the second day, participants returned for the validation session, which aimed to verify the efficacy of the optimized exoskeleton assistance identified during the optimization session. Participants compared four conditions: 1) stair climbing with optimized exoskeleton assistance (OPT), 2) stair climbing with assistance mimicking biological hip joint moment (BIO), 3) stair climbing with the exoskeleton worn but without assistance (No-Assist), and 4) stair climbing without wearing the exoskeleton (No-Exo). Each condition lasted 5 minutes, and the order of the conditions was randomized for each participant to mitigate the influences of measurement noise and trial order through a double-reversal validation test (ex. ABCD-DCBA). Additionally, a 5-minute quiet standing trial was conducted at the end of the protocol to assess the resting metabolic rate. For each trial in which V O˙ 2 and V CO ˙ 2 data were collected, the gross metabolic cost (Pmet) scaled by body mass (m) was computed [23]: $$P_{met} = \frac{(0.278 \times \dot{V}O_2 + 0.075 \times \dot{V}CO_2)}{m}$$ (2) where V O˙ 2 is the volumetric flow rate of oxygen intake, and V CO ˙ 2 is the volumetric flow rate of carbon dioxide exhaust. The calculated gross metabolic cost values were then used to determine the net metabolic cost of walking for each condition by subtracting the participant's resting metabolic rate, which was measured during the quiet standing trial. IV. HUMAN-IN-THE-LOOP OPTIMIZATION We employed a human-in-the-loop optimization approach based on Bayesian optimization to identify the optimal exoskeleton assistance parameters that minimize the user's metabolic cost during stair climbing. Bayesian optimization is a sampleefficient global optimization method that is well-suited for expensive-to-evaluate objective functions, making it particularly appropriate for human-in-the-loop settings where each evaluation requires a significant amount of time and effort from the user [14]. One of the key parameters in Bayesian optimization is the balance between exploration and exploitation. Exploration refers to sampling from regions of high uncertainty in the parameter space, while exploitation focuses on sampling from regions with promising expected performance. In our study, we set the exploration-exploitation balance to 0.5, which means that the algorithm spends equal time exploring and exploiting. This balanced approach ensures that the algorithm efficiently navigates the high-dimensional parameter space while avoiding premature convergence to suboptimal solutions [24]. Another important parameter in Bayesian optimization is the acquisition function, which determines the next set of parameters to evaluate. We used the Expected Improvement (EI) acquisition function, as it is known to perform well in practice and has been widely used in previous studies [25]. The Gaussian process model used in Bayesian optimization is characterized by a kernel function that defines the similarity between data points. In our study, we used the Automatic Relevance Determination (ARD) Matérn 5/2 kernel, which is a popular choice for modeling smooth functions [26]. The ARD Matérn 5/2 kernel allows for different length scales along each input dimension, enabling the algorithm to identify the most relevant parameters for optimization. We also set the number of initial points for the Bayesian optimization algorithm to 10, which means that the algorithm randomly samples 10 points from the parameter space before starting the optimization process. These initial points help the algorithm to build an initial surrogate model of the objective function. The optimization problem is formulated as follows: $$Minimize_x f(x)$$ (3) Where x is the vector of exoskeleton assistance parameters and f(x) is the objective function representing the user's metabolic cost. The exoskeleton assistance parameters to be optimized include the peak timing and duration of hip flexion and extension assistance (tf lexp , tf lexd , textp , textd ) as well as the peak flexion torque magnitude (uf lex). Based on pilot testing with three subjects, we constrained flexion peak timing between 65–92.5%, flexion duration between 20–47.5%, extension peak timing between 15–45%, and extension duration between 20–60% of the gait cycle, while flexion magnitude was limited to 6–13 Nm. The peak extension torque was set to a fixed value of 15 Nm, as determined from the biomechanical analysis of stair climbing in previous studies [1]. Bayesian optimization constructed a probabilistic surrogate model of the objective function using a Gaussian process (GP) [26]. The GP model provided a probabilistic forecast of the objective function's performance across the parameter space, characterized by a mean function μ(x) and a
covariance function k(x, x-), where x and xare parameter settings. The surrogate model was iteratively updated based on observations from user trials, guiding the selection of new parameter values to evaluate. The EI acquisition function at any point x can be expressed as: $$EI(x) = \mathbb{E}[\max(f(x_{\text{best}}) - f(x), 0)] \tag{4}$$ where f(xbest) was the best-observed value of the objective function so far. In each iteration of the optimization process, the algorithm selected the next set of assistance parameters to evaluate by maximizing the acquisition function. The user then walked on the stair treadmill wearing the exoskeleton with the corresponding assistance profiles, and the metabolic cost was measured using an indirect calorimetry system. The measured metabolic cost was used to update the GP model, refining the surrogate model of the objective function [14], [25]. After the optimization process was completed, the optimal assistance parameters were identified and validated in a separate session to assess their effectiveness in reducing the metabolic cost of stair climbing. To compare the effectiveness of personalized optimization with a generalized approach, we conducted a post-hoc analysis to determine a best subject-independent parameters. Rather than simply averaging individual optimal points, we created an average metabolic response surface by combining data from all subjects except the test subject (leave-one-out approach) and found the global minimum using constrained optimization. This methodology provides valuable insights into the best possible control policy when subject-specific data is unavailable, which could be particularly useful in two scenarios: (1) when individual optimization is not feasible or desired, and (2) as an initial seed point for future human-in-the-loop optimization processes to potentially reduce the required number of iterations. This approach is more rigorous than simple averaging as it avoids using a subject's own data in determining their best subject-independent parameters, thereby reducing potential bias in the analysis. We defined metabolic rate convergence in our Bayesian optimization process as the point at which three consecutive iterations exhibited a change of 0.5% or less, representing a more conservative approach compared to previous studies' typical 4% criterion [10]. For exoskeleton control parameters, convergence was defined as a change of 5% or less between iterations. These thresholds were chosen to ensure high precision in identifying optimal hip exoskeleton control parameters while allowing for the complexity of simultaneous multi-parameter optimization. V. METABOLIC RATE ESTIMATION In the pursuit of optimizing exoskeleton assistance for stair climbing tasks, our study aims to adapt a methodology that has been previously applied to level walking, wherein metabolic rate estimation plays a critical role. Previous studies have utilized short-duration metabolic data, specifically 2-minute measurement intervals, to estimate the metabolic rate during level walking tasks [10], [11]. This approach is viable due to the stability of metabolic responses after an initial warm-up period, allowing for consecutive iterations of measurement without significant variance in metabolic rate, thus ensuring reliable predictions from truncated data samples. However, the application of a similar methodology to stairclimbing tasks presents unique challenges. Stair climbing is inherently a more strenuous task, and as such, it is physically challenging for subjects to perform multiple successive iterations without rest. Consequently, each iteration in our stair climbing protocol is followed by a 3-minute rest period, which inherently disrupts the metabolic plateau achieved during warmup, potentially affecting the stability of subsequent metabolic rate estimations. To address the challenge, we conducted a preliminary experiment to determine the optimal duration of metabolic data collection per iteration that would yield accurate estimations within a reasonable error margin. We recruited three healthy adults and measured their respiratory data while they walked on a stair treadmill at a fixed pace of 74.8 steps/minute for 6 minutes. The average of the final 2 minutes of the collected metabolic data was designated as the ground truth. We then compared this against estimates derived from data samples starting from the beginning of the exercise and extending for various durations. The metabolic cost estimation is derived by fitting a first-order dynamic model to the transient data. The model is represented mathematically in the frequency domain as: $$Y(s) = H(s)X(s) \tag{5}$$ with the transfer function H(s) defined as: $$H(s) = \frac{1}{\tau s + 1} \tag{6}$$ Here, X(s) denotes the instantaneous metabolic cost (the input data), and Y (s)represents the estimated metabolic cost (output). The time constant τ is determined to be 42 seconds, based on prior research findings [11]. This first-order dynamic model allows us to estimate metabolic cost from only 2.6 minutes of data, thereby minimizing the overall time investment for the experiment protocol and reducing subject fatigue without sacrificing the accuracy of our metabolic cost estimations. It revealed that utilizing 2.6 minutes of metabolic data for estimation purposes results in an average error rate within 2% (Fig. 3). This finding is pivotal for our stair climbing optimization protocol as it balances the
need for accurate metabolic rate estimation against the practical limitations imposed by the task's physical demands. The adaptation of short-duration metabolic rate estimation methods to stair climbing optimization is feasible with an adjusted measurement interval. By extending the estimation window to 2.6 minutes, we can maintain an acceptable error rate, thereby ensuring the integrity of our optimization process while respecting the physiological constraints of our subjects. VI. RESULTS A. Metabolic Cost Reduction The optimized hip exoskeleton assistance significantly reduced the metabolic cost of stair climbing compared to the No-Exo, No-Assist, and BIO conditions (Fig. 4). The average Fig. 4. Net metabolic cost for the four conditions tested in the validation session. Optimized assistance (OPT) significantly reduced metabolic cost compared to the No-Exo (4.5%, p < 0.05), No-Assist (11.44%, p < 0.01), and Biological hip moment-based assistance(BIO) (5.02%, p < 0.01) conditions. BIO also reduced metabolic cost compared to No-Assist (6.75%, p < 0.01) but was less effective than OPT. Error bars represent the standard deviation of the mean. Asterisks denote statistically significant differences (* p < 0.05, ** p < 0.01). metabolic cost in the OPT condition was 4.5% (p < 0.05) lower than the No-Exo, 11.44% (p < 0.01) lower than the No-Assist, and 5.02% lower than the BIO. The BIO also resulted in a 0.6% increase in metabolic cost compared to the No-Exo (p < 0.01). However, it still achieved a 6.81% reduction compared to the No-Assist (p < 0.01). B. Optimal Assistance Profiles The optimized assistance profiles varied across participants, highlighting the importance of individualized optimization (Fig. 5). The best subject-independent peak flexion torque was $10.3 \pm 2.1$ Nm. The peak flexion timing was $75.9 \pm 4.2\%$ , and the peak extension timing was $25.1 \pm 3.8\%$ of the gait cycle. The flexion duration was $29.0 \pm 3.5\%$ , and the extension duration was $41.2 \pm 7.3\%$ of the gait cycle. Comparison between optimized and bio-inspired assistance profiles revealed significant timing and magnitude differences (Fig. 5). The optimized profiles exhibited significantly later peak flexion timing (76.4 $\pm$ 3.7% vs 65.0% of gait cycle, p < 0.01), shorter flexion duration $(29.2 \pm 3.6\% \text{ vs } 40.0\%, p < 0.01)$ , later peak extension timing $(26.7 \pm 3.8\% \text{ vs } 20.0\%, p < 0.01)$ , and higher peak flexion magnitude (11.1 $\pm$ 1.5 Nm vs 10.0 Nm, p < 0.05) compared to the bio-inspired assistance profile. Extension duration showed no significant difference between conditions (43.8 $\pm$ 7.9% vs 40.0%, p = 0.159). Statistical comparison between individual optimal parameters and subject-independent parameters showed no significant differences across all parameters. Individual optimal parameters versus subject-independent parameters were: flexion timing (76.4 $\pm$ 3.5% vs 75.9%, p = 0.690), flexion duration (29.2 $\pm$ 3.4% vs 29.0%, p = 0.857), extension timing $(26.7 \pm 3.6\% \text{ vs } 25.1\%, p = 0.211)$ , extension duration $(43.8 \pm 3.6\% \text{ vs } 25.1\%, p = 0.211)$ Fig. 5. Comparison of optimized assistance profiles (OPT) and the biological hip moment-based profile (BIO). Orange bold line: Best subject-independent parameters across all participants. Blue bold line: The biological hip moment-based profile used in the BIO condition. On average, OPT exhibited later peak flexion timing, shorter flexion duration, later peak extension timing, and similar extension duration compared to BIO. 7.5% vs 41.2%, p = 0.315) and peak flexion magnitude (11.1 $\pm$ 1.4 Nm vs 10.3 Nm, p = 0.
112). C. Bayesian Optimization Convergence Analysis The Bayesian optimization process for hip exoskeleton control parameters ran for 20 iterations. Metabolic rate convergence, based on our predefined criterion of three consecutive iterations showing a change of 0.5% or less, was achieved after 18 iterations (Fig. 6). Regarding the exoskeleton control parameters, most did not achieve the 5% convergence criterion within the 20 iterations. However, the 'peak flexion timing' parameter showed signs of convergence, with its values stabilizing after 12 iterations (Fig. 6). The optimization process identified control parameters that resulted in a 4.54% reduction in metabolic cost compared to the No-Exo condition, despite not all parameters reaching strict convergence. VII. DISCUSSION This study demonstrates the effectiveness of human-in-the-loop optimization in identifying individualized hip exoskeleton assistance profiles that significantly reduce metabolic rate during stair climbing. The optimized assistance profiles (OPT) resulted in a 4.5% reduction compared to the No-Exo (p < 0.05), an 11.44% reduction compared to the No-Assist (p < 0.01), and a 5.02% reduction compared to the BIO (p < 0.01). The metabolic cost reduction achieved with the optimized assistance (4.54% compared to No-Exo) presents an interesting comparison with both level walking and stair ascent studies [10], [11], [9], [27]. For instance, a previous study reported a 17.4% reduction in metabolic cost with a tethered soft exosuit providing optimized hip extension assistance during level walking [10]. Fig. 6. Convergence analysis of the Bayesian optimization process over 20 iterations. Top five rows: Evolution of exoskeleton control parameters - peak extension timing, extension duration, peak flexion timing, flexion duration, and peak flexion torque. The 'peak flexion timing' parameter showed signs of convergence, stabilizing after approximately 12 iterations. Bottom row: Metabolic cost reduction, demonstrating convergence after 18 iterations based on the 0.5% change criterion for three consecutive iterations. Red dashed lines indicate the 5% convergence threshold for control parameters and the 0.5% threshold for metabolic cost. The difference in metabolic cost reduction between our study and the previous study can be attributed to several factors. First, stair climbing is a more demanding task compared to level walking, which may limit the potential for metabolic cost reduction. Second, the exoskeleton system used in our study is a mobile device with a maximum output torque of 15 Nm, whereas the tethered actuation system used by the previous study can provide up to 3 N of force per kg of user's body weight [10]. Assuming a user weight of 70 kg, their system can generate a maximum force of 210 N, substantially higher than the 15 Nm (approximately 31.5 N for a 0.15 m moment arm) provided by our mobile exoskeleton. The higher assistive forces in the previous study may have contributed to the greater metabolic cost reduction observed. Despite the lower metabolic cost reduction compared to some level walking studies, the results of our study are still significant, considering the challenging nature of stair climbing and the limitations of a mobile exoskeleton system. The OPT resulted in a significant reduction in metabolic cost compared to No-Exo and No-Assist, highlighting the effectiveness of the human-in-the-loop optimization approach in personalizing assistance profiles for stair climbing. Our metabolic reduction results also provide interesting insights when compared to previous stair-climbing studies. Kim et al. achieved higher metabolic reduction (10.2% vs our 4.5%) using more optimized hardware (2.8 kg vs our 3.6 kg) [9]. Despite similar peak torque capabilities (12 Nm vs 15 Nm), their compact form factor potentially facilitated more natural movement patterns. Moreover, their study targeted elderly adults who typically demonstrate increased reliance on the hip joint during locomotion, suggesting that population-specific biomechanical characteristics significantly influence the effectiveness of hip assistance. These differences in hardware design and target population likely account for the observed disparity in metabolic improvement between the two studies. The significantly lower metabolic cost in the OPT compared to the BIO further emphasizes the importance of individualized optimization. This finding is consistent with previous studies that have demonstrated the benefits of personalized assistance profiles over generic profiles based on biological joint moments [28]. The statistical analysis provides quantitative evidence that optimal assistance timing and magnitude differ systematically from biological hip moment patterns during stair climbing. The consistent shift toward later peak timings in both flexion and extension, combined with shorter flexion duration and higher peak flexion magnitude, suggests that effective exoskeleton assistance requires different control parameters than biological hip joint moments. This deviation from biological patterns may reflect the need to account for the mechanical interaction between the user and exoskeleton, as well as the specific demands of powered assistance during stair climbing. This partially aligns with the results of previous studies on level walking, which
have generally reported later peak flexion timing in optimized profiles compared to biological moments [12]. However, our findings differ in terms of flexion duration. While previous level walking studies found longer flexion duration in optimized profiles, our results for stair climbing showed a shorter flexion duration in the optimized profile compared to the biological profile [12]. This difference may be attributed to the specific biomechanical requirements of stair climbing, which involves a more prolonged stance phase and a shorter swing phase compared to level walking. The later peak flexion timing in our optimized profile might help in providing assistance at a more mechanically advantageous point during the swing phase of stair ascent, while the shorter flexion duration and higher peak flexion magnitude could potentially allow for more effective power transfer and rapid transition to the subsequent extension phase, which is crucial for propelling the body upwards during stair climbing [1]. The later peak extension timing in the OPT profiles compared to the BIO profile is consistent with previous findings on level walking [12]. The later peak extension timing may help to provide more effective push-off assistance during the late stance phase of stair climbing, which is crucial for propelling the body upwards and forwards [1]. The similar extension duration between the OPT profiles and the BIO profile suggests that the timing of extension assistance may be more critical than its duration for reducing metabolic costs during stair climbing. Through post-hoc analysis, we employed a best subjectindependent parameters approach that combines each subject's entire metabolic response surface into a single averaged landscape, rather than simply averaging individual optimal points. This comprehensive analysis identified optimal parameters (peak extension timing: 25.11%, extension duration: 41.16%, peak flexion timing: 75.92%, flexion duration: 29.01%, and peak flexion torque: 10.34 Nm). The leave-one-out approach used in our analysis provides two key advantages: first, it prevents potential bias by excluding each subject's own data when determining their best subject-independent parameters, making the results more generalizable to new users; second, it enables a more robust validation of the generalized assistance strategy by simulating the real-world scenario where individual optimization data would not be available. While these parameters suggest the possibility of a generalized assistance strategy that could be more practical for real-world applications, they may not capture the full benefits achievable through individual optimization. The substantial standard deviations across subjects, particularly in parameters like extension duration, indicate significant inter-individual variability in optimal assistance preferences. This finding underscores the potential importance of personalization in exoskeleton control. By examining the complete response surface rather than just optimal points, we gained insights into how different participants respond to various parameter combinations throughout the parameter space. This analysis addresses a fundamental question in exoskeleton control: the extent to which a single set of parameters might be effective across multiple users. While our results suggest some general patterns in optimal assistance preferences, the significant individual variations support the value of personalized optimization approaches. The convergence analysis in our study requires careful interpretation, particularly given the unique characteristics of Bayesian Optimization (BO). Unlike other optimization methods that converge to a single point, BO inherently continues to explore the parameter space through its Upper Confidence Bound (UCB) metric, which balances between exploiting regions of known good performance and exploring areas of high uncertainty. Therefore, our post-hoc convergence analysis should be interpreted as demonstrating reliable metabolic improvement and parameter stability rather than convergence to a singular optimal configuration. Our analysis employed particularly conservative thresholds compared to previous studies. We defined metabolic convergence as achieving less than 0.5% change across three consecutive iterations, substantially more strict than the 4% threshold commonly used in previous studies [10]. For parameter stability, we implemented a 5% threshold. These conservative criteria were chosen to ensure robust performance evaluation, especially given the challenging nature of stair climbing and the limited number of possible iterations. The strenuous nature of stair climbing posed unique challenges to the optimization process. While previous study has demonstrated successful convergence with two control parameters during level walking [10], our five-dimensional parameter space combined with the distinct biomechanical demands of stair climbing presents additional complexity. The 20-iteration limit was specifically established based on quantitative findings from pilot testing, which demonstrated significant participant fatigue beyond this point. While typical metabolic assessments require 5-6 minutes of steady-state data, we employed first-order fitting to estimate steady-state metabolic cost from 2.6 minutes of data, balancing measurement accuracy with experimental feasibility. This practical constraint on iteration count, combined with BO's inherent exploration-exploitation trade-off, influenced our interpretation of convergence. The fact that most parameters did not achieve strict convergence within 20 iterations is consistent with BO's operating principle of continuous parameter space exploration. However, the significant and consistent reduction in metabolic cost achieved demonstrates that the optimization process effectively identified beneficial control settings, even without strict parameter convergence. Although reducing the parameter space to focus solely on either flexion or extension assistance might accelerate convergence, this approach would overlook critical interdependencies between these assistance phases during stair climbing. The apparent stability of the peak flexion timing parameter is particularly interesting, suggesting this parameter may have a more direct relationship with
metabolic cost reduction compared to others. These findings suggest that the parameter space for effective exoskeleton assistance may be relatively broad, with multiple combinations of parameters capable of producing meaningful metabolic benefits. Future studies might benefit from identifying faster-to-measure proxies for metabolic cost, potentially enabling more iterations and thus more thorough parameter space exploration while maintaining the fundamental balance between exploration and exploitation inherent to BO. One limitation of this study is the focus on optimizing assistance profiles for a single stair-climbing speed. Future work should investigate the generalizability of the optimized profiles to different speeds and explore the potential for online adaptation to accommodate variations in user preferences and stair-climbing conditions. Additionally, the long-term effect of the optimized assistance on muscle activation patterns, joint kinematics, and user adaptation warrants further investigation. VIII. CONCLUSION This study demonstrates the successful application of humanin-the-loop optimization to identify individualized hip exoskeleton assistance profiles that reduce the metabolic cost of stair climbing. The results highlight the importance of personalized assistance in enhancing the efficacy of exoskeletons for demanding locomotor tasks and provide insights into the optimal timing and duration of hip flexion and extension assistance for stair climbing.
