FIELD OF THE INVENTION

This invention relates to methods of personalizing the performance of exoskeletons.

BACKGROUND OF THE INVENTION

Exoskeletons that assist leg movement show promise for enhancing personal mobility, but have yet to provide real-world benefits. Millions of people have mobility impairments that make walking slower and more fatiguing, while millions more have occupations that require strenuous locomotion. In research laboratories, exoskeletons can increase walking speed and reduce the energy required to walk, but these benefits have not translated to real-world conditions.

Providing beneficial assistance in the real world is difficult for several reasons: the specialized equipment used to personalize assistance is not available outside the laboratory; unlike walking on a treadmill, everyday walking occurs in many bouts of varying speed and duration; and devices must be self-contained and easy to use. In this invention, the inventors addressed each of these challenges to demonstrate effective exoskeleton assistance under natural conditions.

SUMMARY OF THE INVENTION

A wearable method for personalizing wearable exoskeletons that may substantially improve mobility in patients. Historically, patients have selected wearable exoskeletons via trial-and- error, and current methods for customizing exoskeletons are costly and time-consuming relative to the presented invention. The approach uses a machine learning model and wearable sensors to evaluate the energy costs of exoskeleton assistance conditions involving whole-body coordination. The inventors demonstrate in an exemplary embodiment that for an ankle exoskeleton assistance, a data-driven optimized approach reduced energy consumption in walking across different speeds and inclines in a statistical comparison with existing methods. The presented method used to uncover the relationship between kinematics and energy costs may additionally extend to other assistive devices.

The elderly face significant mobility challenges and decreased physical activity results in rapid decline in physical health. Assistive devices like exoskeletons have shown remarkable promise for maintaining physical health and improving mobility. Optimizing exoskeleton assistance offers remarkable improvements to mobility compared to hand tuned assistance profiles. Existing methods for optimizing assistance require hours of walking and equipment that is infeasible for everyday use. Here the inventors present a wearable method for optimizing ankle exoskeleton assistance to minimize human energy cost during walking. Wearable exoskeletons capable of personalizing assistance may substantially improve mobility for populations like older adults.

A method is provided for personalizing a performance of an exoskeleton. An exoskeleton has an actuator capable of applying torque around a joint. A movement is performed that includes at least two movement cycles. The movement includes motion of the joint. Joint angle, joint velocity and control parameters for a plurality of data points or bins are collected during two of the at least two movement cycles. A trained model has been developed which is a mapping joint angle, joint velocity and control parameters for data points for the movement cycles as input and a single performance value as output. Collected the joint angles, the joint velocities and the control parameters are inputted into the trained model, and obtaining a single performance value for the performed movement. The single performance value is then used to adjust torque actuator control parameters to control the torque actuator of the exoskeleton.

In one embodiment, the trained model is trained at at-least one speed or speed range for the movement cycle.

In one embodiment, the movement cycle is a gait cycle.

In one embodiment, the joint is an ankle joint.

In one embodiment, the method is personalized in a real-world environment.

In one embodiment, the method is personalized during walking in a real-world environment.

In one embodiment, the movement cycles are performed at different speeds.

BRIEF DESCRIPTION OF THE DRAWINGS

The drawings are presented in grey scall converted from color. If desired the reader is referred to the priority document for color interpretation of the drawings/plots/graphs.

FIGs. 1-H show according to an exemplary embodiment of the invention data-driven exoskeleton optimization. We used data from laboratory tests to train a model that can perform optimization in real-time outside the laboratory. FIG. 1A, during optimization, the participant walks with the exoskeleton and experiences a sequence of assistance conditions, or control laws, each defining a pattern of exoskeleton torque. The optimizer's goal is to identify the torque pattern that maximizes performance. FIG. IB, ankle motions for each stride are recorded from sensors on the exoskeleton. FIG. 1C, all possible pairs of conditions are then compared. For each pair, differences in segmented motion data (A) are calculated by subtraction. FIG. ID, differences in motion are multiplied with classifier model weights (W), using a dot product operation, to obtain the pair coefficient ( i . FIG. IE, a logistic function uses the pair coefficient to compute the probability (pij) that the first assistance condition is more beneficial than the second condition. FIG. IF, the score for each condition (5) is computed by summing the probabilities of all pairs that include that condition. FIG. 1G, conditions are then ranked by score and used to update an optimizer. FIG. 1H, the optimizer selects a set of k new control laws to evaluate. This optimization process is repeated until convergence criteria are satisfied, in this case a set number of evaluations having been completed. During real-world experiments, optimization was performed on the exoskeleton's microcontroller.

