Article

Personalizing exoskeleton assistance while

walking in the real world

https://doi.org/10.1038/s41586-022-05191-1 Patrick Slade1,2, Mykel J. Kochenderfer3, Scott L. Delp1,2 & Steven H. Collins1 ✉

Received: 23 February 2022

Accepted: 4 August 2022 Personalized exoskeleton assistance provides users with the largest improvements in
Published online: 12 October 2022 walking speed1 and energy economy2–4 but requires lengthy tests under unnatural
laboratory conditions. Here we show that exoskeleton optimization can be performed
Open access
rapidly and under real-world conditions. We designed a portable ankle exoskeleton
Check for updates based on insights from tests with a versatile laboratory testbed. We developed a
data-driven method for optimizing exoskeleton assistance outdoors using wearable
sensors and found that it was equally effective as laboratory methods, but identified
optimal parameters four times faster. We performed real-world optimization using
data collected during many short bouts of walking at varying speeds. Assistance
optimized during one hour of naturalistic walking in a public setting increased
self-selected speed by 9 ± 4% and reduced the energy used to travel a given distance by
17 ± 5% compared with normal shoes. This assistance reduced metabolic energy
consumption by 23 ± 8% when participants walked on a treadmill at a standard speed
of 1.5 m s−1. Human movements encode information that can be used to personalize
assistive devices and enhance performance.

Exoskeletons that assist leg movement show promise for enhancing Data-driven models may be able to capture important features of
personal mobility but have yet to provide real-world benefits. Millions human performance more simply21–25.
of people have mobility impairments that make walking slower5 and We developed a data-driven model that relates human motion during
more fatiguing6, while millions more people have occupations that exoskeleton-assisted walking to metabolic energy consumption and
require strenuous locomotion7. In research laboratories, exoskele- can be used outside the laboratory. Human movement arises from
tons can increase walking speed1,8,9 and reduce the energy required the interaction between the inertia of our body segments and forces
to walk2–4,10–16, but these benefits have not yet translated to real-world from the environment and our muscles. We hypothesized that careful
conditions17. Providing beneficial assistance in the real world is difficult analysis could extract meaningful information about muscular energy
for several reasons: the specialized equipment used to personalize expenditure from subtle changes in motion. In a previous experiment4,
assistance is not available outside the laboratory; unlike walking on a participants walked with exoskeleton assistance in about 3,600 differ-
treadmill, everyday walking occurs in many bouts of varying speed and ent conditions while data were recorded from both laboratory equip-
duration; and devices must be self-contained and easy to use. In this ment that measure biomechanical outcomes and low-cost, portable
study, we addressed each of these challenges to demonstrate effective sensors on the exoskeleton. We trained a logistic regression model
exoskeleton assistance under naturalistic conditions. using this previous dataset (Extended Data Fig. 1). The data-driven
Maximizing the benefits of exoskeleton assistance requires classification model compared sensor data from two different patterns
personalization to individual needs, which is challenging outside of exoskeleton assistance, each defined by a ‘control law’, and classi-
of a laboratory. The largest improvements in human walking per- fied which control law provided a larger benefit. The model inputs
formance have been achieved by individualizing assistance using were ankle angle and ankle velocity, segmented by gait cycle, and the
human-in-the-loop optimization1–4, a process in which device con- torque parameters for each control law. The model then estimated the
trol is systematically tuned to improve human performance while a likelihood that the first control law resulted in lower metabolic energy
person uses a device. Measuring important aspects of performance, expenditure. In essence, the classifier favoured later, larger exoskeleton
including metabolic rate16, has required expensive laboratory equip- torques and smooth, well timed movements that led to increased ankle
ment and long periods of steady treadmill walking18. Individualizing extension at toe-off. During optimization, the user experienced a set
consumer or medical devices in this way would require several long of control laws, the data-driven model compared all possible pairs of
visits to a specialized clinic, which would be costly and impractical. control laws, the control laws were ranked, and an optimization algo-
If human performance could instead be estimated quickly, using rithm26 updated the estimate of the optimal parameters and generated
low-cost wearable sensors, optimization could be performed as a new set of control laws to evaluate (Fig. 1). This process was repeated
people moved naturally through their daily lives. This might be until convergence criteria were met.
possible using musculoskeletal modelling19, but such simulations Data-driven optimization can use the information embedded in
are computationally intensive20 and require individualization. our movements to identify exoskeleton assistance patterns that are

1
Department of Mechanical Engineering, Stanford University, Stanford, CA, USA. 2Department of Bioengineering, Stanford University, Stanford, CA, USA. 3Department of Aeronautics and
Astronautics, Stanford University, Stanford, CA, USA. ✉e-mail: [email protected]

Nature | Vol 610 | 13 October 2022 | 277

Article
a b Wearable c d
Pairwise comparisons Data-driven classifier
sensor data
Permutations i,j = [1 ... k]
Storage for Control i – Control j = 'ij
k control laws
Θ ( WΔΘ WΔΘ WΔC )
Ankle angle – =

Ankle velocity Θ – = (ΔΘij ΔΘij ΔCij)

C wij
Control laws – =

e Logistic 1
function 1 + e–w

pij Probability
h Apply k control laws g Update optimizer f Score control laws control law
i is better
c = (d1, d2, d3, d4) New control laws than j
S1

...
d2 d4 C = [c1 ... ck] d1
d3 k

...
C S Σ pnj
Torque

Sn =
d1 j=1
New j≠n

...

...
estimated d2 Sk
Gait cycle optimum

Fig. 1 | Data-driven exoskeleton optimization. We used data from laboratory obtain the pair coefficient (wij). e, A logistic function uses the pair coefficient to
tests to train a model that can perform optimization in real-time outside the compute the probability (pij) that the first control law is more beneficial than
laboratory. a, During optimization, the participant walks with the exoskeleton the second. f, The score (S) for each control law (n) is computed by summing the
and experiences a sequence of k control laws, each defining a pattern of probabilities of all pairs that include that control law. g, Control laws are then
exoskeleton torque. The optimizer’s goal is to identify the torque pattern that ranked by score and used to update an optimizer. h, The optimizer selects a set
maximizes performance. b, Ankle angle (θ) and ankle velocity (θ )̇ for each of k new control laws, consisting of d parameters, to evaluate. This optimization
stride are recorded from sensors on the exoskeleton. c, All possible pairs of process is repeated until convergence criteria are satisfied, in this case a set
control laws are then compared (C). For each pair, differences in segmented number of evaluations having been completed. During real-world experiments,
motion data (Δ) are calculated by subtraction. d, Differences in motion are optimization was performed on the exoskeleton’s microcontroller.
multiplied with classifier model weights (W), using a dot product operation, to

as effective as those found with laboratory-based methods, but in steady walking on a treadmill have not provided similar benefits during
one-quarter of the time. We conducted experiments to optimize assis- use under more natural conditions17. The speed-adaptive controller
tance with a tethered exoskeleton emulator (Fig. 2a). The data-driven we developed (Extended Data Fig. 2) interpolated between assistance
optimization evaluated eight sets of control laws in 32 min, four times parameter values previously optimized at different walking speeds
faster than the state-of-the-art approach using indirect respirometry (Fig. 3a) based on the estimated speed of each step (Fig. 3b). We tested
to measure metabolic rate2 (Fig. 2b). Data-driven and metabolic opti- Speed-adaptive assistance on a subset of participants as they walked on
mization approaches identified the same participant-specific adjust- a treadmill with sinusoidally varying speeds. Speed-adaptive assistance
ments to assistance (Fig. 2c). Data-driven Optimized assistance and reduced the energetic cost of walking more than Generic assistance
Metabolic Optimized assistance resulted in similar metabolic cost, with constant parameters (Fig. 3c). Adjusting exoskeleton assistance
which was significantly lower than the metabolic cost of walking with based on speed is an effective strategy for handling speed variations
the exoskeleton in a Zero Torque mode (Fig. 2d). The average of the that occur during normal walking.
Data-driven Optimized parameters matched those of Generic assis- We created an untethered exoskeleton for real-world assistance
tance, which were taken from the best previous study4, but Data-driven using a design approach based on emulation and optimization. Wear-
Optimized assistance provided a larger benefit. This demonstrates the able robotic devices are typically designed using models or intuition,
importance of individualization; even subtle changes in torque can built as specialized prototypes, and then tested. However, humans are
lead to substantial performance enhancements. To test the general- highly complex and diverse, making it difficult to predict the range of
ity of the data-driven model, we conducted experiments at a range characteristics that will be optimal across a population. As a result,
of additional speeds and inclines with a subset of participants. The most devices designed this way are unable to provide optimal assis-
Data-driven Optimized assistance and Metabolic Optimized assistance tance and often provide no benefit at all. To develop the untethered
resulted in similar torque profiles and metabolic cost reductions across exoskeleton used in this study, we first performed experiments with
these conditions (Fig. 2e). This shows that the data-driven classifica- versatile exoskeleton emulators30. These laboratory-based, tethered
tion model captured a fundamental relationship between exoskeleton hardware systems allowed us to perform a wide range of control and
torque, ankle movement and whole-body walking effort. The model optimization experiments (Figs. 1–3) and identify the electromechani-
approximates this biological relationship, precluding statistical guar- cal characteristics that our untethered device would need. Using these
antees of optimality. Nevertheless, our results demonstrate that human design guidelines, we built a specialized, untethered device that pro-
movement encodes information related to underlying physiological vides predictable, meaningful benefits. This emulation and optimiza-
processes, and that data-driven methods can extract this information tion design paradigm can reduce the cost and time required to develop
without laboratory equipment or complex multi-scale models. new wearable robots.
We developed a speed-adaptive controller to adjust assistance based On the basis of the results of our emulator experiments, we designed
on natural variations in walking speed. People vary their walking speed a specialized, untethered ankle exoskeleton. The system consisted of an
widely during the day27 in response to changes in context28 and con- exoskeleton worn on each ankle and a battery pack at the waist (Fig. 4a
straints29. Variations in speed complicate exoskeleton control and may and Supplementary Video 1). The exoskeleton was designed to apply the
help explain why assistive devices that reduce walking effort during range of optimal torque profiles identified in the tethered optimization

278 | Nature | Vol 610 | 13 October 2022

 a b Metabolic c 0.9 Generic
Respirometry
Optimized assistance
20 Data-driven Individual

Optimized parameters (%)

Difference from Metabolic

torque profiles (Nm kg–1)

Optimized participants

Data-driven Optimized
Tether 15

10
Exo-
skeleton
5

0
0
0 32 64 96 128 20 60
Time (min) Gait cycle (%)
Treadmill
1 Slow walk Normal walk Fast walk 10º incline
d e

Torque (Nm kg–1)

Zero Torque
Normal Shoes
Generic Assistance
Metabolic Optimized
0
Data-driven Optimized 20 65 20 65 20 65 20 65
Gait cycle (%) Gait cycle (%) Gait cycle (%) Gait cycle (%)
*
Participant 1
0%
3
0% ** Participant 2 –7%
6
Metabolic cost (W kg–1)

Participant 3
–15% –25%
Metabolic cost (W kg–1)

***
0% –42%
–33% –43%
2 –41% 4 –13%
–43%
0% –35% –44%
–15%
–32% –47%
1 2 0% –16% –23% –42%
–21% –23% –45%

0 0
Normal walk Slow walk Normal walk Fast walk 10° incline
n=9 n=3 n=3 n=3 n=3

Fig. 2 | Data-driven optimization results. a, Exoskeleton assistance was when walking at 1.25 m s−1 (ANOVA, n = 9, *P ≤ 2.7 × 10 −8, **P ≤ 2.4 × 10 −5,
applied using a tethered ankle exoskeleton emulator43. b, Assistance parameters ***P ≤ 0.047). The boxes extend from the lower to upper quartile values of the
optimized using the data-driven method converged to within 5% of the data, with a line at the median and a dot at the mean. The whiskers extend
parameters identified using metabolic optimization, but in one-quarter of the between the minimum and maximum of the data values. e, Optimized torque
time (n = 9). The error bars represent the standard deviation. c, Individual patterns varied with walking condition, with similar changes in Data-driven
participants had unique Data-driven Optimized parameters, centred around Optimized and Metabolic Optimized parameters. Data-driven Optimized
the Generic assistance parameters. d, Data-driven Optimized assistance and assistance and Metabolic Optimized assistance led to similar reductions in
Metabolic Optimized assistance resulted in similar metabolic costs of walking, metabolic rate when walking at 0.75 m s−1 (slow), 1.25 m s−1 (normal) and
significantly lower than with Zero Torque, Normal Shoes or Generic assistance, 1.75 m s−1 (fast), and on a 10° incline at 1.25 m s−1.