Check for updates EXOSKELETON Estimating human joint moments unifies exoskeleton control, reducing user effort Dean D. Molinaro1,2*, Inseung Kang3, Aaron J. Young1,2 Robotic lower-limb exoskeletons can augment human mobility, but current systems require extensive, context-specific considerations, limiting their real-world viability. Here, we present a unified exoskeleton control framework that autonomously adapts assistance on the basis of instantaneous user joint moment estimates from a temporal convolutional network (TCN). When deployed on our hip exoskeleton, the TCN achieved an average root mean square error of 0.142 newton-meters per kilogram across 35 ambulatory conditions without any user-specific calibration. Further, the unified controller significantly reduced user metabolic cost and lower-limb positive work during level-ground and incline walking compared with walking without wearing the exoskeleton. This advancement bridges the gap between in-lab exoskeleton technology and real-world human ambulation, making exoskeleton control technology viable for a broad community. Copyright © 2024 The Authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original U.S. INTRODUCTION Realizing lower-limb exoskeleton technology in the real world would enable human mobility to reach new feats, making heavy boxes feel lighter in the warehouse (1), increasing the success rate of search and rescue operations (2), or even enabling new innovations in athletic training and exercise. To date, lower-limb exoskeletons have had substantial success in improving human mobility, including augmenting human energetics by offloading or adding to the mechanical work done by the underlying human musculature (3–9). Although the tangible benefits and potential societal effects of these devices continue to be found, we fail to see this technology deployed in the real world. So, the question remains, what is preventing exoskeleton technology from being realized "in the wild"? One critical challenge lies within the exoskeleton controller (3, 4, 10, 11). Generally, exoskeleton controllers are divided into three layers: high-level, mid-level, and low-level (12). The high-level layer estimates user and environmental states, such as ambulation mode (9, 13-17) or ground slope (18, 19), used to modulate assistance with changes in user joint demands. The state estimates are passed to the mid-level layer, which computes desired assistance on the basis of predefined control laws, such as spline-based assistance trajectories (8, 20, 21). The low-level layer then converts the desired exoskeleton assistance into actuator commands, often using motorlevel state feedback control. Although human-in-the-loop optimization (8, 20–22) and on-the-fly metabolic cost estimation (23, 24) can optimize and personalize assistance, these methods require reoptimizing controller gains for each high-level state, an inherently timeintensive process. Further, these advances in mid-level control are dependent on accurate high-level state estimates. Although physicsdriven (9, 17, 25) and data-driven (13-16, 18, 19, 26-28) models can estimate one or more high-level states, defining, estimating, and subsequently optimizing high-level and mid-level controllers for all of the possible states needed to parameterize human movement inthe-wild is intractable. Instead, instantaneous estimates of the user's underlying joint moments could replace conventional high-level states (29–39). Because lower-limb joint moments naturally vary across ambulation modes and conditions (40), lower-limb joint moments could serve as a single, continuous high-level state for modulating exoskeleton assistance. However, human joint moments cannot be directly measured but are instead computed post hoc using high-fidelity motion capture and six-axis force plate measurements from stationary in-lab equipment (41). Replacing these in-lab systems with current wearable sensor technology results in incomplete information, such as missing ground shear forces (42), and requires potentially cumbersome kinematic sensing of the distal joints along the limb. These limitations hinder the viability of analytical inverse dynamics solutions from wearable sensors alone, especially for more proximal joints. Recent advances in data-driven approaches have improved the mapping between wearable sensor data and user joint moments, even with little to no user-specific data (29-31, 34-39). Using these methods, researchers have conducted initial experiments using instantaneous joint moment estimates in the exoskeleton control loop (29-31, 37-39). For instance, Gasparri et al. (31) introduced an ankle exoskeleton controller based on a quadratic fit between foot force sensor data and user ankle moments during level walking. This work was extended to incline/decline walking, stair ascent/descent, and even mixed terrain with notable outcomes across able-body and clinical populations (37-39). Although promising, it remains unclear how this approach would extend to joints beyond the ankle or generalize
to additional tasks where the mapping between foot force sensor measurements and joint moments is more complex. Alternatively, energy-shaping methods have been developed for assisting the hip, knee, and ankle during multiple ambulation modes and during sit-stand cycles (29, 30). These approaches have demonstrated impressive offline estimation results and potential benefits in lower-limb muscle activation; however, substantial user benefits and online estimation accuracy relative to ground-truth inverse dynamics have not yet been demonstrated. As such, the development of a unified exoskeleton controller capable of autonomously assisting the user across a wide variety of ambulation modes and intensities has remained an open topic of research-maintaining the divide between in-lab exoskeleton technology and real-world benefits. <sup>1George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA 30332, USA. 2Institute for Robotics and Intelligent Machines, Georgia Institute of Technology, Atlanta, GA 30332, USA. 3Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, Cambridge, MA 02139, USA. *Corresponding author. Email: dmolinaro3@gatech.edu In this study, we present an end-to-end framework for controlling a lower-limb robotic exoskeleton based on instantaneous estimates of the user's joint moments using deep learning, deemed unified joint moment control (Movie 1 and Fig. 1). We validated the unified controller using our autonomous hip exoskeleton (figs. S1 and S2) because hip-centric assistance has substantially improved human energetics during several ambulation modes in previous studies (3, 7, 9, 43, 44), likely by offloading the large positive power requirements of the hip joint (45) while minimizing any metabolic penalties caused by distal-borne mass (46). Computing hip moments from inverse dynamics requires distal joint kinematics (41), suggesting the need for complex, cumbersome sensor suites. Instead, we used a temporal convolutional network (TCN) (47) to estimate the total hip flexion/extension moments of the user, including both the biological hip moment and the exoskeleton torque, on the basis of kinematic data from embedded exoskeleton sensors (fig. S3). In addition, the structure of the TCN leveraged temporal information in the input data as a substitute for multijoint sensing, resulting in accurate hip moment estimates without any user-specific calibration or other user state information, such as ambulation mode. Thus, when integrated into the exoskeleton controller via a scale, delay, and filter at the mid-level layer (fig. S4), the hip moment estimates enabled the exoskeleton controller to autonomously modulate assistance across a wide range of ambulatory conditions without any manual experimenter or user intervention. Movie 1. A summary of the results of this study is provided, including representative trials demonstrating the online performance of the unified controller. Fig. 1. Unified joint moment control. A photograph of an individual walking with the autonomous, robotic hip exoskeleton is shown. A TCN, trained using timesynced exoskeleton data and ground-truth labels (depicted left of the photograph), used data from the encoders and IMUs mounted on the exoskeleton to estimate the user's hip moments. When implemented on the device, instantaneous hip moment estimates from the TCN were converted to desired exoskeleton assistance using a mid-level control layer, which scaled, delayed, and filtered the estimates. Mid-level scaling provided a percentage of the total estimated moment as assistance, delay increased the positive mechanical work done by the exoskeleton, and filtering smoothed the assistance. By training the hip moment estimator with data from a variety of conditions, the controller seamlessly adapted assistance across different users, ambulation modes, and ambulation intensities without the need for user-specific calibration data. To validate our approach, we measured user metabolic cost, the rate of whole-body human energy expenditure needed to move, during level-ground walking and 5° ramp ascent, because both demand large amounts of work from the hip joint (40, 45). Given the lightweight design of our exoskeleton and the natural changes in assistance of the unified controller across participants and across modes, we hypothesized that walking with the exoskeleton using the unified controller (Unified Control) would reduce user metabolic cost during both level-ground and incline walking compared with wearing the exoskeleton without assistance (Zero Torque) and compared with walking without the exoskeleton (No Exo) (H1). As an additional benchmark, we also measured user metabolic cost when walking with exoskeleton assistance from previously optimized splines (20, 21) (Spline Control), representing the state of the art in mode-specific, "off-the-shelf" exoskeleton control. To further decompose the energetic effects of our system on the user, we quantified the effect of the unified controller on user lower-limb positive joint work compared with No Exo. We hypothesized that Unified Control would enable users to substitute
their hip joint work with that of the exoskeleton (48), reducing their lower-limb positive mechanical work compared with No Exo (H2). In addition, we quantified the accuracy of our deep learningbased hip moment estimator when integrated into the exoskeleton controller (deployed online) during level-ground walking, ramp ascent/descent, and stair ascent/descent under 35 total conditions of varying walking speeds, ground slopes, and stair heights. Further simulating real-world conditions, the TCN was also tested during neutral standing, during transitions between walking and standing, and during conditions that were not included in the training set. We tested the TCN online, because closing the loop between estimator outputs and human-exoskeleton dynamics can lead to error propagation undetected in offline analyses (49). We compared our approach with a Baseline method designed to emulate bioinspired exoskeleton controllers from previous studies (7, 26). The Baseline method predicted the average hip moment profile as a function of gait phase, computed a priori from the training set, for each ambulation mode and assumed a perfectly accurate ambulation mode and gait phase oracle (the best-case scenario). Because the TCN should model changes in hip moments across users, across intensities, and across strides, we hypothesized that the TCN estimates would have a lower root mean square error (RMSE) and a higher $R^2$ with respect to the ground-truth hip moments compared with the Baseline method (H3 and H4, respectively). Thus, this work presents and validates a unified exoskeleton controller, which leveraged deep learning to accurately map exoskeleton sensor data to user hip moments. The controller autonomously adapted assistance with changes in user joint demand across a wide range of walking speeds, ground slopes, and stair heights without any tuning or experimenter intervention. Using the unified controller, we measured significant improvements in user metabolic cost, relative to No Exo and Zero Torque, and in lower-limb positive joint work, relative to No Exo (Zero Torque was not collected), during level-ground and incline walking. We also found that the TCN significantly outperformed the Baseline method for estimating user hip moments across a variety of conditions. Further, we have released the time-synced time series data of human biomechanics and exoskeleton sensor data used in this study (34 total participants across the four phases of experimental data collections) in conjunction with this publication, increasing the accessibility of machine learning-enabled exoskeleton technology. With these advancements in exoskeleton control and with the corresponding dataset released with this study, exoskeleton technology moves ever closer to being realized in daily life. RESULTS Augmenting human metabolic cost during level-ground walking and ramp ascent User metabolic cost was measured from 10 participants during level-ground walking and 5° ramp ascent under four conditions: Unified Control using a TCN trained from 14 participants of data (from phases 1 and 2 of the experimental protocol), Spline Control, No Exo, and Zero Torque (Fig. 2). The TCN training set included level-ground walking, ramp ascent/descent, stair ascent/descent, standing, and stand-to-walk and walk-to-stand transitions. During level-ground walking, Unified Control significantly reduced metabolic cost by $0.12 \pm 0.13$ W/kg $(5.4 \pm 5.6\%)$ compared with No Exo (P = 0.0219) and by $0.32 \pm 0.07$ W/kg $(12.7 \pm 2.8\%)$ compared with Zero Torque ( $P = 6 \times 10^{-8}$ ), shown in Fig. 2B. Further, during ramp ascent, Unified Control resulted in significant metabolic cost reductions of 0.57 $\pm$ 0.24 W/kg (10.3 $\pm$ 4.4%) compared with No Exo $(P = 4 \times 10^{-8})$ and of 1.06 $\pm$ 0.31 W/kg (17.8 $\pm$ 5.1%) compared with Zero Torque ( $P = 3 \times 10^{-14}$ ), shown in Fig. 2E. The unified controller had no significant difference in metabolic cost during level-ground walking compared to Spline Control, with a difference in metabolic cost of less than 0.01 W/kg between the two conditions (P = 1.0); however, during ramp ascent, the unified controller significantly reduced metabolic cost by $0.27 \pm 0.15$ W/kg $(5.3 \pm 2.8\%)$ compared with Spline Control (P = 0.0025). Reducing joint-level positive mechanical work of the user To analyze the effect of our controller at the joint level, lower-limb
mechanical work was measured across the same 10 participants under the Unified Control and No Exo conditions. The unified controller was deployed using the same hip moment estimator that was used when measuring user metabolic cost. As shown in Fig. 3, the total positive mechanical work of the user's lower-limb joints (sum of hip, knee, and ankle positive work in the sagittal plane) was significantly lower with Unified Control compared with No Exo during both level-ground walking [change of $0.16 \pm 0.05$ J/kg $(17.5 \pm 5.7\%)$ ; $P = 5 \times 10^{-6}$ ] and ramp ascent [change of $0.16 \pm 0.08$ J/kg $(11.7 \pm 6.2\%)$ ; $P = 2 \times 10^{-4}$ ]. At the individual joint level, we found that Unified Control significantly reduced the positive mechanical work of the hip joint by $0.12 \pm 0.04$ J/kg $(29.2 \pm 10.9\%)$ during level-ground walking $(P = 2 \times 10^{-8})$ and by $0.15 \pm 0.08$ J/kg $(22.8 \pm 12.2\%)$ during ramp ascent $(P = 8 \times 10^{-6})$ compared with No Exo (Fig. 3, A and B). Validating TCN accuracy in the loop The TCN, when trained from a 24-participant dataset (from phases 1 through 3 of the experimental protocol), was evaluated when implemented in the exoskeleton control loop during 35 conditions, including level-ground walking at speeds ranging from 0.6 to 1.9 m/s, inclines and declines ranging from -15° to 15°, and stairs spanning the range of ADA (Americans with Disabilities Act)-compliant step heights of 10.2 to 17.8 cm (4 to 7 inches) (fig. S5). In addition, the Baseline method served as a comparison with the TCN for hip moment estimation, which predicted hip moments on the basis of the ambulation mode–specific, participant-averaged hip moment profiles from the TCN training set. The Baseline method was Fig. 2. Exoskeleton effect on user metabolic cost of walking. (A) User metabolic cost was measured during level-ground walking at 1.25 m/s while the user wore the exoskeleton with the unified joint moment controller (Unified Control), while the user wore the exoskeleton with a spline-based controller (Spline Control), while the user did not wear the exoskeleton (No Exo), and while the user wore the exoskeleton as it commanded zero torque (Zero Torque). (B) The resulting metabolic cost of each condition during level-ground walking is shown, and the percent reduction of each condition relative to Zero Torque is depicted with the arrows. (C) The commanded exoskeleton torque averaged across participants is shown for Unified Control and Spline Control during level-ground walking. (D) User metabolic cost was also measured during 5° ramp ascent at 1.25 m/s. (E) The average metabolic cost resulting from the incline tests is shown. (F) The average commanded exoskeleton torques during ramp ascent are shown. Gait cycles were segmented by heel strike, and hip extension is positive. Bars and curves represent means, error bars and shaded regions represent $\pm 1$ SD about the mean, and asterisks indicate statistical significance (multiple comparisons, P < 0.05, n = 10). Fig. 3. Exoskeleton effect on user lower-limb joint work. The per-stride, positive biological joint work of the hip, knee, ankle, and sum of all three joints (total) during (A) level-ground walking and (B) 5° ramp ascent is shown. Joint work was measured while participants wore the exoskeleton, which provided assistance using the unified controller (Unified Control), and while participants did not wear the exoskeleton (No Exo). Average power of the exoskeleton and biological hip, knee, and ankle joints is shown during level-ground walking and ramp ascent (C to H). Gait cycles were segmented by heel strike. All results were computed with respect to the sagittal plane. Bars and curves represent means; error bars and shaded regions represent $\pm 1$ SD about the mean. Statistical comparisons of total positive joint work were computed using paired t tests, and joint-level comparisons were conducted using multiple comparisons post hoc. Asterisks indicate statistical significance (P < 0.05, n = 10). implemented post hoc on the same trials used to evaluate the TCN. The
resulting RMSE and $\mathbb{R}^2$ of the TCN and Baseline estimates were computed with respect to the ground-truth hip moments from inverse dynamics. The overall RMSE of the TCN averaged across the five ambulation modes was $0.142 \pm 0.021$ N·m/kg, which was significantly lower than the Baseline method [change of $0.035 \pm 0.016$ N·m/kg ( $19.8 \pm 9.3\%$ ); $P = 9 \times 10^{-5}$ ] (Fig. 4), with representative strides of the TCN estimates shown in Fig. 5. Significant reductions in RMSE within modes were also found for the level-ground [change of $0.060 \pm 0.022$ N·m/kg ( $29.1 \pm 10.5\%$ ); P = 0.0012] and ramp descent conditions [change of $0.051 \pm 0.035$ N·m/kg ( $27.0 \pm 18.8\%$ ); P = 0.0092]. In addition, the overall $R^2$ of the TCN was $0.840 \pm 0.045$ , which was significantly higher than that of the Baseline method [change of $0.035 \pm 0.034$ ( $4.4 \pm 4.2\%$ ); P = 0.0088] (fig. S6). Further, when comparing the TCN and Baseline methods within ambulation modes, the TCN significantly outperformed the Baseline method in several conditions (Fig. 4C and fig. S6C), often near the extrema within each mode (P values provided in data file S2). Five of the participants also completed additional trials of neutral standing, stand-to-walk transitions, and walk-to-stand transitions while the unified controller provided assistance on the basis of the TCN estimates. We found that the TCN reduced estimation RMSE during stand-to-walk transitions by 0.081 $\pm$ 0.031 N·m/kg $(34.9 \pm 13.5\%)$ and during walk-to-stand transitions by $0.085 \pm 0.014 \text{ N·m/kg}$ $(35.3 \pm 5.7\%)$ compared with the Baseline method (Fig. 4B), with similar improvements in $R^2$ (fig. S6B). Representative strides during each transition are shown in Fig. 5D. In addition, the estimator naturally turned off assistance during neutral standing, with very little difference in RMSE between the estimates from the TCN and those from the Baseline method, which estimated zero hip moment (Fig. 4B). To further investigate the real-world viability of our hip moment estimator, we also quantified the effects on TCN performance when each of the 35 ambulatory conditions was included versus excluded from the training set (details provided in Supplementary Methods). In general, the TCN generalized well when tested on previously unseen conditions (fig. S7), with significant differences between holdin and hold-out RMSE in only 3 of the 35 total conditions. In addition, there were no significant differences in $\mathbb{R}^2$ between holdin and hold-out conditions. DISCUSSION Deep learning-based hip moment estimates yielded autonomous exoskeleton assistance The unified controller presented in this work autonomously adapted exoskeleton assistance across users, ambulation modes, and ambulation Fig. 4. Hip moment estimation RMSE. (A) The average RMSE of the TCN is compared with the RMSE of the Baseline method during level-ground walking (LG), ramp ascent (RA), ramp descent (RD), stair ascent (SA), and stair descent (SD) (multiple comparisons, n=10). The average RMSE across the five ambulation modes (ALL) is also shown (paired t test, n=10). (B) The average RMSE of the TCN is shown during neutral standing (STAND), stand-to-walk transitions (S2W), and walk-to-stand transitions (W2S) relative to the Baseline method (n=5, no statistical tests performed). (C) The average RMSE of the TCN and Baseline method is shown for each intensity per ambulation mode (multiple comparisons, n=10 for all comparisons except for LG at 1.9 m/s and RD at $-15^\circ$ , which were n=9). All TCN results are based on online estimates used in the control loop. All Baseline results were computed post hoc using the same data. Bars and markers represent