FIGs. 2A-E show according to an exemplary embodiment of the invention data-driven optimization results. FIG. 2A, exoskeleton assistance was applied using a tethered ankle exoskeleton emulator. FIG. 2B, Assistance parameters optimized using the data-driven method converged to within 5% of the parameters identified using metabolic optimization, but in one quarter the time. FIG. 2C, individual subjects had unique data-driven optimized parameters, centered around the generic assistance parameters. FIG. 2D, data-driven and metabolic optimized parameters resulted in similar metabolic costs of walking, significantly lower than a zero torque, normal shoes, or generic assistance, when walking at 1.25 m/s (ANOVA, n = 10, *p < 2.7x l 0' ^x , **7? < 2.4X10’ ⁵ , ***7? < 0.047). The boxes extend from the lower to upper quartile values of the data, with a line at the median and a dot at the mean. The whiskers extend between the minimum and maximum of the data values. FIG. 2E, optimized torque patterns varied by walking condition, with similar changes in data-driven and metabolic optimized parameters. Data-driven and metabolic optimized assistance led to similar reductions in metabolic rate when walking at 0.75 m/s (slow), 1.25 m/s (normal), 1.75 m/s (fast), and on a 10 degrees incline.

FIGs. 3A-C show according to an exemplary embodiment of the invention speed-adaptive control. FIG. 3A, the speed-adaptive controller interpolated between previously optimized assistance parameters to estimate the optimal parameters for each step based on walking speed. FIG. 3B, ground truth and estimated walking speed for a representative participant. Speed was estimated on each step using a model that took stride period as an input (FIG. 1A). The shaded region represents the mean ± one standard deviation. FIG. 3C, when participants (n = 3) walked on a treadmill that varied speed sinusoidally between 0.75 and 1.75 m/s, speed- adaptive control reduced the metabolic cost of walking more than constant, generic assistance.

FIGs. 4A-D show according to an exemplary embodiment of the invention untethered ankle exoskeleton. FIG. 4A, a participant walking in a community setting wearing the exoskeleton. FIG. 4B, the exoskeleton consists of (1) a 0.3 kg battery pack worn on the waist, (2) a motor and drum transmission, (3) electronics to receive sensor data, command the motor, and perform optimization, (4) a carbon fiber frame to transmit forces to the body, and (5) a lightweight shoe. FIG. 4C, the motor can apply a peak torque of 54 Nm when walking at 1.5 m/s, sufficient to match optimized assistance parameters identified in emulator experiments. FIG. 4D, the motor temperature during 30 minutes of walking with maximum assistance reached approximately 35 degrees C, well below the 75 degrees C thermal limit of the motor. An exponential fit indicated the steady-state temperature was 35.4 degrees C.

FIGs. 5A-H show according to an exemplary embodiment of the invention real-world optimization of exoskeleton assistance. FIG. 5A, participant walking on the public validation course. FIG. 5A, map of the 566-meter course used for optimization and validation. FIG. 5C, distribution of self-selected walking speeds and FIG. 5D, walking bout durations during optimization and validation compared to previously recorded distributions of real-world walking data. FIG. 5E, as assistance was optimized over one hour of naturalistic bouts of walking, the convergence parameter (cr) continually improved. FIG. 5F, optimized parameters for each participant were unique. In the original drawing in the priority document red squares depict values for generic assistance. FIG. 5G, during validation under naturalistic walking conditions on the public course, real-world optimized assistance substantially reduced the energy cost of transport and increased walking speed compared to normal shoes (ANOVA, n = 10, *p < 0.039). FIG. 5H, real-world optimized assistance also substantially reduced the metabolic cost of walking compared to normal shoes during benchmark treadmill conditions (ANOVA, n = 10, *p < 0.023). The boxes extend from the lower to upper quartile values of the data, with a line at the median and a dot at the mean. The whiskers extend between the minimum and maximum of the data values.

FIG. 6 shows according to an exemplary embodiment of the invention model weights for the data-driven classification model. The differences in ankle angle and velocity averaged across all pairs of assistance conditions from the training data are shown in black as shown in the original drawing in the priority document. The model associates these differences in motion at each point in the gait cycle with a contribution to a lower or higher metabolic cost, shown as a background color of blue or red as shown in the original drawing in the priority document, respectively. Darker colors indicate greater importance.