study (Fig. 2) while having low mass (1.2 kg per ankle). A brushless motor walking occurs in many separate bouts, most of which are short, with
and custom drum transmission applied torque about the ankle joint, 90% being less than 100 steps in duration34. Speed is relatively consist-
while portable electronics sensed the user’s motion and performed ent within each bout, but varies across bouts27. This fragmentation
real-time control and optimization (Fig. 4b and Extended Data Fig. 3). presents a challenge for collecting optimization data and efficiently
The exoskeleton provided a peak torque of 54 Nm (Fig. 4c), which was fine-tuning assistance. Our data-driven optimization method addresses
about 50% to 75% of the biological ankle torque of participants in this the problem of gathering useful data from short walking bouts by using
study31. Torque was controlled using a mixture of classical feedback kinematic data collected with every step. In pilot tests, we found that
control and iterative learning32, with a tracking error of less than 1% of we could accurately compare control laws based on just 44 continu-
the peak torque. Maximum assistance could be applied continually ous steps, opportunistically captured during natural bouts, allowing
without overheating the motor (Fig. 4d). The battery weighed 0.3 kg and our system to accumulate data from about 77% of steps in a typical
powered the exoskeleton for at least 30 min on a single charge. While day34. We addressed variations in speed by defining speed bins based
the energy cost of carrying mass near a distal joint is high33, locating on observed human behaviour, associating collected data with the
motors and electronics near the assisted joint results in more efficient appropriate bin, noting when sufficient data for any one speed bin had
power transmission, a simpler design and lower total weight, which been accumulated, applying the data-driven classifier to rank assistance
can yield large net benefits. parameters and using these rankings to update the optimal parameter
We used the information encoded in a single walking step to optimize estimates for all speed bins (Extended Data Fig. 4).
exoskeleton assistance while people walked naturally in short bouts of Real-world optimization quickly improved assistance during natural
varying speed. People take thousands of steps per day, but real-world walking conditions. We conducted experiments in which participants

Nature | Vol 610 | 13 October 2022 | 279

Article
a b c
1.0 4 0%

Treadmill speed –15%

0.9 1.8 Speed estimate
Peak torque
Optimized torque parameters

3 –25%

Metabolic cost (W kg–1)

Peak time

Walking speed (m s–1)

Rise time 1.6 –37%
0.8 Zero Torque
Fall time
1.4 Normal Shoes
2
0.7
Generic Assistance
1.2
Speed-adaptive
0.6 Assistance
1.0 1
Participant 1
0.5 Participant 2
0.8
RMSE = 0.06 m s–1
0 Participant 3
0.75 1.25 1.75 0 15 30 Time-varying
Walking speed (m s–1) Time (s) n=3

Fig. 3 | Speed-adaptive control. a, The Speed-adaptive controller interpolated input (Extended Data Fig. 2) with a root-mean-square error (RMSE)
between previously optimized assistance parameters to estimate the optimal of 0.06 m s−1. The shaded region represents the mean ± one standard deviation.
parameters for each step based on walking speed on the previous step. c, When participants (n = 3) walked on a treadmill that varied speed sinusoidally
Optimized values are normalized to the allowable range for each parameter. between 0.75 m s−1 and 1.75 m s−1, Speed-adaptive assistance reduced the
b, Ground truth and estimated walking speed for a representative participant. metabolic cost of walking more than the fixed Generic assistance.
Speed was estimated on each step using a model that took stride period as an

performed one hour of walking in short bouts with exoskeleton assis- throughout the experiment (Fig. 5e), indicating steadily decreasing
tance (Fig. 5a) on a public sidewalk (Fig. 5b and Supplementary Video 2). uncertainty as to which exoskeleton parameters would result in optimal
Participants were given ecologically relevant35 audio prompts28 that performance according to the data-driven model. Post hoc analysis
caused them to self-select walking speeds that matched a ground-truth showed that the optimizer did not reach steady state, suggesting that
distribution27 (Fig. 5c). Prompts were provided in random order and additional time could have provided a better estimate of the optimal
at specific intervals to obtain bout durations that also matched a parameters. Peak torques optimized during naturalistic walking were
ground-truth distribution34 (Fig. 5d). The optimizer steadily converged larger than those from treadmill-based experiments (Fig. 5f). Partici-
pants may have felt more stable during outdoor walking36, allowing
them to benefit from larger torques, consistent with observations from
a other comparisons of outdoor and treadmill walking with exoskeleton
b
assistance37.
6
Real-world Optimized assistance increased self-selected walking
speed and reduced the metabolic energy expended per distance
2
3 travelled during naturalistic walking. In a separate validation experi-
ment, participants performed a fixed set of outdoor walking bouts
with varying durations and speeds, while ground-truth metabolic rate
1 and speed were measured (Supplementary Video 3). Condition order
was randomized (Extended Data Table 1). With Real-world Optimized
assistance, the energetic cost of transport was reduced by 17 ± 5%
(analysis of variance (ANOVA), n = 10, P = 0.039) and walking speed was
increased by 9 ± 4% (ANOVA, n = 10, P = 0.031) compared with Normal
Shoes (Fig. 5g). These energy savings are equivalent to removing a 9.2 kg
5
backpack38, and the increase in walking speed of 0.12 m s−1 is similarly
4 meaningful39. Real-world Optimized assistance provided roughly twice
the benefits of Generic Speed-adaptive assistance, indicating that per-
sonalization was an important contributor to these benefits. Generic
Speed-adaptive assistance may have provided a larger benefit if it had
used the average of the torque parameters optimized during outdoor
c d
75
RMSE = 0.51 Nm Motor limit
54
Measured Fig. 4 | Untethered ankle exoskeleton. a, A participant walking in a
Motor temperature (°C)

Desired community setting wearing the exoskeleton. b, The exoskeleton consists of

40 (1) a battery pack worn on the waist, (2) a motor, drum and rope transmission to
Torque (Nm)

produce assistive torques, (3) electronics to receive sensor data, command the
motor and perform optimization, (4) a carbon fibre and aluminium frame to
transmit forces, and (5) a shoe and (6) a calf strap to transfer forces to the body.
20 Asymptote c, The motor can apply a peak torque of 54 Nm when walking at 1.5 m s−1,
35 sufficient to match the optimized assistance parameters identified in emulator
experiments. Torques were tracked accurately; the shaded region represents
25 Exponential fit the mean ± one standard deviation. d, The motor temperature during 30 min of
0
20 60 0 30 walking with maximum assistance remained well below the 75 °C thermal limit.
Gait cycle (%) Time (min) An exponential fit indicated a steady-state temperature of 35.4 °C.

280 | Nature | Vol 610 | 13 October 2022

 a b 15 m 1.6 c d e
2.3 15 So
Optimization bouts
Validation bouts

Optimization convergence (V)

Ground-truth
distribution

Walking speed (m s–1)

Probability density

Number of bouts
End
Start

Bouts > 210 s

0 0 0
0.8 1.3 1.8 0 200 0 24
0.9 Walking speed (m s–1) Bout time (s) Generations

f g h
Individual participants Normal Shoes
Generic Speed-adaptive
Generic Speed-adaptive *
54 40
Real-world Optimized 0%
8
Optimized peak torque (Nm)

* * –18%
Optimized rise time (%)

5 0%

Metabolic cost (W kg–1)

0%
Cost of transport (J kg–1 m–1)

* 6
–9%

Walking speed (m s–1)

4
1.5 5% 9% *
–17% 0% –23%
0%
3 4 –16%
Rise time
2
Torque
Torque

Peak 2
torque 10 1
35
Gait cycle (%) Gait cycle (%)
0 0 0 0 0
1.07 1.33 1.58 1.07 1.33 1.58 1.25 m s–1 1.5 m s–1 10º incline
Outdoor bouts
Walking speed (m s–1) Walking speed (m s–1) n = 10 n = 10 n = 10
n = 10

Fig. 5 | Real-world optimization of exoskeleton assistance. a, Participant averaged across all participants. Peak torque values in this figure are not
walking on the public validation course. b, Map of the 566-m course used for normalized to emphasize that several participants reached the maximum peak
optimization and validation. Participants walked the course repeatedly during torque that the untethered exoskeleton could provide. The inset torque
optimization. c,d, Distribution of self-selected walking speeds (c) and walking profiles indicate how each torque parameter affected the assistance profile.
bout durations (d) during optimization and validation, compared with g, During validation under naturalistic walking conditions on the public course,
previously recorded ground-truth distributions of real-world walking data27,34. Real-world Optimized assistance substantially reduced the energy cost of
e, As assistance was optimized over one hour of naturalistic bouts of walking, transport and increased walking speed compared with Normal Shoes (ANOVA,
the convergence parameter (σ) continually improved. The error band n = 10, *P ≤ 0.039). h, Real-world Optimized assistance also substantially
represents one standard deviation. f, Optimized parameters for each reduced the metabolic cost of walking compared with Normal Shoes during
participant were unique. The red squares depict the Generic Speed-adaptive benchmark treadmill conditions (ANOVA, n = 10, *P ≤ 0.023). Boxes extend
assistance parameters, consisting of constant values for peak torque from the lower to upper quartiles, with a line at the median and a dot at the
normalized to body mass (Nm kg−1) and rise time (percent gait cycle). For ease mean. Whiskers extend between the minimum and maximum values.
of comparison, we show the Generic Speed-adaptive peak torque in Nm,

walking, rather than treadmill walking. These results demonstrate that Pilot results suggest that the device provides similar benefits under
lower-limb exoskeletons can provide meaningful benefits under natu- other conditions, including walking on a 5° incline, loaded walking and
ralistic walking conditions and provide benchmarks for assessing the stair climbing (Extended Data Fig. 6). Emulator-informed hardware
real-world benefits of future devices. Assistance can be personalized design coupled with opportunistic, data-driven optimization led to
automatically in a natural setting, seamlessly improving human–robot exceptional performance enhancements across walking conditions.
interaction over time. Participants reported that the untethered exoskeleton was easy
Assistance optimized under real-world conditions produced even to use and relatively comfortable. Wearable robotic devices should
larger benefits under standard treadmill conditions. After performing be usable, comfortable and functional for everyday activities to be
optimization in a public setting, we tested our untethered exoskeleton adopted by users41. Participants reported that the exoskeleton was
during standardized laboratory walking conditions to directly compare relatively easy to use (Extended Data Table 3), ranking it in the 65th
with previous devices16. Real-world Optimized assistance reduced the percentile of previously surveyed consumer devices42. Participants
energy cost of treadmill walking by 16% at 1.25 m s−1, 23% at 1.5 m s−1, and found that the exoskeleton did not interfere with their clothing and
18% when walking up a 10° incline (ANOVA, n = 10, P < 0.023) compared had a manageable weight, but were neutral as to whether it would be
with Normal Shoes (Fig. 5h and Extended Data Table 2), approximately comfortable to wear throughout the day (Extended Data Table 4). Par-
twice the benefits of the previous devices with the best performance ticipants reported that it was easy to put on and take off the exoskeleton,
for these conditions (Extended Data Fig. 5). The energy savings during stand while wearing the exoskeleton, and walk indoors and outdoors
inclined walking were equivalent to removing a 15.2 kg backpack40. for extended periods with the exoskeleton (Extended Data Table 5).