means, error bars represent $\pm 1$ SD about the mean, and asterisks indicate statistical significance (P < 0.05). Fig. 5. Representative time series. Examples of hip moments estimated by the TCN and the corresponding ground truth values are shown. (A) Representative strides from 7 of the 13 total level-ground walking conditions are shown (average RMSE of the depicted strides is 0.141 N·m/kg). (B) Representative strides from 10 of the 14 total inclines and declines are shown (average RSME of the depicted strides is 0.138 N·m/kg). (C) Representative strides of stair ascent and descent at each of the tested stair heights are shown (average RMSE of the depicted strides is 0.137 N·m/kg). (D) A representative time series of a stand-to-walk and walk-to-stand transition is shown (average RMSE of the depicted strides is 0.128 N·m/kg). The stand-to-walk and walk-to-stand trials were extended from their original segmentation for visual purposes. Within each ambulation mode, the depicted strides are from the same participant. Gait cycles were segmented by heel strike, and hip extension is positive. intensities without any user-specific calibration data or sensing external to the device. Not only did the deep learning model accurately adapt assistance across all 35 evaluated conditions, but the resulting assistance also provided significant metabolic and joint-level energetic benefits to the user in both ambulation modes evaluated. Specifically, our approach significantly reduced user metabolic cost during both level-ground walking and 5° ramp ascent relative to No Exo and Zero Torque (Fig. 2), confirming hypothesis H1. Although the magnitude of metabolic cost reductions of our controller was similar to those of previous autonomous exoskeleton studies (3), modulating assistance across ambulation modes without any experimenter or user intervention overcomes an important barrier when considering exoskeleton technology for real-world deployment. In addition, further iterations of the exoskeleton hardware or optimizations to the mid-level controller could further improve these results compared with No Exo. Unified control outperformed mode-dependent, off-the-shelf spline control Previous human-in-the-loop optimization studies have found that the optimized shape of spline-based hip exoskeleton assistance is relatively consistent across users (20, 21). As such, averaging previously optimized torque assistance splines from multiple participants (Spline Control in this study) is an effective exoskeleton control strategy for reducing human effort without additional time-intensive optimization (50). Matching metabolic cost reductions compared with Spline Control during level-ground walking demonstrates that the unified controller was as effective as the state of the art for user-independent, mode-specific control but did not require any previous optimization, mode switching, or additional state estimation. During ramp ascent, the unified controller significantly reduced user metabolic cost relative to Spline Control, which was likely a result of the additional 19% of positive mechanical work done by the unified controller compared with Spline Control, despite scaling the two to have near-equivalent peak torque magnitudes (Fig. 2, C and F). This result was unexpected because we expected Spline Control to be a near-optimal controller (20, 21); however, several considerations may explain this result. First, the splines used in this study were optimized on a different device, with differing torque capabilities and actuator dynamics (51); it is possible that human-in-the-loop optimization must be repeated with each new device to maintain energetic benefits. Further, the optimized controller gains from human-in-the-loop optimization may not represent the global optimum, despite often requiring tens of minutes or even hours of walking to converge (8, 20-22). Last, the unified controller introduced in this study instantaneously adjusted assistance across participants and across strides, potentially yielding additional benefits compared with the static profiles of Spline Control. Nevertheless, the result of our metabolic tests confirmed that our unified controller both autonomously modulated exoskeleton assistance across modes and generated exoskeleton assistance as beneficial as or better than the previous state of the art in userindependent, off-the-shelf control, a major step toward real-world human augmentation. Users reduced their joint work to accept exoskeleton assistance In agreement with our metabolic cost findings and hypothesis H2, the unified controller also significantly reduced the total lower-limb positive mechanical joint work of the user during both level-ground walking and ramp ascent (Fig. 3). We found larger relative reductions in metabolic cost between Unified Control and No Exo during ramp ascent but larger relative reductions in positive lower-limb joint work in level-ground walking. This suggests that the additional benefits in metabolic cost during ramp ascent may come from improved muscle-level efficiencies or reduced cocontraction when wearing the device, hinting that future generations of exoskeleton controllers may benefit even more by accounting for the user's underlying muscle dynamics (52). As expected, the benefits of the unified controller on user lower-limb positive joint work were localized at the hip
joint (Fig. 3), with relative reductions of 29.2 and 22.8% relative to No Exo during level-ground walking and ramp ascent, respectively. By delaying the exoskeleton assistance relative to the instantaneous hip moment estimate in the mid-level control layer (Fig. 1 and fig. S4), the unified joint moment controller provided peak assistance torque during the periods of the stride with high hip velocities, increasing the total amount of positive mechanical work provided by the exoskeleton. Thus, the unified controller was able to reduce the positive work at the hip joint by more than 20% despite only scaling the assistance torque to 20% of the total estimated hip moment. This result demonstrates the energetic benefits gained by simply delaying hip exoskeleton assistance relative to the biological joint moment during walking. Deep learning enabled accurate hip moment estimation in the loop Overall, the TCN accurately estimated user hip moments across the 35 ambulatory conditions when deployed online (Fig. 4 and fig. S6). Confirming hypotheses H3 and H4, the TCN significantly outperformed the Baseline method in RMSE and $R^2$ , even though the Baseline method in this study assumed a perfectly accurate ambulation mode classifier and gait phase estimator. In practice, mode classifiers and gait phase estimators also incur error (13–17, 26–28), further increasing the differences between the TCN and Baseline method. Nevertheless, this indicates that the TCN not only captured changes in hip moments as ambulation mode and gait phase varied but also modeled changes in hip moments across participants, across intensities, and/or across strides. Although multiple studies have investigated wearable sensor-based hip moment estimation offline (30, 33-36, 42), this study provides a rigorous validation with respect to ground-truth inverse dynamics when deployed online. Online validation is critical because closing the loop between model estimates and user biomechanics can greatly reduce model performance. For instance, incorrect ambulation mode estimates can lead to controller instability and further propagate error (49). We found that when deployed online, our deep learning-based hip moment estimator obtained similar or even better performance compared with previous offline analyses (34-36, 42). For instance, Forner-Cordero et al. (42) analyzed the effects of replacing force plate measurements with pressure insole estimates when computing inverse dynamics, resulting in a hip moment estimation RMSE of 0.15 N⋅m/kg and correlation coefficient of 0.92 (approximate $R^2$ of 0.85) during level-ground walking. For comparison, the TCN in this study resulted in an estimation RMSE of $0.147 \pm 0.040$ N·m/kg and $R^2$ of $0.879 \pm 0.045$ during level-ground walking across all speeds without the need for any external sensors or data, demonstrating the benefits of our data-driven approach. Further, in our previous work, we used data from a hip goniometer and simulated trunk and thigh inertial measurement units (IMUs) to train a TCN for estimating hip moments across level-ground walking, ramp ascent/descent, and stair ascent/descent (34). This previous study resulted in an average RMSE of 0.131 N·m/kg and $R^2$ of 0.88, which was state of the art for offline, participant-independent human hip moment estimation based on wearable sensors. As shown in Fig. 4 and fig. S6, the hip moment estimator in this study maintained similar performance (average online RMSE of 0.142 ± 0.021 N·m/kg and $R^2$ of 0.840 $\pm$ 0.045) despite additional sensor noise from exoskeleton actuation and closed-loop dynamics between hip moment estimates, exoskeleton assistance, and the resulting kinematics of the user. Thus, using a single regression model for high-level state estimation mitigated error propagation from estimator-controller dynamics, increasing overall controller reliability. The unified controller extended to transient and previously unseen conditions Transitions between standing and walking are extremely common in community ambulation (53) but typically are not accommodated by conventional exoskeleton controllers given the challenge of parameterizing mode transitions. By naturally varying assistance based on the estimated joint moments, the unified joint moment controller seamlessly adjusted exoskeleton assistance through mode transitions, without the need for any additional modifications to the controller (see Fig. 4B and fig. S6B with representative strides in Fig. 5D). This result also aligns with our previous work, which quantified TCN hip moment estimation performance during ambulation mode transitions (34), suggesting that the TCN estimates remain viable even during transient ambulation. Further, the model generalized well when tested on conditions absent from the training set (fig. S7). Model RMSE comparing hold-in versus hold-out conditions was only affected at the extrema
of the dataset, meaning that the model interpolated between conditions in the training set well but began to lose performance during extrapolation. We did not find any significant differences in $\mathbb{R}^2$ between holding in and holding out each condition. This suggests that the TCN maintained the correct "shape" of the hip moments but likely began to improperly scale the magnitudes when extrapolating to high intensities. Nevertheless, this result promotes the need for diverse datasets consisting of large variations in condition and intensity to assist model generalization, which is a key consideration for deploying these systems in the wild. Human-exoskeleton dataset of multimodal ambulation To reduce the barrier to entry of machine learning–enabled exoskeleton research, we have released the complete set of exoskeleton and human biomechanical time series data associated with this study. The dataset was collected with 34 total participants, which was large enough to demonstrate diminishing returns in model accuracy improvements with respect to increasing training set size (fig. S8). Further, this dataset could also be used to develop algorithms that eliminate the need for device-specific data when developing datadriven exoskeleton controllers, perhaps by building from our data transformation approach outlined in Supplementary Methods, which would greatly improve the accessibility of this technology. Limitations This study had multiple limitations. We added a second onboard processor (the machine learning coprocessor) to run the TCN online, which added mass (5% of total), additional power requirements, and software complexity to the system. In addition, the TCN used a substantial amount of device-specific (although not userspecific) training data. Training the TCN from data collected on a different hip exoskeleton did yield a feasible model for initial controller development (see Supplementary Methods); however, a detailed analysis of transferring data across multiple devices was outside the scope of this study. Another limitation of this study was that the metabolic cost and joint work analyses were conducted using a different model (trained from 14 participants of data) compared with the analysis of online model performance (trained from 24 participants of data). It is possible that further metabolic cost and joint work experiments could yield even better results using the updated hip moment estimator, meaning that our results likely represent a lower bound of what is achievable with the unified controller. In addition, although our unified control framework may generalize to other human degrees of freedom, this study only investigated hip flexion/extension moment estimation and assistance. Last, the accuracy and corresponding user outcomes of the unified controller were only evaluated on able-body participants. Although exoskeleton technology could greatly benefit able-body individuals in many real-world applications, translating this technology to individuals with mobility impairments could also lead to substantial benefits. Conclusion In general, this work presents a model-free framework that unifies exoskeleton control across a variety of ambulatory conditions. Where previously proposed unified controllers have used anatomical models calibrated with relatively small amounts of data (29–31, 37–39), we leveraged deep learning with a large amount of labeled training data. The result was a unified controller that effectively augmented user energetics and could adapt to a very broad range of ambulation modes and intensities. This presents a major advancement in the effort of human augmentation with applicability to a broad range of researchers, technologists, and future end users of exoskeleton technology, many of whom may be one and the same. We expect that this technology will enable researchers to ask new questions about human mobility and augmentation that take place off the treadmill and in more realistic settings. For technologists interested in large-scale health monitoring, for instance, our joint moment estimator could also be used as a wearable sensor–based solution to monitor joint kinetics during daily life. Last, for the end user, we hope that this technology will spur safer and more efficient efforts, including those in factories and warehouses, high-endurance missions like search and rescue, and in athletics and exercise at the professional or recreational level. MATERIALS AND METHODS Participants and experimental protocol This study consisted of four phases, involving the enrollment of 25 able-bodied participants and the use of a nine-participant preexisting dataset from our previous work (35). All participant information relevant to this study is provided in table S1, and each participant provided written informed consent according to the protocols approved by the Georgia Institute of Technology Institutional Review Board. To ensure that the hip moment estimator was consistently evaluated on a user-independent basis, meaning trained without user-specific data, each participant only participated in one of the four phases of the study protocol. During each phase, motion capture and ground reaction force (GRF) data were collected (fig. S5), time-synced with the exoskeleton sensor data, and used to compute ground-truth joint moments from OpenSim inverse dynamics (54, 55). A summary of the experimental methods is provided below with a detailed description provided in Supplementary Methods. The first phase of this study consisted of transforming a nineparticipant, preexisting dataset collected using a different hip exoskeleton to be
used as TCN training data compatible with our device (35). Specifically, the dataset was collected with the same exoskeleton sensor modalities as those of our custom hip exoskeleton but had different IMU placements. When collecting the dataset, each participant walked over level ground, along four inclines and declines with ground slopes ranging from 7.8° to 12.4°, and up and down a staircase at four different stair heights ranging from 10.2 to 17.8 cm. The dataset consisted of exoskeleton sensor data and corresponding ground-truth joint moments. To transform the dataset for use in this study, IMU transforms were computed empirically from a single-participant experiment, which adjusted the positions and orientations of the recorded IMU data to match those of our custom hip exoskeleton. All other protocols and analyses were conducted using our custom hip exoskeleton (fig. S1). The second phase of this study consisted of collecting devicespecific labeled data using our custom hip exoskeleton from five participants. During phase 2, the hip exoskeleton provided assistance using the unified controller, in which the hip moment estimator was trained from the phase 1 dataset. Each participant completed the same trials conducted in phase 1, along with additional standing, stand-to-walk, and walk-to-stand trials. The third phase of this study was conducted with 10 participants to evaluate the effects of the unified controller on user metabolic cost and lower-limb positive mechanical joint work. The hip moment estimator used throughout phase 3 was trained on the phase 1 and 2 datasets. The phase 3 protocol was subdivided into two sessions. Session 1 served as a training session for walking with the exoskeleton, provided an additional dataset for model training and evaluation, and served as the session for analyzing user lower-limb mechanical joint work during level-ground walking and 5° ramp ascent with the unified controller (Unified Control) and without wearing the exoskeleton (No Exo). While motion capture and GRF data were recorded, participants walked on a level treadmill at nine speeds ranging from 0.75 to 1.75 m/s, on an inclined/declined treadmill at 10 slopes ranging from −15° to 15°, and up and down a staircase at four stair heights ranging from 10.2 to 17.8 cm using the unified controller. Participants also completed No Exo trials of level walking and 5° ramp ascent at 1.25 m/s, which were used in the joint work analysis. During session 2 of phase 3, metabolic data were measured during level-ground walking and 5° ramp ascent, both at 1.25 m/s after completing 16 min of habituation. Participants completed the metabolic trials under four conditions using a within-participant counterbalanced design: Unified Control; Spline Control, which provided assistance based on previously optimized level ground and 5° ramp ascent assistance trajectories (20, 21); No Exo; and Zero Torque, which consisted of wearing the exoskeleton while it commanded zero torque. When evaluating the effects of the training set size on TCN performance, the model continued to improve when adding the phase 3 data (details in Supplementary Methods). Thus, we conducted phase 4 of the experimental protocol, in which 10 additional participants used the unified controller while motion capture and GRF data were collected. During phase 4, the TCN was trained using the data from phases 1, 2, and 3. Participants walked on a level treadmill (13 walking speeds ranging from 0.6 to 1.9 m/s), on an inclined/ declined treadmill (14 ground slopes ranging from −15° to 15°), and up and down a staircase (four stair heights ranging from 10.2 to 17.8 cm). To test the hip moment estimator on previously unseen conditions, four level-ground walking speeds and two inclines/ declines in this study phase were not collected in any of the previous phases used for training the model. In addition, two of the stair ascent/descent trials were repeated in which each corresponding stair height was withheld from the hip moment estimator training set. Five of the 10 participants also completed trials of neutral standing, stand-to-walk transitions, and walk-to-stand transitions. Robotic hip exoskeleton This study used a custom-designed exoskeleton that provided sagittal plane hip torque assistance of up to 18 N·m [~30% of peak biological hip moments during walking (40)], shown in fig. S1. The system was fully autonomous, meaning that all components were self-contained. The exoskeleton measured angular position from actuator-mounted encoders in addition to linear acceleration and angular velocity from three IMUs mounted on the exoskeleton backplate and left and right thigh struts. Encoder velocity was also computed online using backward finite differencing and was lowpass filtered with a 10-Hz cutoff frequency. A graphics processing unit–enabled secondary processor mounted onboard
the device was used for online TCN deployment. The total exoskeleton mass was 4.8 kg, including electronics and batteries. Additional information is provided in Supplementary Methods and in figs. S1 and S2. Hip moment estimation using a TCN We used a TCN (47) to estimate the exoskeleton user's hip flexion/ extension moments on the basis of measurements from the onboard exoskeleton sensors (fig. S3). TCNs use dilated causal one-dimensional convolution to efficiently map patterns in the input sequence data to the target output (fig. S3C), replacing the need for hand-engineered feature extraction methods (33, 56). In addition, the fixed input sequence length of TCNs has been shown to retain information over longer periods of time than recurrent networks (47), suggesting that they can better leverage time history information on larger timescales compared with alternative neural network architectures. As such, TCNs have been successful in many sequence modeling tasks (34, 47, 57), including achieving state-of-the-art offline performance in human hip moment estimation compared with alternative deep learning models, which we demonstrated in our previous work (34, 35). The TCN used for hip moment estimation in this study was implemented as described by Bai et al. (47) using the hyperparameters optimized in our previous work (34). The input sequence of the TCN consisted of the ipsilateral actuator encoder position and velocity, six-axis ipsilateral thigh–mounted IMU data, and six-axis pelvis-mounted IMU data. Sensor data from the contralateral limb were not included as input to the model because the majority of the training and testing conditions included symmetric movements, which could cause the model to overfit to this type of behavior. On the basis of the selected hyperparameters of the model, the input sequence to the TCN was composed of a ℝ180×14 R187×14 sequence consisting of the latest 930 ms of exoskeleton sensor data. Given an input sequence, the model estimated the instantaneous flexion/ extension moment of the user's left or right hip, scaled by the participant's body mass. Ground-truth hip moments from our biomechanical model, which were used as the labels during model training, were the sum of both the biological hip moment and the exoskeleton torque (the total hip flexion/extension moment at the joint). This means that the TCN was trained to estimate the total hip moment, which helped to preserve the mapping between the kinematic input data and resulting joint moments across variations in exoskeleton assistance. Additional information about the TCN hyperparameters, implementation, and training is provided in Supplementary Methods. Exoskeleton mid-level control Because hip exoskeleton assistance that solely mimics the user's biological hip moments is likely suboptimal for augmenting human energetics (20, 21, 58), we designed a three-step mid-level control layer to convert hip moment estimates into exoskeleton assistance (fig. S4). First, the incoming hip moment estimates were scaled by 20% of their total magnitude. This assistance magnitude has previously been shown to benefit the user (59) and resulted in peak torque assistance close to the maximum assistance the exoskeleton could provide. The estimated hip moments were then delayed using a first-in-first-out buffer before being used to command the exoskeleton (see details below). Last, the delayed torque values were lowpass filtered using a second-order Butterworth filter with a 10-Hz cutoff frequency to smooth the commanded torque, imparting an additional 25 ms of delay to the signal. Delaying peak hip assistance timing relative to the biological hip moment can provide additional metabolic benefits to the user (20, 58). As Ding et al. (58) explained, delayed hip assistance increases the amount of positive mechanical work done by the exoskeleton because the peak assistance of a delayed controller better aligns with peak hip velocities during the stride. Using the data available from Camargo et al. (40), we found that delaying exoskeleton assistance by 125 ms relative to the biological hip moment could theoretically increase the positive mechanical work done by the exoskeleton during level walking by 70% (fig. S4A). In support of this delayed assistance strategy, we found that delaying exoskeleton assistance between 100 and 150 ms was preferred by several novice and expert users during pilot testing compared with smaller delay magnitudes; however, in an N = 3 pilot study (detailed in Supplementary Methods), user metabolic cost was not sensitive to delay magnitudes ranging from 75 to 175 ms (fig. S4B). Specifically, the mid-level control delays that we tested only affected user metabolic cost by a maximum of 2.9% across level-ground walking and 5° ramp ascent. Testing even smaller delays may result in larger metabolic penalties, especially given that the 75-ms condition was the worst metabolically; however, these small delays were omitted from the pilot study because they were uncomfortable to the