FIGs. 7A-C show according to an exemplary embodiment of the invention speed-adaptive control. FIG. 7A, the participant calibrates the walking speed estimation by walking at several known speeds (using a treadmill or GPS). These measured stride durations with ground truth speed measurements are used to fit an affine equation with linear regression. FIG. 7B, These lines of best fit estimate walking speeds for new stride durations. FIG. 7C, the speed-adaptive controller relates these walking speed estimates to exoskeleton assistance parameters by interpolating between assistance parameters previously optimized at fixed speeds.

FIG. 8 shows according to an exemplary embodiment of the invention optimizing assistance during real-world bouts of walking. The exoskeleton assistance parameters were adjusted each step by using stride duration (t stride) to estimate walking speed and perform speed-adaptive control. Assistance parameters were optimized during real-world walking when the participant reached a sufficient number of strides (z) during a bout of walking. If data for the current set of exoskeleton control laws were collected then an optimizer updated its internal parameters and selected a promising set of new control laws, otherwise the condition number was incremented, and the next control law was applied to the user. We simultaneously optimized parameters for different bins of walking speeds (b), due to the variation in speed during natural walking. When any bin performed an update, the mean parameter values (p) of the other bins were also updated based on how much each bin had converged, represented by a change in the sigma which is the step size of the optimizer. This approach rapidly adapted to the participant early in the optimization and focused on updating parameters for common walking speeds as the optimization progresses.

FIG. 9 shows according to an exemplary embodiment of the invention additional untethered exoskeleton treadmill condition evaluations. Walking with data- driven optimized assistance reduced the metabolic cost of walking compared to normal shoes during several additional treadmill conditions, indicating it may perform well during a wide range of common walking activities. These conditions included walking at 1.25 m/s with a 5 degrees incline, walking at 1.5 m/s with a weight vest adding approximately 20% of the participant’s bodyweight, and climbing stairs at a rate of 50 steps per minute.

FIG. 10 shows according to an exemplary embodiment of the invention the method of controlling torque parameters of an exoskeleton using a trained model outputting a single performance value. DETAILED DESCRIPTION

Maximizing the benefits of exoskeleton assistance requires personalization to individual needs, which is challenging outside of a laboratory. The largest improvements in human walking performance have been achieved by individualizing assistance using human-in-the-loop optimization, a process in which device control is systematically tuned to improve human performance while a person uses a device. Measuring important aspects of performance, including metabolic rate, has required expensive laboratory equipment and long periods of steady treadmill walking. Individualizing consumer or medical devices in this way would require several long visits to a specialized clinic, which would be costly and impractical. If human performance could instead be estimated quickly, using low-cost wearable sensors, optimization could be performed as people moved naturally through their daily lives. This might be possible using musculoskeletal modeling, but such simulations are computationally intensive and require individualization. Data-driven models may be able to capture important features of human performance without complex models.

In this invention a data-driven model was developed that relates human motion during exoskeleton-assisted walking to metabolic energy consumption that can be used outside the laboratory. Human movement arises from the interaction between the inertia of our body segments and forces from the environment and our muscles. The inventors hypothesized that careful analysis could extract meaningful information about muscular energy expenditure from subtle changes in motion data. In an experiment, participants walked with exoskeleton assistance in about 3,600 different conditions while data were recorded from both laboratory equipment that measure biomechanical outcomes and low-cost, portable sensors on the exoskeleton. The inventors trained a logistic regression model using this dataset (Extended Data FIGs. 1A-H). The data-driven classification model compared sensor data from two assistance conditions and classified which assistance condition provided a larger benefit. The model inputs were ankle angle and ankle velocity, segmented by gait cycle, and the torque parameters for each assistance condition. The model then estimated the likelihood that the first assistance condition resulted in lower metabolic energy expenditure. During optimization, the user experienced a set of assistance conditions, the data-driven model compared all possible pairs of conditions, the conditions were ranked, and an optimization algorithm (CMA-ES) updated the estimate of the optimal parameters and generated a new set of assistance conditions to evaluate (FIGs. 1A-H). This process was repeated until convergence criteria were met.