Nature | Vol 610 | 13 October 2022 | 281

Article
Six of the ten participants reported that they would prefer using the 15. Kim, J. et al. Reducing the metabolic rate of walking and running with a versatile, portable
exosuit. Science 365, 668–672 (2019).
exoskeleton rather than normal shoes if they were to walk the public 16. Sawicki, G. S., Beck, O. N., Kang, I. & Young, A. J. The exoskeleton expansion: Improving
course again. The device we tested is a research prototype and not a walking and running economy. J. Neuroeng. Rehabil. 17, 25 (2020).
refined product; substantial improvements would be required to allow 17. Fang, Y., Orekhov, G. & Lerner, Z. Improving the energy cost of incline walking and stair
ascent with ankle exoskeleton assistance in cerebral palsy. IEEE Trans. Biomed. Eng. 69,
reliable, unsupervised use during typical daily activities. The survey 2143–2152 (2021).
results suggest that it may be possible to create mobility-enhancing 18. Selinger, J. C. & Donelan, J. M. Estimating instantaneous energetic cost during
products that are easy to use, comfortable and reliable, and that many non-steady-state gait. J. Appl. Physiol. 117, 1406–1415 (2014).
19. Seth, A. et al. OpenSim: simulating musculoskeletal dynamics and neuromuscular
people may opt to use them. control to study human and animal movement. PLoS Comput. Biol. 14, e1006223 (2018).
These approaches to real-world personalization, adaptive assistance 20. Ijspeert, A. J. Biorobotics: using robots to emulate and investigate agile locomotion.
and specialized exoskeleton design could potentially be extended to Science 346, 196–203 (2014).
21. Rosenberg, M. C., Banjanin, B. S., Burden, S. A. & Steele, K. M. Predicting walking
address the needs of workers with physically demanding jobs and peo- response to ankle exoskeletons using data-driven models. J. R. Soc. Interface 17,
ple with mobility impairments. A similar overall development approach 20200487 (2020).
22. Lee, D., Kang, I., Molinaro, D. D., Yu, A. & Young, A. J. Real-time user-independent slope
could be used to address the most important outcomes for each bio-
prediction using deep learning for modulation of robotic knee exoskeleton assistance.
mechanically and neurologically similar group. Assistance could aid IEEE Robot. Autom. Lett. 6, 3995–4000 (2021).
a variety of tasks, such as stair climbing or lifting, and improve other 23. Slade, P., Kochenderfer, M. J., Delp, S. L. & Collins, S. H. Sensing leg movement enhances
wearable monitoring of energy expenditure. Nat. Commun. 12, 4312 (2021).
aspects of performance, such as balance or joint pain. In each case,
24. Matijevich, E. S., Volgyesi, P. & Zelik, K. E. A promising wearable solution for the practical
additional training data could be collected in the laboratory and used to and accurate monitoring of low back loading in manual material handling. Sensors 21,
train new data-driven models, illuminating the information contained 340–265 (2021).
25. Wu, W., Saul, K. R. & Huang, H. H. Using reinforcement learning to estimate human joint
within the body’s movements for each task. With each training dataset, moments from electromyography or joint kinematics: an alternative solution to
the learned models could be made more capable, progressively build- musculoskeletal-based biomechanics. J. Biomech. Eng. 143, 044502 (2021).
ing more general relationships between movement and performance 26. Hansen, N. in Towards a New Evolutionary Computation (eds Lozano, J. A. et al.) 75–102
(Springer, 2006).
outcomes. Data from laboratory-based emulation and optimization 27. Baroudi, L. et al. Estimating walking speed in the wild. Front. Sports Act. Living 2, 583848
experiments could simultaneously provide design guidelines for prod- (2020).
ucts. When used regularly, we expect devices like this to become finely 28. Brinkerhoff, S. A., Murrah, W. M., Hutchison, Z., Miller, M. & Roper, J. A. Words matter:
instructions dictate “self-selected” walking speed in young adults. Gait Posture 95,
tuned to the needs of each individual, resulting in larger performance 223–226 (2019).
enhancements than observed in this study. Longitudinal experiments 29. Brown, G. L., Seethapathi, N. & Srinivasan, M. A unified energy-optimality criterion
will be needed to understand how such assistance affects behaviour and predicts human navigation paths and speeds. Proc. Natl Acad. Sci. USA 118, e2020327118
(2021).
quality of life; as moving becomes easier, we hope to find that people 30. Caputo, J. M. & Collins, S. H. A universal ankle-foot prosthesis emulator for human
will be more active, helping them to lead healthier lives. locomotion experiments. J. Biomech. Eng. 136, 035002 (2014).
31. Moisio, K. C., Sumner, D. R., Shott, S. & Hurwitz, D. E. Normalization of joint moments
during gait: a comparison of two techniques. J. Biomech. 36, 599–603 (2003).
32. Zhang, J., Cheah, C. C. & Collins, S. H. in Bioinspired Legged Locomotion: Concepts,
Online content Control and Implementation (eds Sharbafi, M. & Seyfarth, A.) Ch. 5 (Elsevier, 2017).
33. Browning, R. C., Modica, J. R., Kram, R. & Goswami, A. The effects of adding mass to the
Any methods, additional references, Nature Research reporting sum-
legs on the energetics and biomechanics of walking. Med. Sci. Sports Exerc. 39, 515–525
maries, source data, extended data, supplementary information, (2007).
acknowledgements, peer review information; details of author con- 34. 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, 1077–1090
tributions and competing interests; and statements of data and code
(2008).
availability are available at https://doi.org/10.1038/s41586-022-05191-1. 35. Schmuckler, M. A. What is ecological validity? A dimensional analysis. Infancy 2, 419–436
(2001).
36. Stolze, H. et al. Gait analysis during treadmill and overground locomotion in children and
1. Song, S. & Collins, S. H. Optimizing exoskeleton assistance for faster self-selected adults. Electroencephalogr. Clin. Neurophysiol./Electromyogr. Motor Control 105,
walking. IEEE Trans. Neural Syst. Rehabil. Eng. 29, 786–795 (2021). 490–497 (1997).
2. Zhang, J. et al. Human-in-the-loop optimization of exoskeleton assistance during walking. 37. Haufe, F. L., Duroyon, E. G., Wolf, P., Riener, R. & Xiloyannis, M. Outside testing of wearable
Science 356, 1280–1284 (2017). robots for gait assistance shows a higher metabolic benefit than testing on treadmills.
3. Kim, M. et al. Bayesian optimization of soft exosuits using a metabolic estimator stopping Sci. Rep. 11, 14833 (2021).
process. In IEEE/RAS International Conference on Robotics and Automation (ICRA) 38. Bastien, G. J., Willems, P. A., Schepens, B. & Heglund, N. C. Effect of load and speed on
9173–9179 (IEEE, 2019). the energetic cost of human walking. Eur. J. Appl. Physiol. 94, 76–83 (2005).
4. Poggensee, K. L. & Collins, S. H. How adaptation, training, and customization contribute 39. Perera, S., Mody, S. H., Woodman, R. C. & Studenski, S. A. Meaningful change and
to benefits from exoskeleton assistance. Sci. Robot. 6, eabf1078 (2021). responsiveness in common physical performance measures in older adults. J. Am.
5. Studenski, S. et al. Gait speed and survival in older adults. JAMA 305, 50–58 (2011). Geriatr. Soc. 54, 743–749 (2006).
6. Enoka, R. M. & Duchateau, J. Translating fatigue to human performance. Med. Sci. Sports 40. Lloyd, R. & Cooke, C. B. The oxygen consumption with unloaded walking and load
Exerc. 48, 2228 (2016). carriage using two different backpack designs. Eur. J. Appl. Physiol. 81, 486–492 (2000).
7. Tudor-Locke, C., Leonardi, C., Johnson, W. D. & Katzmarzyk, P. T. Time spent in physical 41. Young, A. J. & Ferris, D. P. State of the art and future directions for lower limb robotic
activity and sedentary behaviors on the working day: the American time use survey. exoskeletons. IEEE Trans. Neural Syst. Rehabil. Eng. 25, 171–182 (2016).
J. Occup. Environ. Med. 53, 1382–1387 (2011). 42. Sauro, J. A Practical Guide to the System Usability Scale: Background, Benchmarks, and
8. Lee, H. J. et al. A wearable hip assist robot can improve gait function and Best Practices (Measuring Usability LLC, 2011).
cardiopulmonary metabolic efficiency in elderly adults. IEEE Trans. Neural Syst. Rehabil. 43. Witte, K. A. & Collins, S. H. Wearable Robotics 251–274 (Elsevier, 2020).
Eng. 25, 1549–1557 (2017).
9. Awad, L. N., Kudzia, P., Revi, D. A., Ellis, T. D. & Walsh, C. J. Walking faster and farther with a Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
soft robotic exosuit: Implications for post-stroke gait assistance and rehabilitation. IEEE published maps and institutional affiliations.
Open J. Eng. Med. Biol. 1, 108–115 (2020).
10. Mooney, L. M., Rouse, E. J. & Herr, H. M. Autonomous exoskeleton reduces metabolic cost Open Access This article is licensed under a Creative Commons Attribution
of walking. J. Neuroeng. Rehabil. 11, 80 (2014). 4.0 International License, which permits use, sharing, adaptation, distribution
11. Collins, S. H., Wiggin, M. B. & Sawicki, G. S. Reducing the energy cost of human walking and reproduction in any medium or format, as long as you give appropriate
using an unpowered exoskeleton. Nature 522, 212–215 (2015). credit to the original author(s) and the source, provide a link to the Creative Commons license,
12. Seo, K., Lee, J., Lee, Y., Ha, T. & Shim, Y. Fully autonomous hip exoskeleton saves and indicate if changes were made. The images or other third party material in this article are
metabolic cost of walking. In IEEE/RAS International Conference on Robotics and included in the article’s Creative Commons license, unless indicated otherwise in a credit line
Automation (ICRA) 4628–4635, (IEEE, 2016). to the material. If material is not included in the article’s Creative Commons license and your
13. Seo, K., Lee, J. & Park, Y. J. Autonomous hip exoskeleton saves metabolic cost of walking intended use is not permitted by statutory regulation or exceeds the permitted use, you will
uphill. In IEEE International Conference on Rehabilitation Robotics (ICORR) 246–251 (IEEE, need to obtain permission directly from the copyright holder. To view a copy of this license,
2017). visit http://creativecommons.org/licenses/by/4.0/.
14. Lee, T. et al. A flexible exoskeleton for hip assistance. In IEEE/RSJ International Conference
on Intelligent Robots and Systems (IROS) 1058–1063 (IEEE, 2017). © The Author(s) 2022