user. In addition, delays below 35 ms could not be tested because of controller limitations from filter delay and worst-case model inference latency. For the remainder of this study, a programmed delay of 100 ms (total delay of 125 ms including the low-pass filter) was used to minimize overall delay and to align with user preference. Analyzing online hip moment estimation accuracy The estimated hip moments recorded onboard the exoskeleton during phase 4 of the experimental protocol were aligned in time with the ground-truth hip moments post hoc to evaluate the accuracy of the TCN when integrated into the unified controller. Two common performance metrics were used to analyze the estimator accuracy (30, 31, 34, 42): the interparticipant average RSME and the interparticipant average of the square of the Pearson correlation coefficient (R2 ). Average RMSE provided an absolute metric of error and is easily compared to previous studies that have investigated wearable sensor–based joint moment estimators. Average R2 provided a nondimensional metric to analyze the goodness of fit of the TCN, meaning the amount of variance in the ground-truth hip moments explained by the TCN estimates via a fitted line. When considering joint moment estimation for exoskeleton control, R2 also provided a metric for analyzing the ability of the model to correctly estimate the shape of the hip moment signal, ignoring error induced from incorrect scaling or bias. In general, this provided a useful metric to evaluate the utility of the hip moment estimator given that scale and bias of the signal could be modified on the fly by the mid-level exoskeleton control layer as needed. Results reported per intensity, as in per walking speed, ground slope, and stair height, were individually computed per condition and then averaged across participants. Results reported per ambulation mode were computed by taking the average of the results computed per intensity within the respective mode, for instance, the average of the RMSE values computed per level-ground walking speed then averaged across participants. Overall results were computed similarly by averaging the results computed per ambulation mode, then averaging across participants. The accuracy of the TCN was benchmarked against a Baseline method designed to emulate conventional exoskeleton controllers that use predefined ambulation mode–specific curves to compute assistance on the basis of gait phase estimates (7, 26, 43). Specifically, the Baseline method was implemented post hoc and estimated the user's hip moments on the basis of a precomputed hip moment curve for each ambulation mode. The hip moment curve for each ambulation mode (level-ground walking, ramp ascent, ramp descent, stair ascent, and stair descent) was computed as the interparticipant average hip moment over the stride from the ground-truth hip moments in the same dataset used to train the TCN in phase 4 (the hip moment data from phases 1, 2, and 3 of the study protocol). During stand-to-walk and walk-to-stand transitions, the Baseline method used the level-ground walking profile. In addition, the Baseline method simply predicted zero hip moment for the standing trials. In all cases, the Baseline method was given access to a perfectly accurate ambulation mode classifier and gait phase estimator that has error in practice (9, 13–16, 26, 27), meaning that our benchmark represented the best-case (yet unrealistic) scenario for estimating hip moments from mode-specific curves. In this case, outperforming the Baseline method meant that the participant-independent TCN captured interparticipant, intercondition, and/or interstride variability that the Baseline method could not represent. Statistical analyses All statistical tests were conducted using Minitab v19 with an α level of significance of 0.05. Further, all statistical tests were computed using repeated-measures (within-participant) methods. When comparing differences among multiple factors and/or multiple withinfactor conditions, a post hoc multiple comparisons test was used to identify significant pairwise differences in the case that significant effects were found from an analysis of variance (ANOVA). All post hoc multiple comparisons were conducted using a Bonferroni correction to control the family-wise error rate. The Bonferroni correction can greatly reduce the statistical power of each pairwise comparison when many pairwise matches exist. In addition, many of the possible pairwise comparisons within each analysis were irrelevant to our hypotheses. Thus, we only evaluated a subset of the possible pairwise comparisons, which were selected a priori (before looking at the results) to limit the amount that each P value needed to be adjusted. Because Minitab did not support this planned comparison approach, we ran a full multiple comparisons test after each ANOVA that yielded statistical significance and adjusted the P values to account for the reduced number of comparisons being evaluated. Metabolic cost comparisons across the four tested exoskeleton assistance conditions (Unified Control, Spline Control, No Exo, and Zero Torque) were analyzed for a main effect using a one-way ANOVA followed by a multiple comparisons test. Differences in positive joint work between the exoskeleton conditions (Unified Control and No
Exo) across the lower-limb joints (hip, knee, and ankle) were evaluated using a two-way ANOVA for level ground and ramp ascent. Pairwise comparisons were only conducted for testing significant differences between Unified Control and No Exo within each joint. In addition, the total positive lower-limb joint work resulting from Unified Control and No Exo were compared separately from the other joints using a paired t test for each ambulation mode. The same statistical tests were run for analyzing both the RMSE and the R2 of the hip moment estimates from the TCN and Baseline method. The overall average results of the TCN across the levelground, ramp ascent, ramp descent, stair ascent, and stair descent conditions were compared with those of the Baseline method using a paired t test. For comparisons at the ambulation mode level, a twoway ANOVA was used to test for significant main and interaction effects across ambulation modes and between estimators (the TCN and Baseline method). A post hoc multiple comparisons test was also used to test for pairwise differences between the two estimators during each ambulation mode. Within each ambulation mode, a two-way ANOVA was used to test for significant main and interaction effects across ambulation mode intensity and between estimators. In addition, a post hoc multiple comparisons test was used to test for significant differences between the TCN and Baseline method within each intensity.
Optimized Mappings from Biological Hip Moment Estimates to Exoskeleton Torque can Personalize Assistance Across Users and Generalize Across Tasks Justine C. Powell, Ethan B. Schonhaut, Dean D. Molinaro, and Aaron J. Young, Senior Member, IEEE Abstract—Recent advancements in data-driven methods have enabled real-time estimation of biomechanical states for exoskeleton control. While biological joint moments can be directly used to scale exoskeleton assistance, this approach is often suboptimal. An optimized mapping between biological joint moments and exoskeleton assistance could enhance end-to-end controllers based on the user's physiological state. We introduce a flexible parametrization of biological moment-based control using delay, scaling, and shaping terms to transform joint moment estimates into commanded torque. We performed human-in-theloop optimization, using metabolic cost to evaluate each iteration's controller parameters, for 9 subjects across three ambulation modes: level walking at 1.1 m/s, 1.5 m/s, and 5° inclined walking. We evaluated three methods of exoskeleton control: 1. Personalized/Task Dependent, 2. Task Dependent/Nonpersonalized, and 3. Task Agnostic/Non-personalized. On average, our personalized approach provided the greatest benefit of 18.3% reduction in metabolic cost compared to walking without the exoskeleton, with the task dependent and task agnostic controllers producing similar reductions of 8.6% and 8.4%, respectively. Our results show that while generalizable, task agnostic control parameters can improve user energetics across cyclic tasks, fully personalized exoskeleton control parameters yield larger metabolic reductions, highlighting the value of personalizing exoskeleton assistance to users across many diverse tasks. Index Terms— Prosthetics and Exoskeletons, Optimization and Optimal Control, Human Factors and Human-in-the-Loop, Deep Learning in Robotics and Automation I. INTRODUCTION Artificially intelligent technologies capable of improving everyday life are becoming increasingly commonplace in society. Wearable devices, such as robotic assistive exoskeletons, can augment user movement to reduce the amount of energy expended throughout the day. These devices, however, are difficult to deploy outside of the lab due to human and task level variability, prompting the need to personalize exoskeleton assistance to the user. Humans continually adapt and optimize our movements to successfully improve task efficiency, as evidenced during steady state treadmill walking [1]. Even with a reduced task space that only encompasses ambulatory tasks, the conditions users may adjust to throughout the day are numerous, with every individual performing these adaptations in unique ways. Assistive robotic exoskeletons have been shown to reduce the energetic demands of human locomotion across discrete ambulatory tasks [2], [3], [4], however these studies often do not scale to the wide array of movements or environments users may encounter throughout the day. For wearable assistive exoskeletons to provide maximum benefit, they must allow users to fine tune and personalize these devices to fit their needs, while simultaneously allowing for adaptability and generalizability to the vastly diverse nature of human movement. Key barriers to translating wearable exoskeletons from the lab to the real world lie in developing control strategies able to handle the transient nature of human mobility. Current exoskeleton control strategies can generally be broken up into three hierarchical levels: high, mid, and low level control [5], [6]. High level control is used to detect and characterize the environmental conditions the user is operating in, such as walking speed or ground slope or the internal state of the user, such as biological joint moment. Mid-level controllers, such as the control framework described in this manuscript (Fig. 1a), typically use environmental information from a high-level controller in conjunction with real-time human-based sensor data to compute the amount of exoskeleton joint torque based on a set of state dependent dynamics. Low level controllers are device dependent [6] attempt to minimize the error between the commanded torque from the controller and the actual torque delivered by the actuators to the subject during deployment. Mid-level lower limb exoskeleton controllers are diverse in their formulation and are characterized by the modality of their input [5], [6]. Kinematic based controllers transform the subject's kinematic information, measured via sensors such as encoders or inertial measurement units (IMUs) placed on board the exoskeleton, to exoskeleton torque using an equation such as impedance control, the user's gait cycle, or a pre-prescribed spline-based timing controller [3], [7], [8], [9]. These controllers, however, are highly task specific and are only This work was supported in part by NSF FRR under Award 2233164 and in part by NSF NRI under Grant 1830215. (Corresponding author: Ethan B. Schonhaut). Justine C. Powell was with the Georgia Institute of Technology Institute for Robotics and Intelligent Machines (IRIM), Atlanta, GA 30332. Ethan B. Schonhaut* is
with the Georgia Institute of Technology George W. Woodruff School of Mechanical Engineering, Atlanta, GA, 30332. (e-mail: eschonhaut3@gatech.edu). Dean D. Molinaro was with the Georgia Institute of Technology, Institute for Robotics and Intelligent Machines (IRIM) and the George W. Woodruff School of Mechanical Engineering, Atlanta, GA, 30332. He is now with Boston Dynamics AI Institute, Cambridge, MA, USA. Aaron J. Young is with the Georgia Institute of Technology, Institute for Robotics and Intelligent Machines (IRIM) and the George W. Woodruff School of Mechanical Engineering, Atlanta, GA, 30332. (e-mail: aaron.young@me.gatech.edu) Fig. 1: (a) Schematic of the experimental design. The experiment was run across three different walking conditions: level ground walking at 1.1 m/s, level ground walking at 1.5 m/s, and 5° incline at 1.1 m/s. The human in the loop optimization (HILO) loop, in blue, was used to metabolically optimize the exoskeleton's control parameters over the course of 21 iterations. Two microprocessors (green and purple, respectively) are used to control the exoskeleton's actuators. A National Instruments MyRio (Microprocessor 1) receives information directly from the exoskeleton's actuators and peripheral sensors and sends the information to an Nvidia Jetson Nano (Microprocessor 2). The Jetson Nano feeds the information through a temporal convolutional network (TCN) to estimate the subject's biological hip joint moment and returns the information to the MyRio. Biological joint moment estimates and the control parameters from the HILO loop are then passed to the MyRio to calculate the instantaneous torque command to send to the actuators. (b) Side view of the hip exoskeleton used during experimentation consisting of a custom carbon fiber hip interface, colocated hip motors, and an electronics backpack containing the system mechatronics. Note, the subject in (b) is an author on this manuscript and has thus consented to the use of this image. successful under strict conditions, which does not accurately represent the diversity of movements and environments that humans operate within. Neural based controllers use sensors that capture signals inside of the human subject to compute exoskeleton assistance. This includes controllers that transform muscle activation measured via electromyography (EMG) to output torque [10], [11], as well as more invasive techniques that transform brain activation via electroencephalography (EEG) to exoskeleton assistance torque through the use of brain computer interfaces [5], [6], [12]. While these approaches directly use the user's physiological signals to influence assistive device control, they can be difficult to incorporate in real-time due to issues with system reliability and robustness [13], [14], and often still require task specific gains and thresholds. A relatively newer class of control uses deep learning (DL) to transform data from the subject sensor domain to the actuator torque domain, usually through a proxy that estimates some internal physiological state of the user, such as biological joint moment [15], [16], [17]. Computing internal physiological states such as lower limb biological joint moments, however, is computationally expensive, often requiring access to motion capture systems and ground reaction forces to compute ground truth biological moment labels [18], [19]. Likewise, assistive exoskeletons often rely on onboard sensing, making real-time computation of these labels within the control loop of a device extremely difficult. Deep learning-based estimators can circumvent this computational limitation by estimating biological joint moments instantaneously from sensors that are easily deployed onboard a wearable assistive device. These approaches, such as the temporal convolutional network (TCN) developed by Molinaro et al. [15], [16] have been shown to be an effective method of continuous exoskeleton control, or end-to-end control, across a wide range of users and tasks [15], [16], [17], [20]. This approach can directly estimate biological joint moments at the hip and knee based on different wearable sensors such as joint encoders, IMUs, and pressure sensing insoles. Furthermore, this approach uses a task and subject agnostic approach by training on a sufficient number of diverse subjects and tasks to improve generalizability across different users and activities [15], [16]. This study employs the TCN developed by Molinaro et al. [16], where biological joint moment estimates are produced instantaneously upon receiving exoskeleton sensor data. The estimates from this network are inputs into a mid-level controller that scales, delays, and filters torque estimates to compute the assistance applied to the user based on the user's body mass and the maximum torque output of the device actuators. The availability of this instantaneous sensor data along with new developments in mid-level control parametrization provide an avenue to use biological torque estimates as a control signal for optimized and improved exoskeleton assistance. While estimating biological joint
moment represents an effective method of generalizable exoskeleton control across different tasks and activities [15], [16], studies have shown that providing purely biological joint moment-based assistance is not metabolically optimal during steady state walking [8], [9]. Similarly, pilot studies from our previous work have shown that delaying exoskeleton assistance relative to the biological hip moment improved metabolic benefit across subjects [16], [20]. During these pilot studies, this benefit was expected because the added delay maximizes the positive work done by the exoskeleton [21], [22], indicating that further optimized parametrizations of exoskeleton mid-level controllers can increasingly improve metabolic benefits for users. Human in the loop optimization (HILO) represents a promising approach to determine the best controller settings and parameters on a per user and per task basis in steady state, cyclic tasks such as walking. This process uses a human outcome measure or cost, such as user energy expenditure, as a metric to assess the effectiveness of a given controller's settings or parameters. These parameters and their associated cost are then fed into an optimizer that determines the next set of parameters to test [8], [9], [21], [22], [23], [24]. As the optimizer iterates, the controller parameters that correspond to the minimum cost for the user and task can be found. Most HILO torque assistance profiles, however, are based on prescribed spline-based torque profiles for a specific task, which does not account for the diversity of human movement. These devices often require a new optimization for each task-specific assistance profile provided by the device, which is not only time consuming to perform, but most importantly ignores the link between the device and the user's underlying physiology. This study expands on previous work from our group by Molinaro et al. [16] by formally optimizing the relationship between biological joint moment and applied exoskeleton torque across multiple subjects and ambulation modes (Fig. 1). Specifically, we aim to understand how sensitive exoskeleton mid-level control is to different parametrizations of the applied continuous torque assistance of our personalized approach during steady state locomotion tasks. Furthermore, we aim to analyze the relationship between positive exoskeleton power and metabolic cost for our continuous end-to-end control approach, as well as how user preference responds and changes with different exoskeleton control parameters. We hypothesize that I: Personalized, task specific biological-torque based exoskeleton control will provide a significant energetic benefit in comparison to non-personalized, task dependent controller settings, II: non-personalized, task dependent controller settings will provide significant metabolic benefit over nonpersonalized, task-agnostic controller settings, III: metabolic cost reductions will be correlated with positive power provided by the exoskeleton during assistance, such that the most effective strategies maximize positive mechanical work provided by the exoskeleton to the user. Here, we introduce our control framework that leverages the advantages of a generalizable, task-agnostic end-to-end exoskeleton controller in conjunction with the benefits of HILO to optimize and unify the user's underlying physiological signals to end-to-end exoskeleton control. II. METHODS A. The Exoskeleton & Controller The device used in this experiment was a fully autonomous, one degree-of-freedom (DOF), untethered hip exoskeleton (Fig. 1b) that was designed, fabricated, and previously validated by our group [16]. The exoskeleton's assistance was designed to assist in the sagittal plane (hip flexion and extension), with torque assistance supplied by actuators with a max output torque of 18 Nm and transmitted to the subject's lower body through custom carbon fiber thigh struts. The exoskeleton's onboard sensors included IMUs placed posteriorly on the pelvis and on the frontal plane of each thigh, as well as encoders available via the device actuators. The backpack attached to the exoskeleton contained batteries, two microcontrollers, and electrical breakout boards required to control the exoskeleton. The total weight of the system was 4.8 kg, with more details on device design available in our previous publication [16]. This study's goal was to determine the optimal relationship between biological joint moment and output exoskeleton torque at the hip joint. Previous work by our group has validated the use of a subject independent TCN model to directly estimate biological joint moments as a means of exoskeleton control using the same hip exoskeleton used in this study with an $R^2$ of $0.840\pm0.045$ and an RMSE of $0.142\pm0.021~N\cdot m/kg)$ [16]. This TCN model was deployed on an Nvidia Jetson (Nvidia, Santa Clara CA) microprocessor which communicates with an NI MyRio (National Instruments, Austin TX) during the experimental loop (Fig. 1a). B. Controller Parametrization The mid-level control parametrization (1) uses three terms to transform estimated biological hip moment $(\