Data-driven optimization can use the information embedded in our movements to identify exoskeleton assistance patterns that are as effective as those found with laboratory-based methods, but in one quarter of the time. We conducted experiments to optimize assistance with a tethered exoskeleton emulator (FIG. 2A). The data-driven optimization evaluated eight sets of assistance conditions in 32 min, four times faster than the state-of-the-art approach using indirect respirometry to measure metabolic rate (FIG. 2B). Data-driven and metabolic optimization approaches identified the same subject-specific adjustments to assistance (FIG. 2C). Assistance optimized using data-driven and metabolic approaches resulted in similar metabolic cost, significantly lower than the metabolic cost of walking with the exoskeleton in a zero-torque mode or with a generic assistance profile (FIG. 2D). The average of the data-driven optimized parameters matched those of the generic condition, taken from a best prior study, yet provided a larger benefit, demonstrating the importance of individualization; even subtle changes in torque can lead to substantial performance enhancements. To test the generality of the data-driven model, we conducted experiments at a range of additional speeds and inclines with a subset of participants. The data-driven and metabolic optimized assistance resulted in similar torque profiles and metabolic cost reductions across these conditions (Fig. 2E). This shows that the data-driven classification model captured a fundamental relationship between ankle movement and whole-body walking effort. This result demonstrates that human movement encodes information related to underlying physiological processes, and that data- driven methods can extract this information without laboratory equipment or complex multiscale models.

We developed a speed-adaptive controller to adjust assistance based on natural variations in walking speed. People vary their walking speed widely during the day in response to changes in context and constraints. Variations in speed complicate exoskeleton control and may help explain why devices that reduce walking effort on a treadmill have not yet provided benefits during real-world use. The speed-adaptive controller we developed (Extended Data FIGs. 2A- E) interpolated between assistance parameter values previously optimized at different walking speeds (FIG. 3A) based on the estimated speed of each step (FIG. 3B). We tested speed- adaptive assistance on a subset of participants (n = 3) as they walked on a treadmill with sinusoidally varying speeds. Speed-adaptive control reduced the energetic cost of walking more than generic assistance with constant parameters (FIG. 3C). Adjusting exoskeleton assistance based on speed is an effective strategy for handling speed variations that occur during normal walking.

We created an untethered exoskeleton for real-world assistance using a design approach based on emulation and optimization. Wearable robotic devices are typically designed using models or intuition, built as specialized prototypes, and then tested. However, humans are highly complex and diverse, making it difficult to predict the range of characteristics that will be optimal across a population. As a result, most devices designed this way are unable to provide optimal assistance and often provide no benefit at all. To develop the untethered exoskeleton used in this invention, we first performed experiments with versatile exoskeleton emulators. These laboratory-based, tethered hardware systems allowed us to perform a wide range of control and optimization experiments (FIGs. 1A-H, FIGs. 2A-E and FIGs. 3A-C) and identify the electromechanical characteristics that our untethered device would need. Using these design guidelines, we built a specialized, untethered device that provides predictable, meaningful benefits. This emulation and optimization design paradigm can reduce the cost and time required to develop new wearable robots.

Based on the results of our emulator experiments, we designed a specialized, untethered ankle exoskeleton. The system consisted of an exoskeleton worn on each ankle and a battery pack at the waist (FIG. 4A). The exoskeleton was designed to apply the range of optimal torque profiles identified in the tethered optimization study (FIGs. 2A-E) while having minimum mass (1.2 kg per ankle). A brushless motor and custom drum transmission applied torque about the ankle joint, while portable electronics sensed the user's motion and performed real-time control and optimization (FIG. 4B). The exoskeleton provided a peak torque of 54 Nm (FIG. 4C), which was about 50% to 75% of the biological ankle torque of participants in this invention. Torque was controlled using a mixture of classical feedback control and iterative learning, with a tracking error of less than 1% of the peak torque. Maximum assistance could be applied continually without overheating the motor (FIG. 4D). The battery weighed 0.3 kg and powered the exoskeleton for at least 30 minutes on a single charge. While the energy cost of carrying mass near a distal joint is high, locating motors and electronics near the assisted joint results in efficient power transmission, simpler design, and lower total weight, which can yield large net benefits.