282 | Nature | Vol 610 | 13 October 2022

Methods following methods from previous studies of non-steady gait23,29. Excess
oxygen consumption and carbon dioxide production during the return
Experimental design to steady state in quiet standing reflect delays between instantane-
The research objective was to personalize exoskeleton assistance dur- ous energy use at muscles and expired gas measurements that arise
ing real-world walking. To achieve this objective, we proposed a method owing to mitochondrial, transport and respiratory dynamics18. Includ-
of data-driven optimization, which uses portable sensors on the exo- ing respiratory data from the period following activity enables more
skeleton to personalize assistance for each participant. We hypoth- accurate measurement of the energy actually expended during short
esized that Data-driven Optimized assistance would provide larger bouts of walking46. The energetic cost of transport was calculated as
reductions in metabolic rate than Generic assistance. We conducted the cumulative metabolic cost divided by the total distance walked.
a power analysis based on previous laboratory-based optimization
experiments and found that a sample size of eight participants was Exoskeleton assistance conditions
necessary for the planned validation experiments. This analysis used A variety of exoskeleton assistance conditions were evaluated to
a power of 1 − β = 0.8, a significance level of α = 0.05, the difference determine the benefits that they provided to the user. These assis-
in mean metabolic rate between optimized (1.44 W kg−1) and generic tance conditions included walking in Normal Shoes and walking with
(1.64 W kg−1) assistance from a previous experiment4, and the variabil- the exoskeletons while they applied Zero Torque, Generic assistance,
ity in metabolic rate (standard deviation 0.15 W kg−1) from the same Speed-adaptive assistance, Generic Speed-adaptive assistance, Meta-
experiment4. In this context, β is the probability of incorrectly accept- bolic Optimized assistance, Data-driven Optimized assistance and
ing the null hypothesis. We collected data from nine participants for Real-world Optimized assistance.
the tethered exoskeleton experiments and ten participants for the We tested walking in Normal Shoes, without the exoskeleton, as a
untethered exoskeleton experiments. We tested more participants baseline condition for the untethered exoskeleton experiments. Ide-
than the minimum number determined by the power analysis to pro- ally, assistance from an untethered exoskeleton would lead to a lower
vide a factor of safety in case data from any participants were found to metabolic cost than walking in Normal Shoes, providing a net benefit to
be unusable during later analysis. All participants that were recruited the user. Separate pairs of the same type of Nike running shoe, weighing
completed the corresponding experiment, and all data from all partici- 0.3 kg per shoe, were used for the Normal Shoes condition and incor-
pants were included in each corresponding analysis. All participants porated into the tethered exoskeleton and the untethered exoskeleton.
had at least 8 h of experience walking with assistance from powered The Zero Torque mode was an exoskeleton condition in which the
ankle exoskeletons, minimizing the effects of training that can occur exoskeleton provided no assistive torques. During this mode, the exo-
while participants learn to walk with an exoskeleton4. All participants skeleton maintained a small amount of slack in the cable transmission
were volunteers and provided written informed consent before com- so that virtually no torque was applied to the ankle. This condition
pleting the protocol (IRB-48749), which was approved by the Stanford was used as a baseline for experiments with the tethered exoskeleton
University institutional review board. Consent was obtained for publica- (but not for experiments with the untethered exoskeleton) because
tion of identifiable images of research participants. The experiments it allowed us to isolate the benefits of exoskeleton assistance from
consisted of human participant testing in both laboratory and outdoor the energetic costs of wearing the emulator, which were expected to
settings. Participants wore bilateral ankle exoskeletons and walked differ from those of an untethered device specialized to provide the
under a series of assistance conditions in a randomized order. Each same assistance.
of the experiments is described in the following sections. We used a The Generic assistance condition used a fixed set of assistance param-
one-way ANOVA to determine whether differences in the metabolic eters identified from a previous optimization experiment. Generic
cost of walking across assistance conditions were different from zero. assistance patterns have been found to reduce the metabolic cost of
walking less than assistance personalized to each individual2–4. The
Measuring the metabolic cost of walking tethered exoskeleton experiments in this study used Generic assistance
The metabolic cost of walking was computed with measurements computed by averaging the optimized parameters from a group of
from respirometry equipment. Respirometry equipment was used participants in a previous experiment using the same tethered ankle
to measure the volume of carbon dioxide and oxygen exchanged on exoskeleton that had provided the largest energetic benefits owing to
each breath. A standard equation was used to compute metabolic exoskeleton assistance so far4. The Generic assistance pattern allowed
energy expenditure in watts for each breath44. Metabolics measure- us to estimate the contributions of personalization through data-driven
ments during the real-world walking experiment and validation were optimization in tests with the tethered emulator.
collected with portable respirometry equipment worn using a vest During Speed-adaptive assistance, the exoskeleton used estimates
on the participant’s torso (K5, COSMED). Metabolics measurements of walking speed to select assistance parameters expected to be
during other exoskeleton experiments were collected with tethered more effective at that speed. The speed-adaptive control approach
respirometry equipment (Quark CPET, COSMED). Metabolics data is described in detail in ‘Speed-adaptive controller and validation
were recorded during a quiet standing condition at the beginning of experiment’. In the Speed-adaptive condition, the controller interpo-
each day of experiments. This quiet standing value was removed from lated between separate sets of assistance parameters that had been
subsequent measurements to isolate the energy cost associated with optimized for the same participant at walking speeds of 0.75 m s−1,
walking and remove any absolute error associated with respirometry 1.25 m s−1 and 1.75 m s−1. This condition was used to test the efficacy of
system calibration. The change in metabolic rate as a percentage of a the speed-adaptive control approach for handling speed variations
baseline condition, measured within the same experiment, is reported during treadmill walking with the tethered exoskeleton emulator.
as the primary outcome to account for differences in respirometry In the Generic Speed-adaptive assistance condition, the
equipment calibration coefficients between data collections. Partici- speed-adaptive controller selected assistance parameters expected to
pants refrained from all food and drink except for water for at least 3 h be more effective at that speed for an average participant. The control-
before experiments that included respirometry measurements to avoid ler interpolated between separate sets of Generic assistance parameters
confounds from the thermal effect of food. Steady-state metabolic for each walking speed, computed by averaging the optimized assis-
cost was computed by averaging data from the last 3 min of each 6-min tance profiles from the tethered exoskeleton experiments in this study
condition. Cumulative metabolic cost was computed as the total energy (Fig. 2). The generic parameters for 0.75 m s−1 and 1.75 m s−1 were com-
expended during the condition45, including the metabolic cost beyond puted by averaging across three participants’ optimized parameters,
that of quiet standing for 3 min following completion of the condition, whereas the generic parameters for walking at 1.25 m s−1 were computed
Article
by averaging ten participant’s optimized parameters. The Generic
Speed-adaptive assistance condition was compared with Real-world Data-driven optimization
Optimized assistance during overground walking with the untethered Data-driven optimization personalized assistance using a data-driven
exoskeleton. Ideally, the Generic Speed-adaptive parameters would classification model to determine which exoskeleton control param-
have been taken as the average values from real-world optimization, eters provided the largest benefits for each person. The participant
but those values were not yet known at the time the experiment was walked while receiving a sequence of different patterns of exoskeleton
conducted. The Generic Speed-adaptive condition provided the best assistance, each defined by the corresponding control law. During
available comparator for isolating the benefits of personalization dur- laboratory-based experiments, participants walked on a treadmill
ing overground experiments with the untethered exoskeleton. for 30 s for each control law. During real-world experiments, partici-
Metabolic Optimized assistance was personalized based on met- pants walked overground for 44 continuous steps for each control
abolic measurements using a previously established optimization law. A fixed number of exoskeleton control laws, comprising one gen-
method2 for identifying the exoskeleton control parameters that mini- eration of the evolution-inspired optimizer, were then ranked using
mize the metabolic cost of walking for a specific person. To perform the data-driven classifier. The optimizer then updated its estimate
metabolic optimization, a participant walked on a treadmill while of the optimal parameters and generated a new set of control laws to
receiving exoskeleton assistance. The same assistance profile, or ‘con- evaluate. The following paragraphs detail what type of data were col-
trol law’, was provided to the participant for 2 min while respirometry lected, how the data were processed, how the data-driven classification
measurements were recorded and the steady-state metabolic cost of model evaluated control laws, how the optimizer was updated based
walking for that control law was estimated18. Participants repeated on the data-driven classifications and how a new set of control laws was
this process of walking on a treadmill for 2 min per control law until a selected for evaluation in the next generation of optimization (Fig. 1).
fixed number of control laws had been evaluated. We refer to the set Exoskeleton torque control parameters, defined by the control law,
of control laws to be evaluated as one ‘generation’ of control laws—the were fixed within each evaluation period. The person experienced sev-
terminology used in evolution-inspired optimization strategies such eral control laws before the data-driven model processed data and the
as the covariance matrix adaptation evolutionary strategy (CMA-ES)26. optimizer updated its estimate of the optimal assistance parameters
Each generation of control laws must be tested before updating the and generated a new set of control laws to evaluate.
estimate of the optimal control law and generating a new generation of The data-driven model input consisted of carefully processed porta-
control laws to test. On the basis of heuristics relating generation size ble sensor data, including ankle angle and ankle velocity measurements
to the number of parameters to be optimized26, the lab-based experi- and the control law parameters that set the pattern of exoskeleton
ments with four optimization parameters had participants complete torque. The angle and velocity measurements were sampled using an
eight control laws per generation. After each generation, an optimizer absolute rotary encoder at the ankle joint of the exoskeleton worn on
(CMA-ES) ranked the control laws in order of metabolic cost, updated the left leg. The control law parameters consisted of four values: peak
the optimization parameters and selected a new set of promising torque magnitude, peak time, rise time and fall time2.
control laws to evaluate. This optimization approach was established Portable sensor data were processed by segmenting the ankle angle
in previous experiments that demonstrated large improvements in and velocity measurements by gait cycle and then discretizing the data
the metabolic energy cost of walking and running with exoskeleton for each gait cycle into a discrete number of bins. The gait cycles were
assistance2,4,47. This optimizer was selected because it is sample effi- segmented whenever a heel strike was detected by the pressure-sensing
cient, meaning that it requires relatively few evaluations to reach a insoles. The first six gait cycles of data were discarded to avoid con-
reliable estimate of the optimal parameters. In the context of exoskel- founds from fast adaptation48 by the person in response to the new
eton optimization, that means fewer exoskeleton control modes to be assistance pattern—in pilot tests, we found that data from these first
experimentally tested on the human participant, which is important six steps exhibited substantial changes in ankle kinematics, while sub-
both to study design and to real-world use of devices that personalize sequent strides were more consistent. The remaining gait cycles were
assistance. Metabolic Optimized assistance was used to validate the discretized by averaging the measurements within each of 30 discrete
data-driven optimization approach in the first experiments with the bins and then averaging each bin across the gait cycles for that control
tethered exoskeleton emulator. law. The processed data were reshaped into a single vector with 64 val-
Data-driven Optimized assistance was personalized using data-driven ues: 30 binned values for the ankle angle across the gait cycle, and 30
optimization. The data-driven optimization used the same optimiza- binned values for the ankle velocity across the gait cycle, and 4 values
tion framework as the metabolic optimization, except that it used the for the torque parameters. The model input consisted of the vector of
data-driven classifier, rather than indirect respirometry measure- data for one control law subtracted from the vector of data for a dif-
ments, to perform the ranking step. The classifier was trained on data ferent control law, which also comprised 64 values. This difference in
from a previous laboratory experiment4 and compared control laws the sensor measurements provided the model with information about
based on the exoskeleton torque parameters applied and the result- how the person’s movements and exoskeleton torque differed between
ing ankle angle and ankle angular velocity (Fig. 1 and Extended Data the two control laws. The choice to segment data by gait cycle follows
Fig. 1). The data-driven optimization condition was applied in tests of our previous findings that data-driven models can more accurately
the data-driven optimization approach using the tethered exoskeleton estimate metabolic energy expenditure from sensors worn by unas-
emulator. sisted humans when the data are formatted in this way23.
Real-world Optimized assistance used speed-adaptive control with The data-driven classification model was trained to compare two
parameters that were personalized using the data-driven optimization control laws at a time, determining which control law was estimated to
approach while walking with the untethered ankle exoskeleton under have provided a larger reduction in the metabolic cost of walking. The
naturalistic conditions. The controller used the same speed estimation data-driven classification model was a logistic regression model. To
and adaptation approach as with the Generic Speed-adaptive condi- train the data-driven classifier, we input previously collected data4
tion, except that the parameters for each speed were personalized that included portable sensor data, in the form of exoskeleton joint
to the individual participant using opportunistic optimization. The angles and velocities, and ground-truth labels, in the form of meta-
Real-world Optimized parameters were computed using the approach bolic measurements, for many exoskeleton control laws. The sensor
detailed in ‘Opportunistic optimization approach’. The Real-world data were taken as input into the model to estimate the likelihood that
Optimized condition was applied in outdoor and treadmill tests with the first of the compared control laws resulted in a lower metabolic
the untethered exoskeleton. cost of walking compared with the second control law. The resulting
probability was a continuous value from 0 to 1, with 1 indicating the The model weights on differences in torque parameter values indi-
highest likelihood that the first control law reduced the metabolic cated that lower metabolic rate was associated with a later time of
cost of walking more than the second control law. The ground-truth peak torque and, to a lesser extent, a larger peak torque magnitude.
labels were computed by subtracting the measured metabolic costs, Exoskeleton assistance was governed by a torque pattern defined by
estimated from 2 min of respirometry data, for the two control laws. A four parameters: peak torque magnitude, peak time, rise time and fall
label with a negative value indicated that the first control law was more time2. These four parameters had allowable ranges of 0 to 1 Nm kg−1, 40%
beneficial, meaning that it reduced the metabolic cost of walking more to 55% stride, 20% to 40% stride, and 10% to 20% stride, respectively.
than the second control law. A positive-valued label indicated that Fall time was further constrained to be at most equal to the difference
the second control law was more beneficial. The previously collected between peak time and the time of toe-off, which prevented application
training data were from an experiment in which 10 participants walked of torque during the swing phase. The largest single contributor to clas-
under approximately 3,600 different exoskeleton control laws4. When sification based on exoskeleton torque, constituting about 60% of the
training the data-driven classifier, we used regularization, a technique total, favoured applying peak torque at a later time in the gait cycle. As
that encourages simpler models and avoids overfitting to training the peak time was constrained, this term had the effect of maintaining a
data, to improve model estimates for new data points that were not peak time close to the upper limit of 55% stride. Consequently, fall time
in the training set. In this case, we used a lasso regularization term was effectively constrained to its lower bound of 10% stride. The large
that penalized the absolute value of the model weights multiplied by model weight on peak time is consistent with previous observations
a regularization parameter with a value of 1. that the timing of ankle exoskeleton assistance is important51, and that
The data-driven classification model was trained to capture a rela- later onset of torque assistance can correspond to larger improvements
tionship between leg movement, exoskeleton torque parameters in metabolic rate52. The data-driven model also favoured larger peak
and the metabolic cost of walking with assistance. The linear weights torque magnitudes, with the associated term contributing about 4%
used by the data-driven classifier are visualized using a colour code in to the total classification. Interestingly, peak torque was not driven to
Extended Data Fig. 1. To aid interpretation, we also overlay the mean its upper limit for most participants or conditions, and the classifica-
difference for each model input as a black line. This was calculated as tion contribution of this term was about ten times less than the sum of
the value from the control law resulting in lower metabolic rate minus contributions from ankle kinematics. This suggests that how a person
the value from the control law resulting in higher metabolic rate, aver- reacts to exoskeleton assistance is more important for determining
aged across all pair-wise comparisons, such that the sign of the mean metabolic rate than the magnitude of the torque and power provided
difference is meaningful. We also provide the cumulative contributions by the exoskeleton.
of each term in the model to classification over the entire training set. Exoskeleton control laws were ranked using the probability values
The percent contributions are calculated as the absolute value of the estimated by the data-driven classifier. Each pair of control laws that
product of the model weight and the difference input, summed over were passed to the data-driven classification model yielded one prob-
all pair-wise comparisons, divided by the sum over all model terms. ability value, defining whether the first control law was expected to have
Even linear data-driven models can be difficult to interpret because of provided a larger benefit than the second control law. All possible pairs
the complex interactions between model terms through the dynamics of control laws were classified with the data-driven model to obtain
of the underlying system, which can be nonlinear and coupled. In this a complete set of probability values. Each control law was scored by
case, the underlying system is a human walking with an exoskeleton, summing the probabilities from all pairs that included that control law.
and we expect strong interactions between exoskeleton torques, joint The control laws were ranked by the magnitudes of their scores, with a
velocities and joint angles, and between states at different times in the larger value indicating that the control law was more likely to provide a
gait cycle. The model may be capturing aspects of these interactions larger reduction in the metabolic cost of walking (Fig. 1). This ranking
in non-obvious ways. Nevertheless, we can gain some intuition about step replaced the previous approach based on metabolic measure-
the relationships that the model may have identified if we consider the ments from indirect respirometry equipment, allowing optimization
effects of key model terms independently. to take place outside the laboratory using inexpensive sensors and a
The model weights associated with differences in ankle kinematics microcontroller on the exoskeleton. Using the data-driven ranking,
suggest that a lower metabolic rate was associated with increased ankle the optimizer then updated its internal parameters and generated a
plantarflexion at toe-off, while guarding against premature onset of new set of control laws to evaluate. The new estimate of the optimal
push-off, excessive plantarflexion velocity and reduced dorsiflexion control law was equal to the weighted average of the best-performing
mid-stance. The largest single contributor to classification based on control laws. New control laws were selected from a distribution around
ankle kinematics, constituting about 10% of the total, favoured a larger this estimate of the optimum, with the shape of the distribution set
plantarflexion angle at 62% stride, the time of toe-off during normal by the covariance matrix and the spread of the distribution set by the
walking. A large negative weight on the difference in ankle velocity at convergence parameter.
48% stride seemed to penalize conditions that resulted in premature The data-driven optimization process can be better understood by
onset of ankle push-off. A sequence of negative weights on ankle plan- working through these steps using example data. Imagine that we have
tarflexion velocity during push-off seemed to favour slower, smoother three control laws, labelled 1, 2, and 3, which happen to be in order of
movement during that phase. Taken together, these velocity regulation increasing metabolic cost and decreasing peak torque. Imagine that
terms constituted about 16% of the total classification. Smaller nega- these control laws had identical torque timing parameters and resulted
tive weights on ankle angle at 38% stride and ankle velocity before the in identical ankle angles and ankle velocities. When performing the
onset of push-off suggest a preference for conditions with greater three pair-wise comparisons (Fig. 1c), the differences between ankle
mid-stance dorsiflexion. The model terms associated with ankle angle angle and ankle velocity would be zero. The differences between torque
and velocity were most informative during late stance, when the con- parameters (ΔC12, ΔC13 and ΔC23) would each be a vector with one posi-
centric contractions of the plantarflexor muscles are less efficient and tive value followed by three zeros. When taking the dot product of the
exoskeleton torque may have the most capacity to reduce metabolic parameter differences with the model weights on control parameters,
cost49,50. During the leg swing phase, model terms contributed little which are all positive (Fig. 1d), the pair coefficients (w12, w13, and w23)
to the total classification, consistent with expectations for an ankle would all be positive scalars. For each of these pair coefficients, the
exoskeleton that produced no torque when the foot was off the ground. logistic function (Fig. 1e) would return a probability greater than 0.5,
The sum of all model weights associated with ankle angle and velocity indicating that the first control law in the pair is likely to have a lower
were 10% and 23% of the total, respectively. metabolic cost than the second control law. Let us imagine that each of
Article
the probability values (p12, p13, and p23) was 0.9. When performing the The second experiment was used to evaluate the same set of assis-
control law scoring step (Fig. 1f), the score for the first control law (s1) tance conditions at additional speeds and treadmill grades. A subset
would be the sum of p12 and p13, or 1.8. The score for the second control of healthy adult participants from the first experiment (n = 3, 3 men;
law (s2) would be the sum of p21 and p23. As p21 is the complement of age, 24.0 ± 2.0 yr; body mass, 66.0 ± 8.0 kg; height, 1.76 ± 0.05 m) com-
p12, p21 = 1 – p12 = 0.1. Thus, s2 would equate to 1. The score for the third pleted the experiment. Participants completed the same experimental
control law (s3) would be the sum of p31 and p32, or 0.2. Thus, the scores protocol used in the first tethered exoskeleton experiment for three
would correctly rank the control laws in terms of metabolic cost. The additional walking conditions: walking at a slow speed of 0.75 m s−1, a
optimizer would then use this ranking to perform an update, estimating fast speed of 1.75 m s−1, and on a 10° incline at 1.25 m s−1.
that the optimal torque parameters were close to those of control law
1, but slightly offset towards control law 2. The optimizer would then Speed-adaptive controller and validation experiment
select new control laws to evaluate, drawn from a distribution around We developed a speed-adaptive controller that adjusted exoskeleton
the new estimate of the optimal parameters (Fig. 1g). assistance based on walking speed. During real-world walking, people
naturally vary their speed27. We hypothesized that adjusting exoskel-
Tethered optimization experiments eton assistance based on walking speed would provide larger metabolic
To compare the efficacy of data-driven optimization to a range of other reductions than a constant pattern of assistance. We estimated the
assistance conditions, we conducted tethered exoskeleton experiments walking speed of each step using a linear model, relating measured
in an indoor laboratory setting. Participants wore tethered bilateral stride durations to measured walking speeds (Extended Data Fig. 2).
ankle exoskeleton emulators43. Exoskeleton assistance was governed Walking speed estimates from each step were used to interpolate exo-
by a torque pattern characterized by four parameters: peak torque skeleton assistance parameters from those optimized at a range of
magnitude, peak time, rise time and fall time2. The exoskeleton control fixed speeds (Fig. 3a).
loop ran at 1,000 Hz on a real-time computer (Speedgoat). Exoskeleton During speed-adaptive control, walking speed from one step was
sensor measurements were recorded at a rate of 2,000 Hz, including used to select the assistance parameters for the following step. We
pressure values from shoe insoles, commanded torque parameters, expect this approach to perform well when changes in walking speed
measured torque, ankle angle, and ankle velocity. Measurements were occur slowly, or when there are rapid changes in speed but they consti-
used to estimate time within the gait cycle as a percentage of the total tute a small portion of total steps, as in natural human gait. Our experi-
gait cycle time, which was used to calculate the desired torque. Torque mental data are consistent with the observation that most acceleration
tracking was accomplished using a combination of classical feedback and deceleration occurs within a few steps at the start and end of each
control and iterative learning, which accounted for errors that consist- walking bout. The expected stance duration was also adjusted based
ently occurred at the same point in the gait cycle on each step30. on speed estimates, following an approach established in previous
Two tethered exoskeleton experiments were used to evaluate the research53. In future studies, the speed-adaptive controller could be
effectiveness of various assistance conditions. The first experiment improved to deliver more effective assistance during rapid changes
compared assistance conditions while participants walked at 1.25 m s−1, in gait speed by incorporating instantaneous estimates of walking
a normal walking speed previously used for metabolic optimization speed54,55 and stance duration. Acceleration regimes could also be
experiments2,4. Healthy young adults (n = 9, 5 men and 4 women; age, considered, with a binning approach analogous to the one used for
24.8 ± 1.8 yr; body mass, 65.3 ± 8.0 kg; height, 1.73 ± 0.07 m) completed speeds in this study, to allow optimization of assistance specific to
a two-day experimental protocol. On the first day, participants per- acceleration and deceleration phases. Other approaches, such as those
formed experiments to personalize assistance parameters with meta- using phase-based control55 or adjusting assistance based on changes
bolic optimization and data-driven optimization, in a randomized in joint kinematics rather than walking speed56, may be beneficial for
order. Participants completed eight generations of optimization for generalizing to a large set of activities.
each approach. Each generation consisted of eight control laws. The We conducted a third tethered exoskeleton experiment to evaluate
optimizations were initialized with the Generic assistance parameters, whether adapting assistance to variations in walking speed could pro-
corresponding to the average of the optimized parameters identi- vide larger reductions in metabolic cost than a fixed generic assistance
fied for a previous group of expert participants4. The optimizations profile. Healthy young adults (n = 3, 3 men; age, 24.0 ± 2.0 yr; body
were initialized with the covariance matrix set to the identity matrix mass, 66.0 ± 8.0 kg; height, 1.76 ± 0.05 m) completed the experiment.
and a scaling factor that corresponded to 20% of the range of the nor- These participants had previously completed the first two tethered exo-
malized assistance parameters (a sigma value of 0.1). The metabolic skeleton experiments, providing Data-driven Optimized parameters
optimization control laws lasted 2 min, which allowed steady-state for walking speeds of 0.75 m s−1, 1.25 m s−1, and 1.75 m s−1. Participants
metabolic cost to be estimated from respirometry data with a good walked on a treadmill while the speed varied sinusoidally from 0.75 m s−1
balance between the time required for each control law and estimation to 1.75 m s−1 with a period of 30 s. Participants completed assistance
accuracy2. This led to a total evaluation time of 128 min of walking. conditions including walking in Normal Shoes and walking with the
During data-driven optimization, each control law was evaluated for exoskeletons under Zero Torque, Generic assistance (which did not
30 s, sufficient to obtain an accurate estimate of participant motions, change in response to changes in speed) and Speed-adaptive assistance
which were nearly steady following the rapid adaptation phase23. This (using the optimized control parameters previously identified for each
required a total evaluation time of 32 min of walking. For each par- participant). The validation tests were randomized and presented in
ticipant, the parameters identified using data-driven and metabolic a double-reversal ABCDDCBA order to mitigate the effects of noise in
approaches were similar. For example, optimized peak torque values the metabolics measurements and trial order.
were well correlated across methods (R2 = 0.76, P = 1.4 × 10−4, n = 9).
On the second day, participants performed a standing rest condition Untethered exoskeleton design
followed by assistance conditions including Normal Shoes and walking The untethered exoskeleton was designed to provide the optimized
with the exoskeletons under Zero Torque, Generic assistance, Meta- assistance parameters from the tethered exoskeleton experiments
bolic Optimized assistance and Data-driven Optimized assistance. under real-world conditions. The maximum peak torque magnitude
The assistance conditions for these validation tests were randomized for the optimized assistance during the tethered exoskeleton study was
and presented in a double-reversal order, as ABCDDCBA, to mitigate 54 Nm when walking at a moderately fast speed of 1.5 m s−1. The motor
the effects of trial order related to within-day adaptation and fatigue. and power transmission elements were designed to robustly provide
Each condition lasted for 6 min and included metabolic measurements. this level of assistance. A portable battery was selected to allow 30 min
of continuous walking on a single charge. The device was designed to maximum voltage of 25.2 V and the battery life experiment was stopped
be lightweight to reduce the metabolic power required to carry the once the battery voltage reached 21.6 V, corresponding to a cell volt-
exoskeleton. age of 3.6 V, the minimum safe level recommended for discharging a
The untethered exoskeleton had a mass of 1.2 kg for each ankle. lithium polymer battery. During testing, cell voltage was monitored by
Many of the mechanical elements were the same as in the tethered a safety regulator and an audio alarm was played once the cell voltage
exoskeleton, including the frame, shoe and pressure-sensor insole. New reached 3.6 V. We found that the 0.3-kg battery used in real-world tests
elements included the portable motor, drum-and-cable transmission, allowed 36.3 min of operation under these conditions.
electronics, and battery (Extended Data Fig. 3). A set of computer-aided The design of the untethered device was guided by previous
design files and a bill of materials are provided as Supplementary Data 2. laboratory-based ankle exoskeletons, incorporating design elements
The brushless motor (AK80-9, CubeMars) contained a single stage 9:1 that allowed for large assistive torques while maintaining comfortable
gear ratio and internal motor driver electronics. This gearmotor has a forces on the body43. The shoe, carbon fibre struts and calf spacers