hat{\tau}{bio})$ to exoskeleton assistance torque $(\tau{cmd})$ at a given time (t). The scale $(\alpha)$ term changes the magnitude of torque assistance as the actuator's maximum torque is much less than that of the hip joint. In this experiment, the scale term was not optimized and instead held constant at 20% of the subject's biological hip joint moment, due to device torque limits. The delay term (d)determines the time difference in milliseconds between when the subject exerts a biological joint moment and when they receive the exoskeleton assistance. In Molinaro et al. [16], the delay term was held to 125 ms based on an N=3 pilot study comparing metabolic cost to several delay terms, which furthermore, was in correspondence to the hypothesis proposed in Camargo et al. [23] that timings between 100 to 150 ms could maximize work the exoskeleton supplies to the user. The shape term $(\lambda)$ is a non-linear transformation based on a power law to change the impulsiveness of the torque profile. The values of $\lambda$ were chosen such that during the optimization process and within any ambulation mode, all combinations of shape and $$\tau_{cmd}(t) = \begin{cases} \frac{[\|\hat{\tau}{bio}(t-d)\|]^{\lambda} * \alpha}{\hat{\tau}{flexion}}, & \hat{\tau}{bio}(t-d) \ge 0\ \frac{[\|\hat{\tau}{bio}(t-d)\|]^{\lambda} * \alpha}{\hat{\tau}{extension}} * -1, & \hat{\tau}{bio}(t-d) < 0 \end{cases}$$ (1) Fig. 2: Steps to transform a subject's biological hip joint moment to exoskeleton assistance torque. The subject produces the biological torque which is measured via onboard sensor data by the MyRio microprocessor. The Nvidia Jetson receives sensor data from the MyRio and passes it through a biological joint moment estimator where it is filtered and returned to the MyRio with a small filtering and communication delay. Step 1: A scaling factor < 1 (α ≈ 0.2) reduces the magnitude of the moment to an actuator torque value within the system limits. Step 2: A delay term 65 ≤ ≤ 240 (including the filter & communication delay) further shifts the exoskeleton control profile away from the original biological hip moment. Step 3: A shaping term . 5 ≤ ≤ 1.6 modulates the sharpness and flatness of the peaks of the control profile. delay parameters would result in similar magnitudes of . However, because the shaping term is exponential, with values ranging from 0.5 to 1.6, there are instances where the magnitude of commanded exoskeleton assistance torque exceeds the capability of the device actuators. To account for this, we estimated the peak net, flexion, and extension joint moments for each subject during a habituation period and adjusted the scaling term, α, to ensure a consistent scaling of 20% peak biological joint moment of the user as parameters are optimized during the experiment. Furthermore, due to device limitations, any torque commanded above peak exoskeleton actuator torque was limited to 18 Nm of torque applied to the user. The effects of each term on an optimal biological joint moment profile adapted from [9] can be seen in Fig. 2. C. Data Collection Nine able bodied participants (8 male, 1 female, average mass 69.14 ± 13.83 kg) were recruited to participate in this study. The study experiment included ambulating while wearing the exoskeleton as the exoskeleton controller parameters were optimized across three different treadmill conditions in three separate 3-hour long sessions. The three treadmill conditions are as follows: level ground walking at 1.1 m/s, level ground walking at 1.5 m/s, and 5° incline at 1.1 m/s. Each session began with a 15-minute habituation period in which the subject was exposed to a range of parameter combinations to condition them to walking in the exoskeleton. The subjects then walked for 25 two-minute recorded trials consisting of two trials with no exoskeleton, two trials walking in the unpowered exoskeleton, and 21 trials walking with exoskeleton assistance. Prior to participating, each participant gave informed written consent to participate in a Georgia Tech Institutional Review Board-approved study. To optimize our control framework, we used metabolic cost measured via indirect calorimetry as our cost function, which measures the user's oxygen intake and carbon dioxide output to compute their energy expenditure [2], [25], [26]. Indirect calorimetry remains the gold standard for measuring user energy expenditure, making it an easily attainable metric to use during HILO. Our optimization occurred for
each ambulation mode across 21 distinct trials, where the exoskeleton assisted the user. During each trial, we collected 90 seconds of sensor data from onboard the exoskeleton and two-minute measurements of metabolic cost via the Parvo metabolic cart (TrueOne 2400, ParvoMedics). Across all subjects and conditions, the first 6 iterations of assistance were made up of pre-selected parameters to seed the optimization process with the same data points across subjects (Fig. 3a). One of these 6 points, the purple triangle in Fig. 3, was selected based primarily on an N=3 study that swept only the delay term for biological joint moment control [16]. A second pre-determined parameter combination, represented by the orange triangle in Fig. 3, was chosen using an offline global optimization of the controller to another optimal control profile, as determined using spline-based exoskeleton controller [8], [9]. Real biological torque data along with a built-in global optimization function in MATLAB (The MathWorks Inc., Natick, MA) was used to determine the shape and delay terms from (1) that minimized the root mean squared error (RMSE) between the optimal control profile from Franks et al. [9] and a control profile generated using the parametrization from (1). The final four pre-selected parameters (pink, light blue, green, and dark blue triangles in Fig. 3) were chosen to provide a broad sampling of the parameter space to seed the Bayesian optimization process. In each session, the order of the 6 preselected points was randomized to mitigate ordering effects and reduce bias between subjects. Note that these 6 control parameters were used to habituate the subject to the exoskeleton and controller during the habituation session, as well as to determine the scaling term α used in the mid-level parametrization. D. Bayesian Optimization We used Bayesian optimization to select the next fifteen iterations of assistance parameter combinations. Bayesian optimization was chosen as it has been found to be a time efficient method of HILO, and has outperformed previously used methods such as gradient descent [24]. Furthermore, Bayesian optimization suits HILO problems well as it is sample Fig. 3: (a) Graphical representation of the exoskeleton control parameter space highlighting the two optimization parameters, shape and delay, for level ground walking at 1.1 m/s. The grey dotted lines represent the minimum and maximum values of each parameter that bounded the human-in-the-loop Bayesian optimization. These values were determined during an N=3 pilot testing of the experiment (apart from the minimum delay which is limited by the time it takes to estimate biological moments). The triangles represent the six control parameter combinations that were used to initiate the Bayesian optimization for each subject and every ambulation mode. The starting points were selected as they represent a wide range of possible control profiles, as well as 2 points that are near optimal points found in previous studies. (b) The exoskeleton assistance torque profiles that are generated by the control parameters in (a). efficient and can directly be controlled to motivate exploration or exploitation to avoid falsely selecting local minima. When the exoskeleton control parameters were updated between iterations, the subject was allotted 30 seconds to adjust to the new parameters prior to beginning collecting data for the next optimizer iteration. Additionally, the subject's preference between the current and previous iterations' assistance was queried after each parameter change. Due to the metabolic mask required by the TrueOne 2400 Parvo system, the subject signaled better, worse, or no difference using hand signals during walking to avoid talking. Due to the physical demands of the data collection, subjects had predetermined scheduled break periods after every six iterations or by subject request and were allowed to rest until ready to continue. Custom MATLAB code using the Gaussian process regressor (GPR) package was created to run the human-in-the-loop Bayesian optimization process during the experiment. The acquisition function used was an expected improvement equation with an additional restriction so that a parameter combination too close to any of the previously tested parameter combinations could not be selected [25]. We restricted the optimizer to selecting delay terms between 65 ms and 240 ms and shape terms between .5 and 1.6 (Fig. 3a); these bounds were determined to be within reasonable range of comfort in pilot testing of the controller. The optimization model was set to be moderately explorative rather than exploitative so that similar parameters weren't tested repetitively and to improve model convergence within the 21 iterations (Fig. 4). E. Metabolic Cost Standard metabolic experiments typically require six minutes of data to compute metabolic cost [26], however due to the physical demand of our experiment, we collected two-minutes of metabolic data where the steady state metabolic cost value was estimated via a first order dynamic model applied to the two minutes of transient metabolic data, representing the current gold-standard within the field [27], [28]. This approach has previously been validated during metabolic and HILO studies that have shown that accurate estimates of metabolic cost require a
minimum of one minute of respiratory data [26], [27]. To make comparisons across both subject and walking conditions, we normalized the metabolic data by the no exoskeleton metabolic value using (2), where all metabolic analyses are presented in units of percent change from the no exoskeleton condition (%Δ No-Exo). For each subject and walking condition, we created metabolic landscapes using the controller parameters and their corresponding normalized metabolic cost from the 21 exoskeleton assisted iterations within each session using cubic interpolation and bounded radial basis function (RBF) extrapolation in areas that were outside of the convex hull of the data points tested. Each subject's best parameters, as approximated by their specific metabolic landscape, are denoted on individual condition surfaces as well as the mean best parameters for that condition. The average condition surface shows the overall best parameter set across all tasks, as well as the best parameters for each task averaged across all subjects. We used a one-way ANOVA in SPSS (IBM, SPSS Inc., Chicago, Illinois) to compare the metabolic cost across different best-case control parameters conditions: personalized control, task dependent control, and task agnostic control. We used the 27 metabolic landscapes determined from our experiment (9 subjects, 3 conditions) to find a single-best case parameter set, consisting of a shape and delay value, for each of the three conditions. Personalized control is both task and subject dependent, with best-case parameters corresponding to the lowest metabolic cost observed on the landscape for a specific subject during a specific task. Task dependent control is solely dependent on a specific task and is not personalized to the user, where parameters for a given subject are determined by averaging the metabolic landscapes from the 8 remaining subjects within that task. Task dependent control parameters are $$Norm. Met. = \frac{Met. -No Exo Met}{No Exo Met}$$ (2) Fig. 4: Convergence analysis of the human-in-the-loop Bayesian optimization across all three ambulation modes. Each plot begins at iteration seven after the six initial pre-selected parameter combinations. The model's convergence to the final (a) delay and (b) shape values in absolute percent change between the Gaussian process regressor's (GPR) believed optimal parameter and the actual value of the optimal parameter, i.e., the percent error at iteration 20 is equal to 0. For both the shape and delay parameters, the model was able to converge in fewer iterations during the inclined walking trials in comparison to the level ground trials. then derived from the corresponding lowest metabolic cost parameters on the task-specific subject-withheld landscape. Task agnostic control is independent of both subject and task, where a given subject's metabolic landscapes for all tasks are withheld and the landscapes from all remaining subjects across all tasks are averaged into one surface. Task agnostic control parameters are then derived from the corresponding lowest metabolic cost parameters found on the subject-withheld average surface. All percentage changes are with respect to the metabolic cost when the subject was not wearing an exoskeleton (No Exo), and all standard deviations are using a 95% confidence interval. Significant differences between both the no exoskeleton condition and between bars within an ambulatory condition cluster were determined using a one-way ANOVA with an alpha level of 0.5 and a multiple comparisons test with a post hoc Bonferroni correction. We compared between metabolic cost, torque, and power using linear regression, where we computed the slope, Pearson's correlation coefficient (r), and p-values generated to determine the significance between each condition and trend. F. User Preference We gave numeric values to the pairwise preference comparisons of the control parameters queried from the subjects during the optimization process. For a given chain of 21 control parameters and corresponding pairwise comparisons, every time a control parameter was rated as worse than another it lost one point of value and every time a control parameter was rated as better than another it gained one point of value. If a subject rated two parameter sets as equal, no points were either gained or lost for both parameter sets. Once every control parameter in each 21-iteration trial was given a numerical value, we created a preference landscape for each task and the overall task average, with a similar process of making the metabolic landscapes based on previously validated preference characterizations [29]. We then separated the parameter space into four quadrants to determine if controller preference could be reasonably predicted using this method. Across all modes, the 20 pairwise comparisons across 9 subjects were grouped by which two quadrants they were comparing, and the percentage of times the subjects preferred one quadrant to another was calculated. Furthermore, we computed Pearson's correlation coefficient between the all-task average preference surface and the all-task average metabolic surface to evaluate if higher reductions in metabolic cost trended with control parameters that were more highly preferred by users. G. Exoskeleton Power & Torque We analyzed how the power and torque provided by the exoskeleton trended with metabolic cost reduction. Average torque, peak positive power, and average positive power were
calculated using unilateral data from the exoskeleton's sensors. Torque $(\tau)$ is directly proportional to the motor torque constant, gear ratio, and current measured by the exoskeleton motors. Positive Power (+ P) was calculated using (3) and (4), where $\theta$ represents the angle of rotation of the motor. Similarly, peak positive power corresponds to the maximum of the positive power observed during the recording period. Average torque and average positive power were calculated by applying the average value formulas to both signals, shown in (5) and (6), where t is time and $t_{max}$ is the maximum time of the data recording. For a given iteration average torque, peak positive power, and average positive power have a corresponding normalized metabolic cost value to which they were linearly regressed. To analyze the metabolic cost individually and across subjects, we used a linear regression to compare the metabolic cost in percent change compared to no exoskeleton to the average exoskeleton torque, peak exoskeleton power, and average positive exoskeleton power. Within these comparisons, the slope of the linear regression, Pearson's correlation coefficient (r), and the p-value indicate trends between the metabolic cost and each presented metric. $$Power = P = \tau \times \frac{d\theta}{dt}$$ (3) $$+P = \begin{cases} if \ P > 0 & P \ else & 0 \end{cases} \tag{4}$$ $$\overline{\tau} = \frac{1}{t_{max}} \int_{0}^{t_{max}} \tau(t) dt$$ (5) $$\overline{+P} = \frac{1}{t_{max}} \int_{0}^{t_{max}} +P(t) dt$$ (6) Fig. 5: Comparisons between the subject-average metabolic benefit of the exoskeleton controller with respect to percent change from the subjects' baseline metabolic cost when walking without the exoskeleton. Results are shown averaged across all ambulation modes as well as for each individual mode. Each bar represents 9 subjects' metabolic cost at control parameters as determined by one of three methods. (i) Personalized control; (ii) Task Dependent control; and (iii) Task Agnostic control. A one-way ANOVA with a post hoc Bonferroni correction was used to determine the significant differences between metabolic cost within each ambulation mode, as well as the significant differences between each metabolic cost and the subjects' no exoskeleton metabolic value. Notably, across all tasks and on average there is no significant difference between (ii) Task Dependent control and (iii) Task Agnostic control. III. RESULTS A. Optimized Metabolic Cost vs. Subject-Task Dependency The personalized control approach significantly (p<0.05) outperformed the no exoskeleton condition, with the only exceptions being the task agnostic and task dependent conditions during level ground walking at 1.5 m/s (Fig. 5). On average across all tasks, the personalized controller condition reduced metabolic cost by 18.3% which was significantly better (p<0.05) than both the task dependent (8.6% reduction) and task agnostic (8.4% reduction) conditions. For the individual tasks, all personalized control parametrizations reduced metabolic cost significantly (p<0.05) as compared to the task agnostic condition. Personalized control reduced metabolic cost significantly (p<0.05) as compared to the task dependent condition during level ground walking at 1.1 m/s (10.3% reduction), but the difference was not significant during level ground at 1.5 m/s (3.95% reduction) and the inclined walking (11.62% reduction) tasks. On average and in each individual walking task, there were no significant (p>0.05) differences between using the task agnostic or task dependent methods. Average metabolic landscapes for each tasks and across all tasks are shown in Fig. 6, with each subject's best parameters and the average best parameters highlighted with colored triangles and stars, respectively. B. Preference Trends in the preference surfaces (Fig. 7) showed relative matches to those of the metabolic surfaces (Fig. 6) with a Pearson's correlation coefficient of r = 0.69, indicating a relatively strong positive correlation. Across all ambulatory tasks, subjects had major bias towards quadrants I and IV, representing longer delays, with reasonably high uniformity. However, subjects preferred smaller shape values, such as quadrant IV over quadrant I, about 69% of the time at the slower walking speed, and during the higher walking speed trial subjects tended to prefer the higher shape quadrant around 57% of the time. For both level ground trials, subjects showed a strong bias against control parameters that have both a low delay and low shape, such as in quadrant III. Across all
tasks, trends showed that users mostly preferred controller parameters within Quadrant IV, representing longer delay values and smaller shape values, between 59% and 74% of the time compared to other quadrants. C. Exoskeleton Power & Torque We used linear regression to observe the relationship between metabolic cost and exoskeleton assistance (Fig. 8). The linear regressor for metabolic cost and average exoskeleton torque had a slope of -1.379 percent-change per Nm/Kg (p<0.05), a Pearson's correlation coefficient of r = -0.2007, and Fig. 6: Average metabolic landscapes for (a) level ground walking at 1.1 m/s, (b) level ground walking at 1.5 m/s, (c) and 5° inclined walking at 1.1 m/s. The X and Y axes show the different exoskeleton parameters, and the color shows metabolic cost relative to the subjects' baseline, no exoskeleton metabolic cost. Each surface seen here is the average of all nine subject's individual metabolic landscape for the specified walking condition. Each triangle represents the parameters that correspond to the minimum metabolic cost of each subject's individual metabolic landscape. The white star on each landscape represents the parameters that correspond to the minimum metabolic cost of the average surface across all subjects. (d) Metabolic landscape averaged over all three walking conditions. This surface represents the results from all 27 metabolic cost surfaces (9 subjects, 3 walking conditions). The red triangles represent the parameters that correspond to the minimum metabolic cost of each walking condition, while the white star represents the parameters that correspond to the lowest metabolic cost on the average surface of all walking conditions. Fig. 7: Preference landscapes for (a) level ground walking at 1.5 m/s, (b) level ground walking at 1.1 m/s, (c) inclined walking at 1.1 m/s, and (d) the average of all tasks using the pairwise comparison queried from subjects during the HILO experiment. Higher values (purple) are preferred over lower values (red). We separated the parameter space into four quadrants to analyze the frequency of which subjects preferred certain quadrants over others. The black arrows superimposed over the surfaces denote the quadrant analysis, where the direction of the arrow indicates quadrant preference, and the number adjacent to the arrow gives the percentage that quadrant was preferred over other quadrants. a statistically significant p value (p < 0.05). The linear regressor for metabolic cost and peak exoskeleton power had a slope of 0.0067 percent-change per W/Kg, a Pearson's correlation coefficient of r = -0.0634, and a statistically insignificant p value (p > 0.05). Lastly, the linear regressor for metabolic cost and average exoskeleton power had a slope of 0.2985 percentchange per W/Kg, a Pearson's correlation coefficient of r = - 0.1653, and a statistically significant p-value (p < 0.05). IV. DISCUSSION This study investigated the effect of personalized exoskeleton assistance on the metabolic cost of 9 users across three cyclic tasks. Overall, our personalized control approach provided significant (p < 0.05) metabolic reductions across all three tested ambulation tasks when compared to both task dependent and task agnostic control approaches (Fig. 5), with the greatest metabolic reduction across all subjects and tasks accomplished during level ground walking at 1.1 m/s in which the subject average metabolic reduction was 21.1% when compared to the no exoskeleton condition. Our personalized control approach provided an average reduction in metabolic cost of 18.3%, whereas the task dependent approach provided an average reduction of 8.6%, thus verifying our primary hypothesis I. These results were expected based on previous literature that investigated similar personalized exoskeleton assistance via HILO when compared to generalized control parameters such as an end-to-end, deep learning driven biological torque-based controller [8], [9], [27], [30], [31], [32]. For example, the generalizable control approach presented in Molinaro et al. [16] on the same hip exoskeleton used in this study, provided a reduction in user metabolic cost of 5.4% and 10.3% for level ground and 5° inclined walking relative to the no exoskeleton condition, respectively. While this approach represents a novel method of exoskeleton control that can adapt to users in realtime, an essential requirement for translating exoskeletons out of the lab, the metabolic benefits of our personalized approach largely outperformed the purely generalizable controller across similar cyclic tasks, indicating the added value of personalizing exoskeleton control. Across all modes, there was no statistically significant metabolic difference between using non-personalized taskdependent and task-agnostic controller settings (Fig. 5), thus Fig. 8: (a) Multi-modal peak