We used the information encoded in single walking steps to optimize exoskeleton assistance while people walk in natural ways, with short bouts of varying speed. People take thousands of steps per day, but real-world walking occurs in many separate bouts, most of which are short, with 90% being less than one hundred steps in duration. Speed is relatively consistent within each bout, but varies across bouts. This fragmentation presents a challenge for collecting optimization data and efficiently fine-tuning assistance. Our data-driven optimization method addresses the problem of gathering useful data from short walking bouts by using kinematic data collected with every step. In pilot tests, we found that we could accurately compare conditions based on just 44 continuous steps, opportunistically captured during natural bouts, allowing our system to accumulate data from about 77% of steps on a typical day. We addressed variations in speed by defining speed bins based on observed human behavior, associating collected data with the appropriate bin, noting when sufficient data for any one speed bin had been accumulated, applying the data-driven classifier to rank assistance parameters, and using these rankings to update the optimal parameter estimates for all speed bins (Extended Data FIGs. 3A-C).

Real-world optimization quickly improved assistance during natural walking conditions. We conducted experiments in which participants performed one hour of short, exoskeleton-assisted walking bouts (FIG. 5A) on a public sidewalk (FIG. 5B). Participants were given audio cues based on ecologically relevant prompts that caused them to self-select walking speeds from a typical distribution (FIG. 5C). Cues were provided in random order and at specific intervals to form a typical distribution of bout durations (FIG. 5D). The optimizer steadily converged throughout the experiment (FIG. 5E), indicating improvements in exoskeleton control. Post-hoc analysis showed the optimizer did not reach steady state, suggesting that additional time could have provided a better estimate of the optimal parameters. Peak torques optimized during naturalistic walking were larger than those from treadmill-based experiments (FIG. 5F). Participants may have felt more stable during outdoor walking, allowing them to benefit from larger torques. These results show that robotic devices can collect useful optimization data under real-world constraints.

Real-world optimized assistance increased self-selected walking speed and reduced the metabolic energy expended per distance traveled during naturalistic walking. In a separate validation experiment, participants performed a fixed set of outdoor walking bouts with varying durations and speeds, while ground-truth metabolic rate and speed were measured. Energetic cost of transport was reduced by 17 ± 5% (ANOVA, n = 10, p = 0.039) and walking speed was increased by 9 ± 4% (ANOVA, n = 10, p = 0.031) compared to normal shoes (FIG. 5G). These energy savings are equivalent to removing a 9.2 kg backpack, while the increase in walking speed of 0.12 m/s is clinically meaningful. These results demonstrate that lower-limb exoskeletons can provide ecologically relevant benefits and provide benchmarks for assessing real-world performance of future devices. This shows that optimization can be conducted in a natural setting, seamlessly improving human-robot interaction over time.

Assistance optimized under real-world conditions produced even larger benefits under standard treadmill conditions. After performing optimization in a public setting, we tested our untethered exoskeleton during standardized laboratory walking conditions to directly compare to other devices. Optimized exoskeleton assistance reduced the energy cost of treadmill walking by 16% at 1.25 m/s, 23% at 1.5 m/s, and 18% when walking up a 10° incline (ANOVA, n = 10, p < 0.023), compared to normal shoes (FIG. 5H), approximately twice as effective as all existing devices (FIGs 4A-D). The energy savings during inclined walking were equivalent to removing a 15.2 kg backpack. Pilot results suggest that the device provides similar benefits under other conditions, including walking on a 5 degrees incline, loaded walking, and stair climbing (Extended Data FIGs. 5A-H). Emulator-informed hardware design coupled with opportunistic, data-driven optimization led to exceptional performance across walking conditions.

Participants reported that the untethered exoskeleton was easy to use and relatively comfortable. Wearable robotic devices should be easy to operate, comfortable, and functional for everyday activities in order to be adopted by users. Participants reported the exoskeleton was relatively easy to use (Table 1), ranking it in the 65 ^th percentile of previously surveyed consumer devices. Participants found that the exoskeleton did not interfere with their clothing and had a manageable weight, but were neutral as to whether it would be comfortable to wear throughout the day (Table 2). Participants reported that it was easy to put on and take off the exoskeleton, stand while wearing the exoskeleton, and walk indoors and outdoors for extended periods with the exoskeleton (Table 3). Six of ten participants reported that they preferred using the exoskeleton rather than normal shoes while walking on the public course. The device we tested was a research prototype and not a refined product. Our survey results suggest that it may be possible to create mobility-enhancing products that are easy to use, comfortable, and reliable, and that a substantial portion of the intended population may opt to use them.