rated peak torque of 18 Nm, a no-load speed of 25 rad s−1, and a mass of were designed to be interchangeable to fit different participants, fol-
0.5 kg. We selected this motor based on simulations with a simplified lowing best practices for fitting62. The motor-and-drum transmission
model that predicted it would be capable of applying the patterns of and heel spur were designed to be one size fits all, with interchange-
ankle torque and velocity that corresponded to optimized assistance able shoes and spacers accommodating differences in foot size and
in the tethered exoskeleton experiments, assuming an additional 5:1 mediolateral dimensions of participants’ legs. It might at first appear
gear ratio from the drum to the heel spur. that the force applied by the cable between the drum and heel spur
The custom drum was machined from 7075 aluminium, with a radius would pull the exoskeleton down the leg, but the rigid exoskeleton
of 0.020 m. A cable connected the heel spur to the motor drum. The frame allows the axial component of this force to be reacted out at the
heel spur had a maximum lever arm (the distance from the centre exoskeleton joint rather than as shear on the person’s skin43. Thus, only
of the ankle joint to the rope tie-off point) of 0.115 m. The lever arm a normal force is applied to the shank of the leg, which allows for more
decreased as the ankle plantarflexion angle increased, with a singular- comfortable application of high torques63. The carbon fibre frame of
ity at a maximum plantarflexion angle of 55° ensuring that no ankle the exoskeleton used stiff material and a cross-section with a high-area
torque could be applied to hyperextend the ankle joint. The torque moment of inertia to prevent meaningful deflection during loading.
assistance profile of the exoskeleton was not impacted by changes in As the system regulated exoskeleton joint torque, rather than motor
the lever arm because torque was measured directly at the ankle; strain current or velocity, and as torque was measured directly at the joint,
gauges on the superior and inferior surfaces of the heel lever directly compliance and dissipation in the transmission, exoskeleton frame and
sensed bending moment independent of cable force. This allowed for human–exoskeleton interface did not affect the accuracy or consist-
accurate torque control without explicitly correcting for joint angle. ency of the applied torque.
When the motor applied torque to the drum, a force was generated in The design of the untethered exoskeleton required several trade-offs.
the cable, which then transmitted this force to the heel lever, creating The highest design priority was providing a peak torque of 54 Nm during
a torque about the ankle joint of the exoskeleton. The drum-and-cable walking at 1.5 m s−1, specified from previous optimization experiments,
transmission had the added benefit of being backdrivable, avoiding with the least mass possible. We considered several factors to ensure
the possibility of force spikes that can be produced by classically stiff that the motor would provide 54 Nm during operation. We simulated
actuators57,58. The cable could also be driven to a slack state to allow the torque needed to provide the desired assistance, overcome trans-
the person to move freely when desired, an important capability that mission inefficiencies, and accelerate the mass of the motor rotor and
prevents interference when not providing assistance59. drum as required to track ankle movements during walking at 1.5 m s−1.
The untethered exoskeleton electronics consisted of a microcon- The motor had to operate at a safe steady-state temperature to prevent
troller, portable sensing elements, a motor driver integrated into the damage to the windings. A brushless motor was selected for its rela-
motor and a rechargeable battery. The untethered exoskeleton used tively high efficiency and peak torque. This untethered exoskeleton
a Raspberry Pi 4b microcontroller to read sensor data and perform was designed for the optimized parameters of our experimental par-
real-time control and optimization at a rate of 200 Hz. A breakout board ticipant group, and other participants may require a different device
enabled sensors to interface with the microcontroller. A step-down with different balance between torque and weight to provide the same
voltage converter enabled the electronics to be safely powered by a reductions in the metabolic cost of walking.
portable battery. The portable sensing elements included a rotary Another important decision was whether to place the motor and
encoder in the ankle joint that measured ankle angle and velocity, a electronics near the assisted joint or closer to the torso. The energy
pressure-sensing insole in the shoe, a set of strain gauges in a full Wheat- cost of carrying mass at distal joints is high33, suggesting a relocated
stone bridge configuration applied to the heel spur to measure torque, drive approach with heavy motors carried more proximal to the centre
and an amplifier (IAA100, Futek) to allow measurement of strain-gauge of mass of the body. We considered mounting the motor and electronics
signals. The pressure-sensing insole had pressure sensors located at the at the hip and using a Bowden cable to transmit forces to the ankle joint.
heel, fifth metatarsal, distal phalanx of the great toe and the first meta- Bowden cables have an inner cable that moves relative to an outer con-
tarsal. Fusing information from these different sensors enabled robust duit like a bicycle brake. This introduces complex transmission dynam-
estimation of stance and stride period while providing measurements ics, including stick–slip friction, history dependence and a dependence
to extract information for optimizing assistance. This choice of sensors on leg posture, making torque control more challenging, reducing
was guided by the design heuristic that multiple modes of sensing are control bandwidth and decreasing energy efficiency. The cables and
important for effective exoskeleton control60. Muscle electrical activity additional electrical wires also add to the weight of the system. For
could have provided additional information for control, but with the these reasons, we selected a drum-and-cable transmission located on
added challenge of handling noise from sensors placed on the skin61. the shank of the leg. Locating motors and electronics near the assisted
The total weight of electronics was 0.15 kg. joint resulted in more efficient power transmission, lower transmission
The entire system was powered by a lithium polymer battery with a compliance, better control bandwidth and less total weight.
nominal voltage of 24 V, a capacity of 1,300 mAh, and a weight of 0.3 kg. Our untethered exoskeleton was designed to allow tests of real-world
Battery life was experimentally evaluated under the most demand- personalization and resulting mobility benefits during naturalistic
ing assistance pattern, characterized by a peak torque of 54 Nm and walking in a community setting. A significant amount of additional
late timing of peak torque. Tests were conducted while walking on a engineering would be required to make this device ready for everyday
treadmill at a speed of 1.5 m s−1. The battery was initially charged to a use by consumers. Everyday use would require easier donning and
Article
doffing, a more comfortable interface, more robust electronics hard- that set the spread of the distribution from which to draw parameters
ware and more intuitive, independent control, for example, utilizing a for the next generation. Optimizations were performed for three bins
smartphone app. In addition, the exoskeleton would have to be tested of walking speed: less than 1.22 m s−1, between 1.22 m s−1 and 1.38 m s−1,
to ensure functionality during additional common activities such as and greater than 1.38 m s−1. These speeds were chosen based on the
navigating stairs, and to ensure that it did not interfere with common 33rd and 66th percentile of real-world walking speed distributions27,
activities such as sitting and driving. While we did not directly evaluate resulting in an equal expected likelihood for the participant to walk
descending stairs in this study, we did notice that the long heel spur in each bin. Speed-adaptive control interpolated assistance based on
required participants to walk carefully to avoid hitting the previous the speed of each individual step (Extended Data Fig. 2). When a suffi-
step. This design choice was made for convenience, allowing us to use cient number of steps were collected for one control law, the estimated
as many elements from our previous tethered exoskeleton design as walking speeds for all steps during that control law were averaged, the
possible. A less obtrusive transmission would be needed for a con- corresponding speed bin was selected, and data were stored for the
sumer device. The commercially available Dephy ExoBoot64 provides optimization process. When a complete generation of control laws were
an example of a more streamlined design; it has no spur behind the heel collected for a speed bin, control laws for that bin were ranked and the
of the shoe, has simple donning and doffing features, and has minimal optimization parameters for that bin were updated. The estimate of
structure on the medial side of the leg, making it a good candidate for the optimal assistance parameters for the other speed bins were also
extended use in a large range of activities. Other autonomous ankle adjusted by a lesser amount, with the magnitude of the adjustment
exoskeletons10,17 demonstrate complementary ways of designing hard- being proportional to the value of the convergence parameter, σ, for
ware that is more compatible with everyday use. With increased torque that bin (Extended Data Fig. 4). This allowed parameters in all speed
capacity, more accurate torque control and real-world personalization bins to update more quickly at the beginning of the optimization, with
using the approach described here, we expect commercial devices decreased across-speed influences as the optimizations within each
could achieve similar reductions in metabolic rate. speed bin converged.
We chose to optimize a set of assistance parameters for each of three
Opportunistic optimization approach bins of walking speed, but it is possible to formulate this optimiza-
We overcame the challenges of optimizing assistance during short tion in different ways. The data-driven classifier requires comparisons
bouts of walking at varying speeds by opportunistically accumulat- of control laws at similar walking speeds. A larger number of bins of
ing data across many bouts and binning by speed. This opportunistic walking speeds could be used to provide more granular speed-based
optimization approach used the same data-driven classification model adaptation, at the expense of additional time to optimize a larger num-
and optimization method that were validated in the tethered experi- ber of assistance parameters. It may also be possible to simultane-
ments, with the addition of a check that sufficient consecutive steps ously solve for a larger set of control parameters that fully define the
had been collected for each control law and a method for addressing speed-adaptive controller, but this would introduce challenges related
a wide range of speeds (Extended Data Fig. 4). to the larger parameter space, interaction effects between parameters,
The opportunistic optimization method checked that sufficient and poorly conditioned maps between parameters that have a strong
steps had been collected before moving on to the next control law. We effect on assistance at one speed and little effect on assistance at differ-
chose the requirement of 44 steps to approximate the durations used ent speeds. Instead, we opted for a small set of speed bins, with a rela-
in the tethered data-driven optimization experiments. If sufficient tively simple approach to updating the optimal parameter estimates.
continuous steps were not collected before the end of the walking
bout, the optimizer would start over with the same controller on the Real-world optimization experiments
next bout. Once sufficient strides were collected, the next control law In the real-world optimization experiments, we used the untethered
was applied for that speed bin. As with the tethered experiments, the exoskeleton to optimize assistance during naturalistic bouts of walking
first six strides of data were discarded to avoid confounds related to and then evaluated the optimized assistance profiles under real-world
rapid adaptation to a new exoskeleton control law. and treadmill conditions.
The same data-driven classification model used in the tethered Healthy adult participants (n = 10, 6 men and 4 women; age,
exoskeleton experiments was used for the real-world optimization, 24.2 ± 1.8 yr; body mass, 67.0 ± 8.2 kg; height, 1.72 ± 0.07 m) completed
but a different set of assistance torque parameters were optimized. a two-day protocol. On the first day, participants walked outside in a
The torque parameters for peak time and fall time were fixed to the public setting along a path consisting of concrete, asphalt and brick
average values of the Data-driven Optimized parameters from the first sidewalks (Fig. 5b) for approximately 1 h while the untethered exoskel-
tethered exoskeleton experiment (54.6% of the gait cycle and 10.0% of eton provided assistance and performed data-driven optimization. To
the gait cycle). We fixed the values of peak time and fall time because emulate natural walking, the participants received audio cues to tell
the optimized values changed little across speeds and participants, them to start and stop walking bouts. The durations of these bouts were
indicating that fixed values may be sufficient. The optimized values randomly drawn from a preselected distribution (Fig. 5d) that matched
of peak torque and rise time varied substantially across speeds and naturally occurring bout durations34. Participants stood at rest between
participants, and so these parameters were optimized in untethered bouts for a randomized duration of 5 s to 10 s. To encourage a normal
exoskeleton experiments. Optimizing two, rather than four, torque range of speeds, we provided participants with audio prompts, such
parameters reduced the dimensionality of the optimization, requir- as “Walk as if you were walking to catch a bus” and “Walk as if you were
ing only six, rather than eight, control laws to be collected for each walking a small dog”, at the start of each bout. A previous study28 demon-
generation of optimization. Reducing the number of control laws to be strated that these prompts were associated with different self-selected
evaluated per generation allowed for more generations to be completed walking speeds, and we expected that participants would adopt similar
within a set experiment time, providing more frequent optimization speeds. We randomly sampled from a distribution of speeds (Fig. 5c)
updates and a better estimate of the optimal values. This may have that mimicked natural walking patterns measured in a previous study27.
come at the cost of suboptimal assistance timing parameters for some On the second day, participants performed outdoor and treadmill
participants. validation tests to evaluate the benefits provided by Real-world Opti-
Once data for all the control laws in a generation were collected, the mized assistance. For the outdoor validation, participants walked along
data-driven classification model ranked the control laws. The optimizer a 566-m path in the same public setting with a fixed ordering of bouts
used this ranking to update its estimate of the optimal parameters and of specific distances and corresponding speed prompt commands that
to adjust internal parameters, such as the convergence parameter (σ) were selected to match real-world distributions27,34. Distances were
set using cones to mark stopping locations, which ensured consistent outcomes, so we considered studies conducted at within 10% of the
distances for each bout. Participants completed this outdoor course speeds and inclines that we tested. Before conducting our final experi-
once for each condition, including Real-world Optimized assistance, ment, we selected the speeds (1.25 m s−1 and 1.5 m s−1) and inclines (10°)
Generic Speed-adaptive assistance and Normal Shoes. The ordering that captured the largest percent reductions in metabolic rate that had
of the conditions was randomized to minimize effects of testing order previously been observed for any exoskeleton study in the literature. We
(Extended Data Table 1). The double-reversal protocol, used in the compared with previous exoskeleton studies with at least five partici-
first three laboratory experiments, was not used because the outdoor pants, because studies reporting data from fewer tests are difficult to
experiments took significantly more time owing to the longer trial interpret owing to measurement noise and inter-participant variability.
time, varying self-selected walking speeds, short bouts of walking, and We compared with previous studies in which the metabolic cost of walk-
rest periods between bouts and conditions. Each real-world condition ing was calculated using standard techniques, by averaging respirom-
required about 15 min, compared with about 8 min for each treadmill etry measurements during the last 2 min or 3 min of a 5-min or 6-min
condition. Outdoor and indoor tests of Real-world Optimized assis- steady-state treadmill condition. One previous exoskeleton study65
tance were conducted on the same day to avoid confounding effects was excluded because steady-state metabolic cost was computed by
from differing respirometry system calibrations. The total walking taking the median of respirometry measurements. We found that using
time for these two experiments was about 1.5 h, and we found that the median rather than the mean to compute metabolic rate in our
participants were not able to complete the additional 1.5 h of walking untethered exoskeleton study increased the magnitude of the reduc-
that would have been required for a double-reversal approach without tions in metabolic cost by an average of 7% across participants. This is a
experiencing fatigue. For the 3 min following completion of the path, large amount compared with the total improvement of 23%, indicating
participants stood at rest while respirometry data were collected to that the median and mean measurements are not equivalent. We were
capture the total metabolic cost of completing the course. The duration not able to obtain the data from the previous study that would have
of walking for each bout was timed with a stopwatch. Walking speed for allowed computation of the mean percent change in metabolic rate.
each bout was computed by dividing the fixed distance for that bout To keep Extended Data Fig. 5 legible, we only depict studies report-
by the time spent walking during that bout. Walking speed for each ing results within a 5% change in metabolic rate of the best previous
condition was calculated as the total distance travelled divided by the value for that condition category. There are several other untethered
total time spent walking while navigating the course. exoskeletons that have provided some reduction in metabolic rate
The indoor validation consisted of a standing rest condition followed under conditions similar to those tested in this study. For example,
by six treadmill conditions, each lasting 6 min. Participants walked on the Dephy ExoBoot, the commercially available exoskeleton with the
a treadmill at 1.25 m s−1, at 1.5 m s−1, and on an incline of 10° at 1.25 m s−1. most similar features to the prototype tested in this study, can provide a
Participants completed each treadmill speed and grade twice, once 5.2% reduction in metabolic energy consumption compared with walk-
with Real-world Optimized assistance, as identified during the outdoor ing with Normal Shoes while walking on a treadmill with time-varying
optimization period, and once with Normal Shoes. The ordering of speed64. Another technologically mature untethered exoskeleton, the
conditions was randomized, with a constraint that the exoskeleton MyoSuit Beta, has shown that hip assistance during outdoor uphill
would only be donned and doffed one time to reduce experiment time walking can reduce metabolic rate compared with wearing the exo-
(Extended Data Table 2). We did not use the double-reversal protocol skeleton in Zero Torque mode37. Sufficient data are not yet available to
in these tests because we found that participants could not reliably estimate the benefits compared with walking without the exoskeleton.
complete the additional 1.5 h of walking that would have been required In the interests of clarity, we did not include the results of all previous
without experiencing fatigue, and so instead used the more typical exoskeleton experiments in Extended Data Fig. 5.
approach of single presentations with randomized order. We compared the results of this study against all types of lower-limb
One pilot participant completed additional indoor conditions, walk- exoskeleton, including devices that assist the knees and hips, to provide
ing at 1.25 m s−1 with an incline of 5°, walking at 1.25 m s−1 with a load of the clearest understanding of the relative benefits of this design and
20% of their body weight carried in a weight vest, and stair climbing personalization approach. Considering instead only ankle exoskeletons
on a stairmill at 50 steps per minute. The results (Extended Data Fig. 6) would allow for a more mechanistic comparison of system compo-
were used to test the generality of the approach. Owing to the small nents and biomechanics outcomes, at the cost of reduced generality
sample size (n = 1), this figure and the numerical results for change in of the high-level findings. As exoskeleton technologies mature and
metabolic rate are not included in the main text. address more tasks and populations, joint-specific benefits or restric-
We performed a naturalistic overground experiment in an outdoor, tions related to specific conditions may make it more sensible to apply
suburban community setting. People require assistance in many differ- joint-specific comparisons in some contexts.
ent settings and for a variety of additional activities, and future work Our untethered exoskeleton provided the largest reductions in the
should extend the approaches presented in this study to optimize metabolic cost of walking primarily owing to the way it personalized
assistance and evaluate assistive device benefits for a wider range of assistance to individual users, but hardware design differences may
tasks. For example, future devices could sense, adapt to and optimize also have contributed to its efficacy. Design differences between the
assistance for various grades55, during stair navigation17 and over rough untethered exoskeleton and some previous devices include: directly
terrain54. These future studies will provide additional translational measuring joint torque, rather than inferring it from motor current;
impact for daily mobility. providing slack in the transmission to avoid interference during leg
swing and Zero Torque mode; and larger peak torque capabilities,
Comparison with other untethered exoskeletons such that benefits were limited more by the user’s ability to accept
We compared the benefits of Real-world Optimized assistance with assistance than by limitations in the hardware. Directly measuring
the untethered exoskeleton to the best results of comparable previous joint torque requires additional electronics hardware for sensing and
studies10–16. To allow direct comparison, we considered only studies signal processing but enables more precise control of applied torques,
that tested untethered devices, report data for normal walking, tested which eliminates errors owing to model mismatch and power losses
similar walking conditions, tested sufficient participants and used in the transmission and interface with the body. This helps provide
standard data-processing techniques. For untethered exoskeletons, users with a consistent assistance pattern. Placing slack in the trans-
the most relevant outcome is the percent change in the energy cost mission cable during periods when zero torque is desired prevents
of walking with exoskeleton assistance to walking in normal shoes the inadvertent application of the small damping torques needed for
without the exoskeleton. Changes in walking conditions can affect linear feedback control. Although they may seem small, these damping
Article
torques can substantially increase user effort. Allowing for larger peak 46. Blokland, I. J. et al. Estimation of metabolic energy expenditure during short walking
bouts. Int. J. Sports Med. 42, 1098–1104 (2021).
torques, in this case approximately twice the value of previous unteth- 47. Witte, K. A., Fiers, P., Sheets-Singer, A. L. & Collins, S. H. Improving the energy economy
ered ankle exoskeletons10,17,64,66, allowed for a larger range of possible of human running with powered and unpowered ankle exoskeleton assistance. Sci. Robot.
assistance parameters. This makes it more likely that the global opti- 5, eaay9108 (2020).
48. Snaterse, M., Ton, R., Kuo, A. D. & Donelan, J. M. Distinct fast and slow processes
mum for a given participant and walking speed lie within the range of contribute to the selection of preferred step frequency during human walking. J. Appl.
hardware-feasible control. Larger torques require a rigid frame to react Physiol. 110, 1682–1690 (2011).
out transmission forces in the exoskeleton joint43, rather than through 49. Jackson, R., Dembia, C. L., Delp, S. L. & Collins, S. H. Muscle-tendon mechanics explain
unexpected effects of exoskeleton assistance on metabolic rate during walking. J. Exp.
shear on the skin63, to maintain user comfort. The present results would Biol. 220, 2082–2095 (2017).
therefore seem to favour devices that can apply higher torques to 50. Nuckols, R. W. et al. Individualization of exosuit assistance based on measured muscle
achieve greater benefits from assistance, at the cost of greater worn dynamics during versatile walking. Sci. Robot. 6, eabj1362 (2021).
51. Galle, S., Malcolm, P., Collins, S. H. & De Clercq, D. Reducing the metabolic cost of
mass. This relationship, however, will be sensitive to the populations walking with an ankle exoskeleton: interaction between actuation timing and power.
and tasks that are assisted. The above design decisions enabled the J. Neuroeng. Rehabil. 14, 35 (2017).
untethered ankle exoskeleton in this study to provide accurate, reli- 52. Moltedo, M. et al. Walking with a powered ankle-foot orthosis: the effects of actuation
timing and stiffness level on healthy users. J. Neuroeng. Rehabil. 17, 98 (2020).
able and substantial assistance to the user, which enabled participants 53. Sun, D., Fekete, G., Mei, Q. & Gu, Y. The effect of walking speed on the foot inter-segment
to obtain large net benefits from real-world personalized assistance. kinematics, ground reaction forces and lower limb joint moments. PeerJ 6, p.e5517
(2018).
54. Zihajehzadeh, S. & Park, E. J. Experimental evaluation of regression model-based walking
Participant surveys on exoskeleton usability
speed estimation using lower body-mounted IMU. In Conference of the IEEE Engineering
Participants completed a series of surveys to evaluate the ease of use, in Medicine and Biology Society (EMBC) 243–246 (IEEE, 2016).
comfort and functionality of the untethered exoskeleton after comple- 55. Quintero, D., Lambert, D. J., Villarreal, D. J. & Gregg, R. D. Real-time continuous gait phase
and speed estimation from a single sensor. In IEEE Conference on Control Technology
tion of all the experiments. Participants completed a System Usability and Applications (CCTA) 847–852 (IEEE, 2017).
Scale survey67 to determine how easy it was to operate the untethered 56. Martinez, A., Lawson, B. & Goldfarb, M. A controller for guiding leg movement during
exoskeleton. Users reported that the exoskeleton was relatively easy to overground walking with a lower limb exoskeleton. IEEE Trans. Robot. 34, 183–193 (2017).
57. Ham, R. V., Sugar, T., Vanderborght, B., Hollander, K. & Lefeber, D. Compliant actuator
use, with an overall score of 72.5 (Extended Data Table 3), placing it in designs. IEEE Robot. Autom. Mag. 3, 81–94 (2009).
the 65th percentile of 5,000 devices previously surveyed42. Participants 58. Vanderborght, B. et al. Variable impedance actuators: a review. Robot. Autom. Syst. 61,
also completed surveys adapted from the Orthotics and Prosthetics 1601–1614 (2013).
59. Dollar, A. M. & Herr, H. Lower extremity exoskeletons and active orthoses: challenges and
Users’ Survey68, which acts as a self-report instrument for evaluating state-of-the-art. IEEE Trans. Robot. 24, 144–158 (2008).
the outcomes of prosthetics and orthotics services in a clinically useful 60. Novak, D. & Riener, R. A survey of sensor fusion methods in wearable robotics. Robot.
manner. Among comfort-related outcomes, participants were most likely Autom. Syst. 73, 155–170 (2015).
61. Lenzi, T., De Rossi, S. M., Vitiello, N. & Carrozza, M. C. Intention-based EMG control for
to agree that the weight of the device was manageable, that it was easy to powered exoskeletons. IEEE Trans. Biomed. Eng. 59, 2180–2190 (2012).
put on and that their clothes were free of wear (Extended Data Table 4). 62. Stirling, L. et al. Static, dynamic, and cognitive fit of exosystems for the human operator.
Participants were more likely to be neutral or to disagree that the exo- Hum. Factors 62, 424–440 (2020).
63. Yandell, M. B., Quinlivan, B. T., Popov, D., Walsh, C. & Zelik, K. E. Physical interface
skeleton would be comfortable throughout the day. Among outcomes dynamics alter how robotic exosuits augment human movement: implications for
related to functionality, participants found standing, walking indoors optimizing wearable assistive devices. J. Neuroeng. Rehabil. 14, 40 (2017).
and outdoors, and donning and doffing the exoskeleton to be easy or 64. 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,
very easy (Extended Data Table 5). Participants found picking objects up 3571–3577 (2022).
from the ground and walking up steep ramps to be slightly difficult. When 65. Lim, B. et al. Delayed output feedback control for gait assistance with a robotic hip
asked whether they would prefer to use the exoskeleton or normal shoes exoskeleton. IEEE Trans. Roboy. 35, 1055–1062 (2019).
66. Mooney, L. M., Rouse, E. J. & Herr, H. M. Autonomous exoskeleton reduces metabolic cost
if they had to complete the outdoor walking course again, six out of the of human walking during load carriage. J. Neuroeng. Rehabil. 11, 151 (2014).
ten participants reported that they would prefer to use the exoskeleton. 67. Brooke, J. in Usability Evaluation in Industry (eds Jordan, P. W.) 189–194 (Taylor & Francis,
1996).
68. Heinemann, A. W., Bode, R. K. & O’Reilly, C. Development and measurement properties of
Reporting summary the Orthotics and Prosthetics Users’ Survey (OPUS): a comprehensive set of clinical
Further information on research design is available in the Nature outcome instruments. Prosthet. Orthot. Int. 27, 191–206 (2003).
Research Reporting Summary linked to this article.
Acknowledgements We thank G. R. Tan, D. Miller and V. Chiu for assistance with exoskeleton
hardware; K. Poggensee for assistance with accessing previous experimental data; W. Peisch
Data availability for assistance with mechanical design; and W. Gu, O. Chaudhuri, A. Okamura, B. Roth, S. Slade
and L. Lau for editorial suggestions. Funding: National Science Foundation Graduate Research
All study data necessary to replicate this work are available in the Source Fellowship DGE-1656518 (P.S.), Stanford Graduate Fellowship (P.S.), Wu Tsai Human
Data included with the paper. Computer-aided design files and a bill Performance Alliance Postdoctoral Fellowship (P.S.), NIH Grant P41EB027060 (S.L.D.) and
National Science Foundation Grant No. CMMI-1734449 (S.H.C.).
of materials for the untethered ankle exoskeleton are provided in Sup-
plementary Data 2. Source data are provided with this paper. Author contributions P.S. contributed to conceptualization, methodology, investigation,
visualization and writing. M.J.K. contributed to conceptualization and writing. S.L.D.
contributed to conceptualization and writing. S.H.C. contributed to conceptualization,
methodology, visualization, and writing.
Code availability
Optimization code samples are provided in Supplementary Data 1. Competing interests S.H.C. is an inventor on a patent application (US patent no. 10,537,283)
that covers the emulator systems used in this study.
This code uses Python version 3.6.1. The required python packages are
numpy (1.17.4), scikit-learn (0.21.3), scipy (1.3.2) and matplotlib (2.0.2). Additional information
Supplementary information The online version contains supplementary material available at
44. Brockway, J. M. Derivation of formulae used to calculate energy expenditure in man. https://doi.org/10.1038/s41586-022-05191-1.
Hum. Nutr. Clin. Nutr. 41, 463–471 (1987). Correspondence and requests for materials should be addressed to Steven H. Collins.
45. Seethapathi, N. & Srinivasan, M. The metabolic cost of changing walking speeds is Peer review information Nature thanks Carlos Rodriguez Guerrero, Zachary Lerner and the
significant, implies lower optimal speeds for shorter distances, and increases daily other, anonymous, reviewer(s) for their contribution to the peer review of this work.
energy estimates. Biol. Lett. 11, 20150486 (2015). Reprints and permissions information is available at http://www.nature.com/reprints.
Extended Data Fig. 1 | Weights, inputs, and effects for the data-driven differences in inputs. To generate this average, we ordered each pair-wise
classification model. The data-driven classifier decoded latent information comparison by metabolic rate, such that inputs from the control law with a
from human movement that was not otherwise interpretable, allowing higher metabolic rate were always subtracted from those with a lower
exoskeleton assistance to be optimized without laboratory-based metabolic rate. Typical values of the model inputs differ, in part because of
measurement equipment. Top row: Model weights and mean inputs. The model differences in units, and so the magnitudes of model weights do not
compares data from two control laws at a time and associates inputs with correspond well to the contributions of those terms to the classification
higher or lower metabolic rate to estimate which control law resulted in a lower overall. Bottom row: The classification contributions of each term in the model,
metabolic rate. Inputs comprised differences in ankle angle and ankle angular averaged over the entire training set. The percent contribution is calculated as
velocity at 30 different points in the gait cycle and differences in the four the absolute value of the product of the model weight and the input difference,
control law parameters of peak torque magnitude, peak time, rise time, and fall summed over all pair-wise comparisons, divided by the sum over all model
time. The data-driven model weights that multiply these differences are shown terms. For the x-axes, 0% and 100% of the gait cycle refer to the instant of heel
as a background colour of blue or red. Blue indicates that a positive difference strike of the assisted limb at the beginning and end of one stride. Toe-off occurs
is associated with lower metabolic rate, while red indicates that a positive at about 62% of the gait cycle. For a discussion of the intuitive meaning of the
difference is associated with higher metabolic rate. Darker colours indicate weights and contributions, please see the Methods subsection “Data-driven
greater influence. Black lines depict the average, across all training data, of the optimization”.
Article