exoskeleton torque for each subject. The goal was to keep the peak torque relatively constant across all iterations within each ambulation mode. Note: As subject two's exoskeleton data was corrupted for their slower level ground mode, that data was not included in this analysis. Across 9 subjects with 21 iterations of differing control parameters and 3 ambulation modes, and withholding subject two's missing data, there are 540 data points to relate exoskeleton torque and power data to subject metabolic cost. This data was used to regress metabolic cost versus (b) average torque provided by the exoskeleton, (c) peak power provided by the exoskeleton, and (d) average positive power provided by the exoskeleton. As expected, metabolic cost reduces as exoskeleton torque and power increase, however the p-values indicate that this relation is significant only for average exoskeleton torque and average positive exoskeleton power. rejecting hypothesis II. Overall, this indicates that for this type of biological moment-based hip exoskeleton controller in the absence of subject specific data, a controller setting that is task agnostic would be as robust as a controller that is task dependent across steady state, cyclic tasks. This is likely because in this controller paradigm there are areas of the landscape (specifically near Quadrant IV) that the user's metabolic cost has low sensitivity to small changes in the controller parameters (Fig. 7), indicating that using parameters from a multi-task average metabolic landscape does not have a drastic metabolic penalty as compared to using single-task metabolic surfaces. When comparing the average metabolic and preference landscapes across both tasks and subjects, areas such as Quadrant IV have considerable overlap, with a relatively strong positive correlation (r = 0.69), indicating that as metabolic cost reductions increase, user preference similarly increases. For example, the most preferred set of parameters on average were found in Quadrant IV (shape = 0.5 and delay ≈ 212.72 ms) corresponding to a metabolic cost reduction of 6.23% on the average metabolic surface, suggesting that further HILO studies could improve personalized control methods by optimizing for user preference as a proxy for metabolic cost. The controller parametrization used in this study was based on previously optimized spline-based hip exoskeleton control profiles observed during level ground walking and ramp ascent [8], [9], with our parametrization representing a profile that is flatter at times where the biological torque is near zero and sharper at peak biological torques. Offline optimizations based on this optimal torque profile showed that the addition of a shaping parameter, λ in (1), and a delay term to our parametrization enables our controller to further transform our estimated biological torque profile to a nearly identical profile to that of [9], with an R2 of 0.92 for level ground walking and R2 of 0.96 for ramp ascent. When tested on three expert-level users using a high-torque (~200 Nm) tethered exoskeleton with only hip assistance, this previously validated optimal torque profile provided an average metabolic reduction of 6.72% relative to a no exoskeleton condition [9]. Our approach consisting of personalized exoskeleton assistance when deployed on a fully autonomous hip exoskeleton capable of providing 18 Nm of hip flexion and extension assistance, and across three cyclic tasks and 9 subjects with only 15 minutes of training, provided an average 18.3% reduction in user metabolic cost, representing an 11.6% difference in reductions when compared to the high-torque exoskeleton used in [9]. This highlights the importance of personalizing exoskeleton assistance on autonomous devices to maximize the benefits of assistive technology when deployed in the real world. In our controller parametrization, we observed that the shaping term had a larger effect on the average power provided by the exoskeleton than the scaling and delay terms. This can be seen in Fig. 3, where lower shape values flatten the peak torques of the control profile, making the integration from (6) larger. Increasing positive power provided by an exoskeleton has been shown to improve metabolic cost reductions in users [22], thus optimizing the shape of the provided torque is highly important for personalized exoskeleton control. Likewise, when looking at the correlation between subject metabolic cost and average exoskeleton torque or positive power (Fig. 8), as exoskeleton average torque and power increases, metabolic cost significantly decreases, partially confirming our hypothesis III. It is interesting to note that peak exoskeleton power did not significantly correlate with metabolic cost, which may be due to the subjects trending towards higher average power as opposed to higher peak power delivered by the exoskeleton, i.e., as the shape term is increased there are higher peaks in the exoskeleton control profile, however the overall area under the profile's curve is reduced. Similarly, at the slower ambulation tasks, such as level ground and incline walking at 1.1 m/s, the average best parameters lie in a place where average positive exoskeleton power delivered to
the subject is very high. This benefit can be seen in the most physically demanding mode of the experiment at 5°-degree incline walking, where most metabolic reductions occurred at extremely low shape values during walking. Additional evidence is further reflected in user preference, where the level ground and inclined 1.1 m/s preference surfaces show quadrant IV having the strongest preference of all quadrants. There are several limitations to this approach. First, human subjects' experiments centered around metabolic cost are timeconsuming and difficult to collect. Due to the strenuous nature of our experimental protocol, we limited metabolic trials to two-minute iterations, as opposed to the standard six-minute trials to ensure subjects were able to complete the entirety of the protocol, which may have reduced the accuracy of our metabolic landscapes. Secondly, our experiment focused only on steady-state cyclic activities, such as level-ground and inclined walking, which represents a small subset of the dynamic tasks and activities that humans perform throughout the day. While our mid-level control approach significantly reduced user metabolic cost during steady-state cyclic activities, it is difficult to know whether these results change for different or non-steady state ambulation modes such as running, sit-to-stands, stairs, or lifting weights without further investigation. In particular, the insensitivity of the metabolic cost landscape of the task agnostic controller may have more sensitivity outside of the tested cyclic locomotion tasks and requires further studies on the topic. Lastly, our exoskeleton was limited by the device hardware to provide a maximum of 18 Nm of assistive torque at the hip joint. While studies focused on high powered tethered exoskeletons have shown increased metabolic benefits for users, these are inherently difficult to translate outside of the lab due to their size and weight. Our device was autonomous and portable, providing relatively close metabolic reductions to these tethered devices while allowing for a more lightweight and autonomous device. V. CONCLUSION This study used personalized exoskeleton control parameters derived from human-in-the-loop optimization to map the relationship between biological joint moment and optimal exoskeleton assistance. Overall, we found that our personalized control approach reduced the metabolic cost of walking by an average of 18.3% when compared to walking without an exoskeleton. Our results indicate that our personalized control approach can significantly reduce user metabolic cost across all three tested ambulation modes, as well as when compared to non-personalized task-dependent and task-agnostic control approaches. Furthermore, our results showed that there was no statistical significance between a task-dependent and taskagnostic control approach, suggesting that parameters derived from a multi-task metabolic HILO can potentially be used in a task-agnostic or generalizable control approach with minimal penalty to user metabolic cost. Our results further show that user preference has a relatively strong correlation with metabolic cost during steady state tasks, indicating the potential for future HILO studies to use user preference as a proxy for metabolic cost, thus avoiding time consuming and tiring optimization experiments. Furthermore, this approach can expand on end-to-end exoskeleton controllers such as a realtime biological joint moment estimator, to help improve and fine tune exoskeleton assistance for each user, representing a critical step to translating these devices to the real world.
REVIEW Open Access Review of control strategies for lower-limb exoskeletons to assist gait Romain Baud1 , Ali Reza Manzoori1* , Auke Ijspeert1 and Mohamed Bouri1,2 Abstract Background: Many lower-limb exoskeletons have been developed to assist gait, exhibiting a large range of control methods. The goal of this paper is to review and classify these control strategies, that determine how these devices interact with the user. Methods: In addition to covering the recent publications on the control of lower-limb exoskeletons for gait assistance, an efort has been made to review the controllers independently of the hardware and implementation aspects. The common 3-level structure (high, middle, and low levels) is frst used to separate the continuous behavior (midlevel) from the implementation of position/torque control (low-level) and the detection of the terrain or user's intention (high-level). Within these levels, diferent approaches (functional units) have been identifed and combined to describe each considered controller. Results: 291 references have been considered and sorted by the proposed classifcation. The methods identifed in the high-level are manual user input, brain interfaces, or automatic mode detection based on the terrain or user's movements. In the mid-level, the synchronization is most often based on manual triggers by the user, discrete events (followed by state machines or time-based progression), or continuous estimations using state variables. The desired action is determined based on position/torque profles, model-based calculations, or other custom functions of the sensory signals. In the low-level, position or torque controllers are used to carry out the desired actions. In addition to a more detailed description of these methods, the variants of implementation within each one are also compared and discussed in the paper. Conclusions: By listing and comparing the features of the reviewed controllers, this work can help in understanding the numerous techniques found in the literature. The main identifed trends are the use of pre-defned trajectories for full-mobilization and event-triggered (or adaptive-frequency-oscillator-synchronized) torque profles for partial assistance. More recently, advanced methods to adapt the position/torque profles online and automatically detect terrains or locomotion modes have become more common, but these are largely still limited to laboratory settings. An analysis of the possible underlying reasons of the identifed trends is also carried out and opportunities for further studies are discussed. Keywords: Exoskeleton, Lower-limb, Control, Review Introduction Powered lower-limb orthotic devices, also called powered exoskeletons, are often considered as tools in rehabilitation and the assistance of the human gait. A signifcant amount of research in diferent felds has been dedicated to developing and improving the performance of these devices, and there are many challenges in this area of Biorobotics laboratory (BioRob), EPFL, Lausanne, Switzerland Full list of author information is available at the end of the article *Correspondence: ali.manzoori@epf.ch research due to inherent requirements of portability and safe interaction with the user and the environment. One of the most important aspects for improving the performance of these devices is their control [1]. Currently, there are two main types of exoskeletons for gait assistance: the ones for full mobilization, and the ones for partial assistance. Full mobilization exoskeletons are designed to move the legs of people sufering from a severe loss of motor control or motor disorders, typically in people with spinal cord injury SCI. Te actuators must have a high torque capability because they provide the entire torque required for the movement. Such devices are available commercially since 2011, when the ReWalk (ReWalk Robotics, Israel) was released on the market. Tey could be developed quickly because their control strategy can be simply position control over time. Tere is no need to collaborate with an existing voluntary movement of the legs, because there is none (or it is very weak) and thus the user's legs are assumed to be passive. Te start of the gait is often triggered by the upper body movements or buttons pressed by the fngers, which is simple to implement. Tese exoskeletons seem more successful because they dramatically improve the bipedal ambulation capability (from no gait at all to some slow gait). Partial assistance devices are generally lighter, targeting various less severe handicaps. Tese could be the loss of stamina because of aging [2], the loss of strength or coordination because of incomplete spinal cord injury SCI, stroke, neurodegenerative diseases, etc. Tese devices can also assist the gait of healthy people, which can be useful for endurance augmentation purposes. Tis is more challenging because the device has to assist more than it is hindering its user, given the complex nature of the interaction with the user. People who can already walk independently also have higher expectations for the performance (e.g. higher gait speed). A major
subcategory of partial assistance exoskeletons are the devices that are intended for rehabilitation purposes.1 Here, the ultimate purpose is to train the users to become independent of the assistance ofered by the device. A fundamental distinction can thus be made between the desired outcomes of these exoskeletons versus the ones that are used to directly assist the mobility only when wearing the device. Actually, a training strategy for rehabilitation may consist in resisting the user movement [3]. Notwithstanding this diference in the end goals, there is a lot of commonality between the two applications in terms of the techniques used for control. Several reviews already exist on diferent aspects of exoskeletons and gait assistance devices, but very few are focused on control. Te two most exhaustive reviews of control strategies to date are the ones of Tucker et al. [4] and Yan et al. [5]. However, these are already 5 years old at the time of writing this paper, and many new developments deserve to be mentioned, since this feld is evolving fast. More than 190 new publications addressing control strategies have been identifed since the publication of the two previous reviews in 2015, and advancements have been made with new control methods and device designs, resulting in major performance improvements in terms of metrics such as metabolic cost reduction and capabilities such as crutch-less dynamic walking. Te review of Tucker et al. is broad and considered both orthoses and prostheses. A "generalized control framework" was proposed with a 3-layer hierarchical controller, and also the environment, the user and the hardware of the device. But this review did not provide much detail on the mid-level layer of control. Te article by Yan et al. focuses on the control of exoskeletons and orthoses, but it is mostly organized around the devices themselves, and how they are built (e.g. single/multi joints). Some reviews have also been recently published on gait assistance devices [6–9], but none of them comprehensively address the control aspect. A recent review by Sawicki et al. [7] focused on comparing the results of partial assistance for the gait, and only considered the successful orthoses with respect to metabolic cost reduction. Tis excludes all the devices that did not undergo such testing and also full mobilization exoskeletons. Also in this article, few details are given on the details of the control part. A more broad review by Kalita et al. [8] studied the existing exoskeletons and orthoses in the literature, categorizing them according to joint structure, actuation and control strategy. Control strategies are roughly divided into 9 categories, each one only briefy explained without going into the details. In this review, the various control approaches of gait assistance devices are thoroughly addressed, focusing on the lower-limb exoskeletons designed to enhance the locomotion of disabled or healthy people. Compared to the existing reviews, a stronger emphasis is placed on the control methods and separating them from the hardware and implementation details as much as possible. Based on the existing control methods in the literature, a modular classifcation framework consisting of 3 layers is proposed. Te purpose of the framework is to enable describing all of the existing control strategies with the minimum number of functional elements. Tis paper also shortly reviews the metrics used to characterize the 1 Although for severely disabled patients, full mobilization is also used in the early stages of rehabilitation, most rehabilitation devices ft in the partial assistance category better. performance of these robots when worn by a user. However, the assessment of the performance of the cited controllers and their comparison are beyond the scope of this review. Assistive strategies From the control perspective, the main challenge for gait assistance is to contribute to the intended movement, since the device cannot directly communicate with the wearer to clearly recognize the intention and collaborate efectively. Efective collaboration can be interpreted in diferent ways, depending on the context and application. In general, for partial assistance it would mean synergy in forces or torques between the user and the device, and for full mobilization it would be coordination between the movements of the exoskeleton and those of the user's upper body. Many strategies are used to identify the user's intent, and apply an appropriate torque or motion accordingly. In the rest of this section, the existing strategies will be reviewed and discussed. Before getting into the review of these strategies, the rationale behind the criteria that were used for screening the literature and the proposed classifcation method will be explained, and the methodological steps will be described. Methods Scope and methodological steps Te main question to be addressed in this part is: what approaches have been used in the literature up to now for controlling lower-limb exoskeletons with the purpose of directly assisting the wearer's gait? Target devices for the controllers in this review do not need to provide an improvement of the user's health. Although the devices are typically anthropomorphic, exceptions also exist (such as [10–13]). Te so-called "soft exoskeletons" (ex