Extended Data Fig. 2 | Speed-adaptive control approach. a, To calibrate the to estimate walking speed based on measurements of stride duration alone.
walking speed estimator, data are collected while the participant walks at c, The speed-adaptive controller relates estimated walking speed to
several prescribed speeds, each within the range of speeds associated with a exoskeleton assistance parameters by interpolating between assistance
set of assistance parameters to be optimized. The measured stride durations parameters specified at a set of chosen speeds. In this case, there are three sets
and ground-truth speed measurements from those tests are used to fit an of optimized parameters corresponding to three different walking speeds.
affine equation with linear regression. b, The resulting model can then be used
Extended Data Fig. 3 | Diagram of untethered exoskeleton electromechanical passing through a plastic tube embedded in the sole of the shoe. A Vectran
hardware. These computer-aided design drawings depict the hardware transmission cable (not pictured) transmits force from the drum to the tip
elements of the untethered exoskeleton. The primary components are of the heel spur. At the top of the calf strut, Velcro straps (not pictured) are
labelled. An image of the entire device, including textile components, can be connected to the strut through slots. These straps adhere to a separate
found in Fig. 4b. A running shoe (not pictured) is attached to the toe strut with Velcro strap (not pictured) worn on the shank of the leg, just below the knee.
pins that extend from the tip of the toe strut into a carbon fiber plate embedded A complete bill of materials and set of computer-aided design files for this
in the sole of the shoe. The heel of the running shoe is attached to the heel spur untethered exoskeleton assembly is included as Supplementary Data 2.
by a rope (not pictured) tied into holes on either side of the heel spur and
Article

Extended Data Fig. 4 | Optimizing assistance during real-world walking. the user. After six control laws had been applied for a given walking speed bin,
The exoskeleton applied speed-adaptive control, which adjusted exoskeleton forming one generation for the optimizer, the stored data were used to update
assistance parameters on each step. Stride duration (t stride) was used to estimate the optimization parameters associated with that speed bin. When any bin
walking speed (v) as described in Fig. 3. While the participant walked, portable performed an update, the estimate of the optimal parameter values (μ) for the
̇
sensor data (d) were collected, which included ankle angle (θ), ankle velocity (θ ), other bins were also updated. Bins that were closer to convergence, indicated
and the control law defining exoskeleton assistance torque (C). If sufficient by a small value of the convergence parameter (σ) for that bin, were adjusted
continuous strides (z) were not collected before the bout finished, the data less. This approach allowed the optimizer to rapidly adapt to the participant
were discarded and evaluation of the same control law began anew on the next early in the optimization, then to fine-tune the speed-specific parameters as
walking bout. If sufficient continuous strides were collected, then data were the optimization progressed. Following the update, the optimizer selected a
stored for the associated control law number (n) and walking speed bin (b), promising set of new control laws to be sequentially evaluated in the next
selected based on the average walking speed for the collected strides. The generation for the associated walking speed bin.
control law number was incremented and the next control law was applied to
Extended Data Fig. 5 | Exoskeleton comparison for standardized walking comparison to the largest prior improvements in metabolic rate. For legibility,
conditions. We compared the benefits of Real-world Optimized assistance in this figure we depict only results within a 5% reduction in net metabolic cost
from the untethered exoskeleton under standardized laboratory conditions to of the best prior results for each category. Please see the Methods subsection
those of prior untethered exoskeletons10–16. We considered only the results of “Comparison to other untethered exoskeletons” for a complete explanation of
tests that: compared exoskeleton-assisted outcomes to walking in normal the methods used to select amongst prior exoskeleton experiments.
shoes without an exoskeleton; used standard indirect respirometry procedures; Real-world Optimized assistance from the untethered exoskeleton resulted in
had sufficient sample sizes; and applied walking conditions within 10% of the large improvements in energy cost.
chosen walking speeds and inclines in this study, which were chosen to allow
Article