osuits) are included too, even if these are not really stif "skeletons", but closer to "tendons and muscles". Te papers that do not deal directly with an exoskeleton, but suggest a sensing method that could be useful for them are included as well. As explained previously, many gait assistance devices are presented in the context of rehabilitation. In light of the similarities from the control perspective, we did not limit the scope of this review to a specifc application; as long as the described controller is supposed to assist the user during gait, the method was included in this review regardless of the long-term goal. Tis review aims to address wearable gait assisting exoskeletons, because they have the potential to be used for real-life applications out of the laboratory. However, the articles involving fxed-frame devices designed to explore such control strategies (e.g. LOPES [14], ALEX [15], the exoskeleton emulator of Collins et al. [16], etc.) are also included in this review. In addition, if at least part of the control strategy proposed for a fxed-frame rehabilitation device also assists the user's gait and is applicable to assistive exoskeletons, it is included (for example [17]). Te strength augmentation devices are excluded because they are not designed to enhance the walking mobility. Te main consequence is that they are of no use for people afected with gait defciencies, or healthy people willing to improve their ability to walk (higher speed and/or endurance) with no load. Tey also mainly focus on load lifting so the control strategies may be diferent, and may also involve upper limbs. Te task of carrying a load while walking (e.g. [18]) is closer to the topic of this article, but such devices still do not assist in moving the user's legs or relieve the user from the bodyweight. In addition, it makes comparing the performance even more difcult, because the assistance beneft depends on the amount of payload. However, a strength augmentation device that would enable its wearer to jump higher or run faster would have been included, but such reference could not be found. Similarly to the fxed-frame rehabilitation devices, a strength augmentation device can be still be included if at least part of the control strategy could be applicable to the assistance of the gait with no carried load (e.g. [19]). Te inclusion and exclusion criteria used in the screening process of this review are summarized in Table 1. Most of the publications were found using the following Google Scholar query:robot* assist* control* (exoskele* OR orthosis) and a similar query on Scopus: "robot* assist" "control" AND ( exoskele* OR orthosis ) AND NOT ( "upper limb" OR "upper-limb" OR "hand exoskelet*" ) among the records published since January 2000 up to the end of August 2020. Te references cited in the two previous review papers by Yan et al. [5] and Tucker et al. [4] were also included. First, the references were screened with the title, then the abstract, and fnally the full-text to check if they ft the inclusion/exclusion criteria. Ten, they were read entirely and entered in a database. Te relevant articles cited by the ones already in the database were also added. A fowchart of the methodology is shown in Fig. 1. For each entry in the database, the following felds (as long as they were relevant/applicable) were entered: highlevel control method, mid-level control method, lowlevel control method, type of actuator, short controller description, intended application, assisted joints, device name, and remarks. Proposed classifcation Tis review is centered on control strategies, being hardware-agnostic as much as possible. To be accurate enough in describing the diferent control strategies features, but Table 1 Summary of the inclusion and exclusion criteria used for screening the articles Includes description of controller(s) applicable to lower-limb exoskeletons with the purpose of helping wearer's gait, or detection methods applicable to such controllers Inclusion criteria Exclusion criteria Describes a controller that is specifc to other devices such as prosthetics, upper-limb exoskeletons, fxed-frame rehabilitation devices (such as Lokomat [326] and MotionMaker [327]) or portable devices that operate as external units rather than wearable robots (e.g. WalkTrainer [328]) or devices that were designed only for animals Date of publication: January 2000 to August 2020 Describes a controller for devices that assist locomotion by using another movement than the natural movement of the human leg (such as rolling devices, jumping stilts in which the blade moves below the foot, jetpacks [329] and portable inertial devices [330, 331]) Language of publication: English Describes a controller that is intended for assisting load-carrying or strength augmentation without signifcantly afect
ing the gait itself Type of publication: peer-reviewed journal or conference article, patent Describes a controller that is impossible to apply outside of a simulated environment Does not give enough details about the control method to fully describe it (typically the case for papers reporting clinical trial outcomes) Gives inconsistent information about the controller and/or the device Only reviews control methods with no redundancy in the descriptions, it was chosen to break the behavior into smaller functional units. Indeed, an initial assessment of the literature revealed that even among diferent control strategies, shared elements exist. Compared to describing each control strategy as an atomic entity, this classifcation method allows for reusing the same elements to represent several strategies. Te literature shows a considerable number of diferent controllers, with diferent structures, designs and actuation methods. However, the ultimate requirements in terms of performance and desired behavior are mostly similar. In an attempt to classify them, we will separate the controllers into smaller functional units that are comparable. Each functional unit can be used in several diferent combinations to form various controllers. Terefore, these functional units can be considered as the building "blocks" of the controllers. Based on their role in the hierarchy of the control system, all of these blocks can be classifed into three categories: high-level, midlevel and low-level control (see Fig. 2). Tis hierarchical classifcation is similar to the one used in [4]. Within each level, various methods and approaches thus form the diferent blocks. Some of the blocks within the same level perform the same function (in terms of outputs) using diferent methods, while others have a dissimilar functionality. Hence, even though the blocks in diferent levels may be used together, they are not always compatible. All of these blocks are shown in Fig. 3 and will be explained in detail later in the paper. It should also be noted that the reviewed control strategies do not necessarily cover all the three levels, with most of the research being focused on mid-level control. Tis review will then focus on mid-level control mostly. Our analysis of the high- and mid-level layers is also implementation-agnostic, which means it focuses on the external behavior of the device rather than the way to program it or make the hardware design. Most of the hardware-specifc aspects will be separately discussed in the low-level layer. Te results obtained by all these controllers are not compared, because the target users are diferent (healthy, elderly, paraplegic, stroke, etc.), the tasks are diferent (walking, running, ascending stairs, etc.), and even for the same task, the experimental protocol is often diferent. Such comparison is possible, but only with a narrower scope. For example, the review of Sawicki et al. [7] focuses on the partial assistance for the gait, to decrease the metabolic cost of locomotion for healthy people. High‑level control Te high-level control determines the general behavior of the exoskeleton. Exoskeletons can usually switch between several operating modes, depending on the desired type of activity, and the environment (e.g. walking on fat terrain, climbing stairs, and sit-to-stand transitions). Often, this change of mode does not occur frequently, and there is typically a gap of at least several seconds between two consecutive changes. Tis makes it possible to be selected by the user. Relatively few papers are dedicated to high-level control. For most research purposes, the focus is on a certain mode of operation, and the experiments take place in controlled lab settings and are based on well-defned scenarios. However, reliable high-level control is crucial for the usability of exoskeletons for people in real-world situations and everyday life, where a variety of movements and gaits in diferent environments and terrain types are required and short transition times are necessary. Te inputs to the high-level controllers can come from the user (via input devices and/or sensors), the environment, or a combination of both. Te output is usually a mode of operation. Artifcial intelligence and machine learning methods are being increasingly used as a substitution for the user choice. Te main motivation is to make the operation more automatic for the user, and possibly faster than manual input. Fundamental criteria for the usability of such methods are the real-time operation and short processing times, since decisions need to be made fairly quickly to allow enough reaction time for the lower-level controllers. Existing high-level control strategies are discussed in more detail below. Explicit/manual user input (MUI) Te user directly determines the mode of operation of the exoskeleton, using input devices such as buttons [20–34] or voice commands [35, 36]. Tese methods are currently the most common due to their ease of implementation, higher predictability, and lower risk of errors. However, these advantages come at the cost of additional participation required from the user, which makes the user experience less natural, increases the cognitive load, and can slow down the operation. Moreover, this method is also prone
to human errors which are more likely to happen during demanding tasks, long operation times, or with novice/distracted users. In this case, the challenge is both to make the user interface easy to use to minimize the learning time and the risk of manipulation errors, and also quick to use to avoid losing time in transitions. Tis is not trivial since the interface has to be used in a standing position, and the hands often have to hold crutches at the same time. Te explicit user input is commonly used in full-mobilization exoskeletons for complete spinal cord injury (SCI) patients, because no input can be obtained from the legs. It is also the most predictable for the user, which is important for trusting the device. In this case, buttons on the crutch handle, or a special wristwatch can be used. Voice command is not common because it requires speaking, which may feel awkward in public spaces. It is also more error-prone in noisy environments. Brain‑computer interface (BCI) Te user's brain activity is measured using electrodes, amplifed and analyzed to determine the mode of operation [37–39]. Among the diferent brain signal recording methods, currently electroencephalography (EEG) is predominantly used since it is non-invasive and therefore Fig. 3 Block diagram of the proposed classifcation of the control strategies subparts. The idea of this classifcation is that any controller in the literature can be represented by a path that joins the used control blocks. The path does not have to start from the high-level layer, and may start directly in the mid-level. A controller can have several parallel paths if the controller combines several strategies at the same time, or successively during the gait. Connecting lines show the common paths identifed in the literature. However, it should be noted that the lack of a line between two blocks does not mean they cannot be related. For instance, the outcome of the high-level layer, the "operation mode", could afect any of the blocks of the middle-level, but it is not connected to them for the sake of readability safer and easier to use. Despite the promising features of these methods, there are many practical challenges associated with them, including high levels of concentration required from the user (and therefore limiting simultaneous cognitive activities such as speech), artifacts with muscular activation (EEG signals at the surface of the scalp have an amplitude close to 100 μV [40], while electromyography (EMG) signals are several millivolts), rather lengthy procedures for electrodes placement, the need of training for the user and the algorithm, and being very slow (in the order of seconds) or limited to very few commands [39, 41–44]. A thorough review of braincomputer interfaces BCIs for lower-limb gait assistance devices in general can be found in [45], and an in-depth review of methods based on EEG in [46]. Movements recognition (MOV) Tis type of controller changes the behavior automatically depending on how the user moves or is intending to move. Te main advantage of this method is that it does not require any cognitive load or direct input from the user, making the interaction more intuitive and natural. For this method, generally joint sensors and IMU data (often from the upper body in persons with paraplegia) are processed by a machine learning or fuzzy logic algorithm to recognize the situation [47–64], although simpler threshold-based methods have also been proposed [65]. Sometimes, other types of signals such as the ground reaction forces or electromyography (EMG) are also used to infer the movement or the intention of the user [66–71]. Capacitive electromyography was also investigated [53]. In practice, often additional inputs are also required to complement these controllers (e.g. to disable them when the user needs to perform other activities while standing still in the device) since the movements of the user are not always sufcient to correctly determine the intention. In [72], the discrimination between walking and jumping is performed with a threshold on the phase diference between the two legs (shank segment), computed with the angle-speed diagram. Moreover, standing is detected if the magnitude of the phase vectors for the two legs is below a certain threshold. Terrain identifcation (TER) Generally, the most decisive factor in determining the mode of operation and high-level behavior of gait assistance devices is the terrain. Information about the terrain can hence be used to construct a high-level controller for such devices. In these controllers, embedded sensors are used to recognize the terrain type or obstacles in front of the user, in order to plan the steps accordingly [73].2 Sensors used for these high-level controllers are most often cameras (either usual visible-light cameras [74, 75] or 3D depth-sensing [41, 76–82]), but other sensors such as infrared distance sensors [83] or fusion of laser distance sensors and inertial measurement unit (IMU) [73, 84, 85]
have also been utilized. Terrain identifcation has recently gained attention in the felds of orthotics and prosthetics, and the body of literature exploring it is relatively small. Even the existing papers are limited to proof of concept implementations, demonstrating the performance of terrain identifcation algorithms without actually integrating them into the high-level controller of a device [73, 75, 79, 80, 83–85]. Tese techniques are usually computationally expensive because of the image or point-cloud processing. However, promising results have been demonstrated and with the advances in pattern recognition and machine learning methods, successful implementations of such controllers are to be expected in future research. Mid‑level control Te mid-level is defned here as the continuous behavior of the robot, which computes the joints target torque or position, at each timestep of the main control loop. Te mid-level controller plays the most important role in shaping the interaction of the device with the user, and the majority of the research on the control of exoskeletons is dedicated to this level. Although the output of the high-level controller also afects the behavior, it often only changes some parameters of the mid-level controller without fundamentally altering the essence of the interaction with the user. In the proposed classifcation, the mid-level control blocks have been separated in two sublayers. As shown in Fig. 2, the "detection/synchronization" sublayer estimates the gait phase or gait state, which is a piece of information commonly needed by the "action" sublayer that actually computes the motor command. Te frst sublayer uses external inputs (from sensors and/or user interface) to determine the continuous phase or discrete state of gait. In the second sublayer, the desired physical output of the device is decided. An exoskeleton controller can have a diferent control scheme for each joint. Tis is for example the case in [19], in which a simple spring is used for the ankle, an active damper for the knee, and torque control on the hip joint. Another example in [86] is an adaptive-frequency 2 Many of the papers cited in this section are taken from prosthetics literature, but the fact that the terrain identifcation systems have been designed with a prosthesis in mind does not afect the outcome and all of the results are equally applicable for orthoses as well. oscillator AFO-based impedance control for the hip, fxed position or zero torque control for the knee (depending on stance/swing), and event-triggered torque sequence for the ankle. Detection/synchronization sublayer Te desired outcome of this sublayer is either the accurate gait phase (0–100%), or the gait state. Gait states are generally subphases of the gait cycle (e.g. stance/swing or fner divisions such as loading response/foot-fat/pushof), the kind and the number of which depend on each controller. Manual trigger by user (MAN) Tis lets the user explicitly trigger the movement. Tis block is usually followed by the "Linear increase of the gait phase" and "Position profle". Tis method is simple and used frequently to trigger the steps of a full mobilization exoskeleton. Te trigger is generally a button ([20, 24, 26–28, 31, 87, 88]), but steps can also be triggered by EEG [42], although very slowly. It is worth mentioning that controllers in which the user manually triggers the start and stop of locomotion (and not the individual steps) such as [89, 90] do not belong in this category. Impose the movement (IMP) Instead of synchronizing to the user, the robot imposes the movement continuously. So, it is the user's responsibility to stay synchronized with the robot. Tis is sometimes the case with early-stage full-mobilization exoskeletons that test the continuous gait without providing a user interface to use them in real-use conditions [91–95]. Other common cases are brain-computer interface (BCI)-controlled exoskeletons that do not need crutches, with start and stop commands instead of having to trigger each step [41, 43, 44]. As opposed to the rest of the blocks, this one does not represent an actual function in the controller, nor does it have an output for its following block. Rather, this block is only used to emphasize the lack of synchronization. It is always followed by "Simple linear increase of the gait phase", which then usually feeds the "Position profle" or "Torque profle" blocks. Event trigger (EVT) Tis method can be found in many exoskeletons for partial assistance and full mobilization. It consists in using an event of the gait to start a step, a torque profle or to transition a state machine. Te most common event is the heel strike, detected with a foot switch at the heel or (rarely) with an instrumented treadmill [96–107]. If the pressure sensor is located under
the forefoot, the late stance can be detected instead of the heel strike [108]. Te reference instant can also be recognized with an inertial measurement unit (IMU) on the shank, when crossing the zero angular speed [109]. A variant is to detect the point of "negative-to-positive power" of the ankle by looking at the ankle speed (one IMU on the foot, one IMU on the shank) [110], or with a classifer [111]. An alternative is to use an inertial measurement unit (IMU) in the foot sole [112–114]. Similarly, it is possible to detect the lift-of [48, 115]. A set of thresholds on the "analog" ground reaction force can also be used to discriminate several phases in the gait cycle [116–118]. Events in the kinematics can also be used. Te peak value of the hip angle is used in [119–124], or similarly the peak ankle dorsifexion angle [125]. In [126] the state machine is transitioned with thresholds on the knee angle and velocity. In [10], there is a threshold on the time-derivative of the pressure of the passive pneumatic actuator, which relates to the joint speed. In [127], a hidden Markov model is used to detect the gait phases from trunk and segment angles measured with an inertial measurement unit (IMU). For full-mobilization exoskeletons, the steps can be triggered by weight shifting measured by the load cells under the feet [21, 128–130], by leaning toward the front or on the sides which is measured by the inertial measurement unit (IMU) [30, 128, 131], with a combination of the crutches load cells and the feet load cells [32], or a combination of the trunk tilt and the feet load [29]. Adaptive frequency oscillators (AFO) AFOs are dynamical systems with an oscillatory behavior that are capable of learning the features of a periodic input signal [132]. Due to the periodic nature of the gait, they can be used to determine the gait frequency and the phase. Tey can adapt quickly to a change of cadence, and do not need any prior knowledge on the shape of the gait pattern, except the fact that it is periodic. Tis makes them robust and makes the controller suitable for almost any user without the need for extensive parameter tuning or gait pre-recording. AFOs are usually fed with joints angles, but can also be used with any other periodic signal, such as the muscular activity, estimated using capacitive sensing [133] or interaction forces between the device and the user [134, 135]. AFOs can produce several useful pieces of information: the current progress in the gait cycle (0-100%), the frequency, and a fltered version of the input signal with no lag. Actually, the whole trajectory over the full gait cycle is modeled by the adaptive-frequency oscillator (AFO). Tese can be used in further action blocks, typically "Torque profle", or "Impedance control". Te output has occasionally been directly used as a position reference as well [134, 135]. While AFOs are able to compute precisely the frequency and the joint angle value function over the gait cycle, the reference moment (usually the heel strike at 0%) is unknown so the absolute gait cycle progress cannot be determined. Several techniques exist to solve this issue: Foot switches can measure the instant of the heel strike [51, 136, 137]. Tis method is accurate, needs no heuristics, but requires an additional sensor. An inertial measurement unit (IMU) can also be used instead [138]. A special feature in the joint trajectory (e.g. minimum or maximum value, or maximum slope) at a known gait phase can also be recognized, but this is subject-dependant and less reliable [122] (and probably [50]). Instead of a sine wave as the frst harmonic, a known average human gait trajectory can be used [139–141]. Tis is less accurate if the user is walking in a nontypical way. Such an oscillator is called "PSAO" (particularly shaped adaptive oscillator) by the development team of the GEMS exoskeleton [139]. Finally, strategies that do not use the absolute gait cycle progress can be selected, so that there is no need to obtain this information. Tis is the case for force felds that attract the joint towards its predicted position [56, 86], or compensation from a physical model (weight, inertia) [142].3 Note that "attracting toward the predicted position" is equivalent to using an impedance controller with the AFO-identifed movement with a time ofset (to follow the future) as the reference. In [143], AFOs are also used, but the reference determination method is not explained. Te Honda Stride Management Assist is also using a special AFO method according to a patent [144