Extended Data Fig. 6 | Additional untethered exoskeleton treadmill

condition evaluations. For one pilot participant (n = 1), walking with
Real-world Optimized assistance reduced the metabolic cost of walking
compared to Normal Shoes during several additional treadmill conditions.
These results suggest that Real-world Optimized assistance may perform well
during a wide range of common walking activities. The conditions were walking
at 1.25 m s−1 on a 5° incline, walking at 1.5 m s−1 while wearing a vest weighing
approximately 20% body weight, and climbing stairs at a rate of 50 steps per
minute. These results are not included in the main text due to their preliminary
nature compared to the primary study outcomes.
Extended Data Table 1 | Randomized condition ordering for the outdoor evaluations in the real-world exoskeleton
optimization experiment

Subject Condition 1 Condition 2 Condition 3

1 Real-world Optimized Generic Speed-adaptive Normal Shoes

2 Generic Speed-adaptive Normal Shoes Real-world Optimized

3 Normal Shoes Generic Speed-adaptive Real-world Optimized

4 Real-world Optimized Normal Shoes Generic Speed-adaptive

5 Normal Shoes Real-world Optimized Generic Speed-adaptive

6 Generic Speed-adaptive Real-world Optimized Normal Shoes

7 Real-world Optimized Generic Speed-adaptive Normal Shoes

8 Normal Shoes Real-world Optimized Generic Speed-adaptive

9 Normal Shoes Real-world Optimized Generic Speed-adaptive

10 Generic Speed-adaptive Normal Shoes Real-world Optimized

Participants completed a quiet standing condition followed by Normal Shoes, Generic Speed-adaptive exoskeleton assistance, and Real-world Optimized exoskeleton assistance conditions,
presented in random order.
Article
Extended Data Table 2 | Randomized condition ordering for the indoor evaluations in the real-world exoskeleton
optimization experiment

Subject Condition 1 Condition 2 Condition 3 Condition 4 Condition 5 Condition 6

1 Incl. Exo 1.25 Exo 1.5 Exo 1.5 NW 1.25 NW Incl. NW

2 1.25 NW 1.5 Exo Incl. Exo 1.25 Exo Incl. NW 1.5 NW

3 1.25 Exo Incl. Exo 1.5 Exo 1.25 NW 1.5 NW Incl. NW

4 1.5 NW Incl. NW 1.25 NW Incl. Exo 1.25 Exo 1.5 Exo

5 Incl. Exo 1.25 Exo 1.5 Exo 1.25 NW Incl. NW 1.5 NW

6 1.5 NW Incl. NW 1.25 NW Incl. Exo 1.5 Exo 1.25 Exo

7 1.5 Exo Incl. Exo 1.25 Exo 1.5 NW 1.25 NW Incl. NW

8 Incl. NW 1.5 NW Incl. Exo 1.5 Exo 1.25 Exo 1.25 NW

9 1.25 Exo 1.5 Exo Incl. NW 1.25 NW 1.5 NW Incl. Exo

10 1.5 NW Incl. NW 1.25 NW 1.5 Exo Incl. Exo 1.25 Exo

The labels ‘1.25’, ‘1.5’, and ‘Incl.’ represent the treadmill conditions of walking at 1.25 m s−1, walking at 1.5 m s−1, and walking on a 10° incline at 1.25 m s−1. The labels ‘Exo’ and ‘NW’ represent the
Real-world Optimized assistance and Normal Shoes conditions. Conditions were presented in random order, with a constraint that the exoskeleton would only be donned and doffed once to
reduce experiment time.
Extended Data Table 3 | Usability survey results for exoskeleton participants

Question text (4 = Strongly Agree, 3 = Somewhat Agree, Untethered

2 = Neither Agree nor Disagree, 1 = Somewhat Disagree, 0 = Strongly Disagree) Exoskeleton
(Mean ± SD)

I think that I would like to use this system frequently. 2.8 ± 1.2

I found the system unnecessarily complex. 0.5 ± 0.5

I thought the system was easy to use. 3.2 ± 0.6

I think that I would need the support of a technical person to be able to use this system. 1.4 ± 1.3

I found the various functions in this system were well integrated. 3.0 ± 0.5

I thought there was too much inconsistency in this system. 1.9 ± 0.7

I would imagine that most people would learn to use this system very quickly. 2.6 ± 1.0

I found the system very cumbersome to use. 1.1 ± 1.0

I felt very confident using the system. 2.9 ± 0.9

I needed to learn a lot of things before I could get going with this system. 0.6 ± 0.7

Total usability score (out of 100) 72.5 ± 14.5

The System Usability Scale67, which uses a Likert scale, was used to evaluate the usability of the untethered exoskeleton. Participants (n = 10) completed this survey after completing all walking
experiments. The untethered exoskeleton was in the 65th percentile of a distribution of 5000 devices previously evaluated with the System Usability Scale42.
Article
Extended Data Table 4 | Survey results on the comfort of the untethered exoskeleton

Question text (4 = Strongly Agree, 3 = Agree, 2 = Neutral, Untethered Exoskeleton

1 = Disagree, 0 = Strongly Disagree) (Mean ± SD)

My skin is free of abrasions and irritations. 2.3 ± 1.4

My exoskeleton is pain free to wear. 2.5 ± 1.4

My exoskeleton is comfortable throughout the day. 1.5 ± 1.1

My exoskeleton looks good. 2.3 ± 0.9

My clothes are free of wear and tear from my exoskeleton. 3.8 ± 0.4

It is easy to put on my exoskeleton. 2.9 ± 0.6

My exoskeleton fits well. 2.8 ± 0.6

My exoskeleton is durable. 2.1 ± 1.0

The weight of my exoskeleton is manageable. 2.9 ± 1.0

Average Score 23.1 ± 4.6

This survey was adapted from the Orthotics and Prosthetics Users’ Survey68, which acts as a self-report instrument for evaluating clinically useful outcomes of prosthetics and orthotics services.
Participants (n = 10) completed the survey after completing all walking experiments.
Extended Data Table 5 | Survey results on the functionality of the untethered exoskeleton

Question text (4 = Very Easy, 3 = Easy, 2 = Difficult, Untethered Exoskeleton

1 = Very Difficult, 0 = Cannot Perform Activity) (Mean ± SD)

Walk up to two hours. 3.3 ± 0.8

Walk up a steep ramp. 2.3 ± 0.8

Stand one-half hour. 3.6 ± 0.7

Walk out-doors on uneven ground. 2.7 ± 0.5

Pick up an object from the floor while standing. 2.4 ± 0.7

Balance while standing. 3.8 ± 0.4

Put on and take off exoskeleton. 3.0 ± 0.7

Walk indoors. 3.5 ± 0.5

Average Score 24.6 ± 5.1

This survey was adapted from the Orthotics and Prosthetics Users’ Survey68, which acts as a self-report instrument for evaluating the outcomes of prosthetics and orthotics services in a
clinically useful manner. Participants (n = 10) completed this survey after completing all walking experiments.
β ɑ