], but the details are not clearly documented. Te AFOs strategy is limited to the partial assistance paradigm, since the user needs to be able to initiate the gait and maintain it at least for a few steps. Simple linear increase of the gait phase (LNP) Tis is the simplest way to determine or impose the gait phase. It consists of increasing linearly the gait phase over time, knowing in advance the step duration. If the movement is imposed all the time (IMP), the gait phase is looping continuously [38, 90, 91, 145]. If triggered manually [26, 35, 87] or with an event such as foot contact with the ground [97, 101, 146], lateral weight shifting [89], tilting the trunk [30, 147], or muscle activation (sensed via Fig. 4 Example of angle-speed phase diagram. The data plotted is the hip angle during a few gait cycles of a test session with the exoskeleton SPRIINT (see [325]) electromyography (EMG)) [82], it only runs once per trigger. Te output of this block then feeds a position or torque profle. Time-interpolated gait phase (TBP) Tis is the same as LNP, except that the gait cycle duration is determined automatically from the duration of the previous steps. Tis is very accurate if the gait is periodic and with a small inter-step variability. Tis method is very common for partial assistance [16, 96, 98–100, 102, 103, 105, 106, 108, 110, 113–115, 121, 123, 125, 148–155]. An extension of this method is to use a Gaussian probability density to reject outliers [156]. Angle-speed plot phase (ASP) Tis technique consists in determining the gait phase from the angle and speed of a single joint. Intuitively, the function that maps a joint angle to the gait phase is surjective but not injective, because there are at least two solutions, due to the backand-forth movement. So at best, if the joint trajectory is not bouncing, there are two possible gait phases for a given joint angle. However, the speed has the opposite sign for the way back, so it gives enough information to disambiguate the gait phase. In practice, these states are plotted on an angle-speed graph, and the phase angle can be extracted (Fig. 4). Te center of the trajectory must be defned by prior calibration. Te main advantage of this method is that it keeps its accuracy even if the gait cadence changes rapidly. However, it is very sensitive to bouncing, and is inaccurate if a joint moves little during part of the gait cycle. Tis is why it is not used with the knee joint. Tis method is used in [72, 157]. Machine learning phase (MLP) Te gait phase can also be estimated using machine learning, with techniques such as support vector machine (SVM) or neural networks. Te machine learning methods are diverse and complex, so they will not be explained here. All the found 3 Tis second variant should technically not be part of this review, because it was proposed for an upper-limb exoskeleton, even though the paper also addressed lower-limb assistance with another controller. It could probably be applicable to lower-limbs also, but no publication could be found with this principle. references used a diferent machine learning method and diferent inputs. A neural network fed with the trunk IMU data and hip encoder angles is used in [158]. An online Gaussian process regression is fed with the joints angles and interaction forces with the thigh cufs in [59]. In [159], the gait phase is estimated with a decision tree, from the segments IMU data and the feet loads. In [160], deep learning is used on the shank and thigh IMU data and feet loads. In [111], a SVM is used with the shank IMU data. In [53], a quadratic discriminant analysis allows to get the gait phase from capacitive sensors measuring the thigh muscles contraction. Finally, in [161], a computer vision classifer can estimate the gait phase from the data of depth cameras located on the crutches. Other gait phase estimator (OTP) Te gait phase can also be estimated by other less common methods. One of the controllers proposed in [162] ("State Estimation" controller in the paper) estimates the gait phase by ftting the recorded joint angles and the foot loads to a reference model, using least-square regression based on the method from [163]. In addition to this method, another variant is also suggested for comparison in [163], which determines the gait phase based on minimizing the squared error between the instantaneous ankle angle and contact forces at toe and heel with those of a reference model (the frst method is called "cross-correlation" and the second "k-nearest neighbors"
in the original paper). However, the estimated gait phases have not been used in a controller, but have only been compared to evaluate the estimation accuracy. State machine (FSM) Controllers can switch behavior depending on transitions triggered by events. Tis may be useful because some states of the gait are noncontinuous. Te best example is the foot contact, which is binary (swing/stance) and changes the dynamics of the leg. Many controllers use a state machine and diferent criteria have been utilized for transitioning between the states. Most commonly, the ground contact state of the feet, or equivalently the ground reaction force (GRF), is used either for the entire foot to only distinguish between stance/swing [164–169] or considering local components (e.g. at the heel and under the toes) to further diferentiate between stance subphases [48, 67, 86, 117, 148, 162, 163, 170–174]. Te gait state can also be determined by computing the center of pressure (CoP) position of the stance leg with four load cells per foot, then applying a threshold to identify four states [175, 176]. In one paper, the subphases of stance were detected only based on the total ground reaction force (GRF) [116]. For some state machines, the ground contact status has been used as the only factor for transitioning the states [67, 117, 162–166, 171, 172], but it has also been used in combination with joint angle(s) [116, 167, 170], joint angular velocities [173], segment angles and angular velocities [48, 168], or the relative position of the feet [177]. In [148], the linear acceleration of the shank is also used in addition to ground contact data to improve the accuracy of heelstrike detection. Te amount of time elapsed since the onset of swing has also been used in addition to ground reaction force (GRF) data to further detect subphases of swing [169]. Joint angles and angular velocities have also been used without the ground contact information to transition states [126, 178–182]. In [47], in addition to the angle and angular velocity of the knee joint, the moment at the joint and the angular velocity of the leg are involved in state transitioning. Te authors in [19] have augmented joint angles with the forces and moments sensed in the exoskeleton segments to transition the state machine. In an alternative method, the diference between left and right joint angles (hip and knee) are used along with zero-crossing events of hip angular velocity to transition between the states [151, 152]. In [10], thresholds on the derivative of the pneumatic actuator pressure (which indicates the direction of movement intended by the user) are used for the transitioning. Surface electromyography (EMG) has also been used as another indicator of user's intention to transition the states [66]. In [89], the estimated projection of the center of mass (CoM) on the ground relative to the feet is mostly used to transition between the states, but direct user input (via buttons) is required for transitioning in and out of the initial and fnal states, while transitioning between others (e.g. between shifting the weight to the stance leg and swing of the opposite leg) is initiated automatically. Diferent states may only change the parameters and/ or inputs to a controller (for example [48, 89, 117, 172, 183, 184]) or change the control strategy completely (for example [19, 61, 66, 86, 173, 185, 185]). It is also worth mentioning that sometimes the state machine does not involve any electronics, and is implemented using mechanical components only [164, 178, 179]. Action sublayer Te goal of this second sublayer is to generate a motor command, that can either be kinematic (angle or speed), or kinetic (torque or force). Position profle (PPR) Te goal of the position profle is to assist the user to move according to a predefned trajectory, supposed to be the intended one. Te trajectories can be described in joint space or Cartesian space, often called "foot locus" for this second case. Tese trajectories are usually completely predefned based on recorded gait data from healthy people [48, 89, 91, 145, 186–188]. Databases of recorded trajectories from diferent healthy people have also been used in some strategies, where the controller chooses which trajectory to use depending on the situation [67]. In another approach, the trajectories have been recorded as a therapist manually guided the subject's legs to achieve a desired gait pattern [187]. In [189], recorded trajectories from each subject walking in the exoskeleton in passive mode are averaged and used as reference. Some small modifcations are generally necessary to account for user-specifc and device-specifc differences before
actually using the trajectories recorded from healthy people for patients. In many cases, the trajectories are signifcantly changed or fully generated at runtime, and some papers are completely dedicated to the problem of optimization/ generation of trajectories [190–193]. In some studies, model-based computations [194–197] or polynomial minimum jerk trajectory generation methods [94] have been used to generate the trajectories ofine. Trajectories can be generated so as to reach a certain target position/ orientation in task space as well [191, 198]. For simpler implementations, the trajectory may also be defned approximately by a fnal target angle and a speed limitation instead of the complete path, and has been used for pneumatically actuated exoskeletons [88, 174]. However, the trajectories are not necessarily fxed or predefned. Online modifcations can be applied to the baseline trajectories, as is the case in [199] and for the hip trajectories in [20] and [89] (only abduction/adduction angle in the latter). In [181], the user is free to move the legs during stance, and the baseline swing trajectory (from healthy subjects) is adapted at every step to match the leg confguration at the end of stance. More advanced methods have recently been proposed to automatically adapt the recorded gait trajectories from healthy people to the environment, and generate new trajectories for different types of terrain [82]. Te trajectories could also be generated online, for example synthetic and parametrized trajectories can be used to adapt the foot clearance, step length and duration, peak joint fexion, etc. [77, 200]. Te authors in [130] have proposed to generate the leg movement online to match the step length measured by a walker which is moved manually by the subject. In [192], a method is proposed to calculate the joint trajectories as a function of the movement of the crutches by the user's arms, based on synergies extracted from the data of healthy subjects walking with crutches. Some controllers that are based on AFOs predict the joint trajectories online based on the estimated gait frequency and phase [190], and the future positions could be used as the reference for the actual joint [13, 56, 142]. Phase information estimated by AFOs has also been used to generate a custom trajectory in order to approximately achieve the desired power output [137]. In a diferent approach, the trajectory is generated online before each step based on the spring-loaded inverted pendulum (SLIP) model, taking the dimensions of possible obstacles into account [201]. For exoskeletons targeted at hemiplegic people, the movement of the nonparetic side at each step has also been recorded and used as the reference trajectory for the paretic leg [202, 203]. In a similar approach, kernel-based nonlinear flters have been used to learn the movements of the nonparetic leg as a function of gait phase online, and the learned functions are then used to generate the reference trajectory for the paretic leg [204]. Using position profles is often associated with rigid position control in the full mobilization case. Ten, the position profle is simply played back over time [27, 32, 33, 131]. Te challenge is then to generate a set of gait trajectories that are comfortable, stable and able to overcome obstacles. For partial assistance, it is associated with impedance control [13, 145, 182, 205–207]. Tese trajectories can be played back over time [89, 205], or may be time-invariant (a tunnel or force feld around the nominal path) [17, 181, 197, 208–211]. In [137], the reference profle is artifcially generated and tuned to achieve a certain pattern of assistance. A combination of rigid trajectory tracking for some degrees of freedom and partial assistance around a trajectory for others has also been used [196]. Te major drawback of the fxed-position-profle-based methods is their lack of fexibility, especially in the case of full mobilization. Even with many of the online modifed or generated trajectories, the user is still forced to walk with the given gait pattern, which may not be suitable, and the trajectories are often specifc to a particular terrain. For the partial assistance paradigm, even though the user has the freedom to diverge from the profle, it is still imposed and the controller will try to push in that direction, which might not necessarily help the user. Torque profle (TPR) Using a torque profle is the most simple and common method for partial assistance. A torque profle can be played back over time when it is triggered by an event [48, 96–100, 104, 111, 115, 119, 121, 123, 125, 151, 154, 165]. As the timing is very important, the torque profle may have (possibly online) tunable delay at the beginning of the torque profle
. Te torque profle itself may change over time, and be optimized online [105]. Te torque profle can be as simple as a square pulse [103]. In some studies, the torque profles are fne-tuned ofine based on subjective feedback from the users [136, 212] or previous measurements from the users [169]. In others, they are optimized online for metabolic cost reduction [105, 213, 214]. Te gait phase can also be estimated continuously, so the torque is applied as a direct function of the phase, independently of the time [50, 122, 136, 138–140, 143, 215–217], or combined with other inputs [51, 137, 141]. Probably the simplest case is the constant extension torque profle applied to the knee joint, when the leg is in single stance [176]. Impedance controller (ZCT) Impedance control is a widely used method in rehabilitation robotics and many other felds where the mechanical interactions with the user and the environment are signifcant [218]. As already mentioned, this method is used mostly in partial assistance paradigms where the human limbs are considered as active elements. Impedance control is often implemented such that the user gets the assistance torque only in case of a large deviation from the intended movement. Tis is usually called "assist-as-needed" and is mainly used for rehabilitation training, since it is believed to induce more active participation from the user compared to constant assistance or full mobilization, thus improving the learning and recovery. In practice, impedance control can be implemented as an M/K/B (inertia/stifness/damping) based dynamical system relating joint angles to torques [47, 49, 89, 127, 137, 153, 210, 219–223]. Either a reference target trajectory is played back over time [38, 145, 206], or the target is fxed and changes (also the stifness and damping) only when the gait state changes [137, 166, 172, 180, 224–226]. In both cases, the target trajectory is generally in joint-space. Another type of implementation is to use a force feld with the joint states (angle, speed, acceleration, etc.) as inputs [181, 204, 227, 228]. A variation of the force feld is the fow feld controller proposed by Martinez et al. [229], which can also use the "state" given by several joints, while applying torque only at one [168]. A combination of both the force feld and the fow feld is suggested by Jabbari Asl et al. [230]. Note that using a multi-dimensional force-feld in foot-locus-space to assist the leg to follow a pre-defned trajectory (such as the strategy used in [211]) is time-invariant, and is not the same as playing back a reference trajectory, even if both are classifed as impedance control. Te impedance controller is usually implemented in software by changing the motor torque depending on the position and movement of the joint, but it can also be implemented using mechanical elements only (see "Torque control"). In [231], a negative impedance is tuned to compensate that of the leg to make walking less demanding for the user, since less efort is required to generate the same movement of the legs. Finally, another possible strategy is to "attract" the joint to its future position with a virtual stifness feld [86, 142]. Te future position can be predicted by exploiting the periodicity of the gait. Te trajectory is typically identifed online with an adaptive-frequency oscillator (AFO). Tis is equivalent to impedance control with the time-shifted identifed trajectory as a target. Muscles activity amplifcation (MYO) A joint torque that depends directly on the measured muscular activity is simple and can be very efective, since it can detect the intention of the user before the movement starts. However, it is usually limited by the fact that electromyography (EMG) sensors are time-consuming to set up, the signal amplitude may change because of changes of skin conductivity and muscles fatigue, and that some muscles are not accessible with surface electrodes. In addition, this technique becomes even more difcult in case of neurologic impairment. In fact, the muscles may have a lower contraction which reduces the amplitude of the measured voltage and hence the signal-to-noise ratio (SNR). Tis method is simply not usable with people afected with complete paraplegia because there is no voluntary stimulation of the muscles. It also cannot help people afected with coordination troubles, which would just be amplifed by the device. In this method, generally the calculated torque is directly applied to the joint [96, 116, 232–234, 234–236], but in some papers the torque is fed to an admittance model to generate position commands for the low-level controller [237, 238]. In
terms of the approach to calculating the intended torque from muscle activity, several variants can be distinguished: Te amplifcation of independent muscles activities is typically implemented with one artifcial muscle per biologic muscle [239]. Its advantage is that the cocontraction of the biologic muscles also produces cocontraction of the artifcial muscles, which allows to amplify both the torque and stifness of the muscles. Tis can also be implemented with a single muscle; however, in this case, the biologic co-contraction will make the orthosis produce net torque. Tis approach has also been used in ankle exoskeletons with only unidirectional actuation (e.g. plantarfexion assistance only) [239, 240] Te diferential amplifcation of muscular activity computes the assistive torque by computing the difference of the activations [241]. Co-contraction just results in less torque. However, it may be approximative because of the non-linear activity/torque relationship. A variant is to let one activation inhibit the other [242]. Instead of assuming the joint torque proportional to the raw measured activation, an alternative is using a calibrated musculoskeletal model to compute the joints torques from the measured activation [243]. In some approaches, muscle activity amplifcation is guided by a gait phase estimation method, where the activity of a certain muscle causes assistance only during a specifc period of the gait cycle [116, 233]. A thorough review of these techniques can be found in [244]. In [245], the EMG-torque relation is estimated online during swing using a physical model. In [133], capacitive sensing is used instead of electromyography (EMG). Direct joint torque estimation (JTE) Te biological joint torques required for performing a certain movement can be estimated approximately using a simplifed model with several weighted segments, and then (completely or partially) applied with an exoskeleton. Such a method has been used to assist squatting [92, 246, 247] or stair ascent [248] assuming quasi-static movement (neglecting inertia terms and only compensating the weight). A similar approach has been used in [49] to assist gait in diferent terrains (level ground, stairs, ramp) in conjunction with other strategies. In [249] the authors have used inverse dynamics (4 sets of equations depending on the contact point(s) with the ground) to estimate the joints torque. Another method that does not rely on an accurate model, is using ground reaction forces, shank angle, and shank length [250, 251]. It has also been proposed to use a spring-loaded inverted pendulum (SLIP) model to estimate the required biological hip and knee torques [252]. Te point foot approximation is made, and the controller requires hip/knee joints angles, ground reaction forces, and center of pressure (CoP) position obtained with an instrumented treadmill. Similarly, in [253] the required stance ankle torque to compensate the efect of gravity has been derived based on a simple 2-DoF compass gait model. A mass model and ground reaction forces are used in [254] to estimate the hip and knee torques during gait, but the exoskeleton is actually not actuated. Model-computed action to keep balance (BAL) Some control strategies address the issue of balance during gait based on diferent mathematical models of walking. For full mobilization exoskeletons, provided they have enough actuated degree of freedoms (DoFs) and that the user does not interfere, walking controllers developed for humanoids have been used [20, 194, 255, 256]. In [89] hip abd/adduction trajectories during swing are adapted online to improve lateral balance based on the "extrapolated center of mass" concept. In another approach, the diference between model-computed and actual GRFs have been fed to an admittance model to update the predefned trajectories online [198]. In the partial assistance paradigm, Zha et al. [257] have developed a controller only assisting in case of loss of balance, which is detected based on a quantitative balance metric. A model-based assistive torque is then calculated as the weighted sum of gravity, Coriolis, and inertial terms with weights determined using fuzzy logic. Neuromuscular model (NMM) A class of bio-inspired controllers attempt to mimic the human neuromuscular system, consisting of virtual neurons and muscles. Tese virtual muscles are mathematical models based on the Hill-type muscle model [258] that generate torques as a function of the activation signal and the current muscle states (which are in turn a function of joint angles and angular velocities). Te torque applied to each joint is then obtained as the algebraic sum of the torques generated by the virtual muscles acting on that joint. Some of these controllers are based on the neuromuscular refex model from Geyer and Herr [259]. Tis bio-inspired model works based on feedback loops, or "ref

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.