These approaches to real-world optimization and device design could be applied to many assistive technologies. Real-world optimization could improve the effectiveness of robotic devices that assist people in diverse contexts, from workers with physically demanding jobs to people with mobility impairments. Assistance could aid a variety of tasks, such as stair climbing or lifting, and improve other aspects of performance, such as balance or joint pain. In each case, additional training data could be collected in the laboratory and used to train new data-driven optimizers, illuminating the information contained within the body’s movements for each task. With each training data set, the learned models could be made more general, progressively connecting to more fundamental relationships between movement and performance outcomes. Data from laboratory-based emulation and optimization experiments could simultaneously provide design guidelines for new products. When used regularly, we expect devices like this to become finely tuned to address the needs of each individual, resulting in larger performance enhancements than observed in this study. Longitudinal experiments will be needed to understand how such assistance affects behavior and quality of life; as moving becomes easier, we expect people will be more active, helping them to lead healthier lives.

Methods

Experimental design

The objective was to personalize exoskeleton assistance during real-world walking. To achieve this, a data-driven optimization was used, which uses portable sensors in the exoskeleton to personalize assistance for each participant. We hypothesized that data-driven optimized assistance would provide similar reductions in the metabolic cost of walking as metabolic optimized assistance, and significantly larger reductions than generic assistance or normal shoes. A power analysis that eight participants were the necessary sample size to validate the data- driven optimization. This analysis used a power of 0.8, alpha value of 0.05, and previous experimental results where metabolic optimized assistance (1.44 ± 0.15 W/kg) provided significantly larger metabolic reductions than generic assistance ⁹ (1.64 W/kg). We stopped data collection after reaching nine participants for the tethered exoskeleton experiments and ten participants for the untethered exoskeleton experiments, in case of sensor failures or a participant’s failure to complete all conditions. All participants had at least 8 hours of experience walking with assistance from powered ankle exoskeletons to minimize the training effects that occur while participants leam to walk efficiently with the exoskeleton. The experiments included human subject testing in both laboratory and outdoor settings. Participants wore bilateral ankle exoskeletons and walked under a series of assistance conditions in a randomized order. Each of the experiments is described in the following methods sections. A one-way analysis of variance (ANOVA) with a fixed effect for the assistance condition was used to determine whether one assistance condition reduced the metabolic cost of walking significantly more than another assistance condition.

Estimating the metabolic cost of walking

The metabolic cost of walking was computed with measurements from respirometry equipment. Respirometry equipment was used to measure the volume of carbon dioxide and oxygen exchanged per breath. The Brockway equation was used to compute metabolic energy expenditure in Watts from each breath of carbon dioxide and oxygen measurements taken by the respirometry equipment. Metabolics measurements during indoor exoskeleton experiments were measured with tethered respirometry equipment (Quark CPET, COSMED). Metabolics measurements during outdoor exoskeleton experiments were measured with portable respirometry equipment worn on the participant’s back (K5, COSMED). Participants refrained from all food and drink except for water for at least three hours before experiments that included respirometry measurements. The steady-state metabolic cost was computed by averaging the respirometry measurements during the last three minutes of the six-minute condition. The cumulative metabolic cost was the total energy expended during the condition, including the metabolic cost beyond quiet standing for three minutes following the condition. The energy spent during the return to quiet standing accounted for any deficits in oxygen during walking and delays in respirometry measurements. The cost of transport was calculated as the cumulative metabolic cost divided by the total distance walked.

Exoskeleton assistance conditions

A variety of exoskeleton assistance conditions were evaluated to determine the benefits that they provided to the user. These assistance conditions included walking in normal shoes and walking with the exoskeletons in a zero-torque mode, with a fixed generic assistance profile, with metabolic optimized parameters, and with data-driven optimized parameters.

Walking in normal shoes acted as a ground truth condition measuring the energy required to walk without the exoskeleton. Ideally, exoskeleton assistance would require a lower metabolic cost than walking with normal shoes, providing a benefit to the user.

Zero-torque mode was a tethered exoskeleton condition where the exoskeleton provided no assistive torques. The zero-torque condition evaluated the effort required to walk with the added weight of the exoskeleton and without the benefits of assistance. Comparing the walking effort during zero torque to a condition with assistance provided an estimate of the metabolic savings that an idealized assistive device could provide if it did not add any additional weight.

A generic assistance profile was a fixed set of assistance parameters. In experiments, generic assistance reduced the metabolic cost of walking, but was less beneficial than assistance personalized for each individual. The tethered exoskeleton experiments in this invention used a generic assistance condition computed by averaging the optimized parameters from a previous experiment. The untethered exoskeleton experiments used generic assistance parameters computed by averaging the optimized profiles for three speeds of walking from the tethered experiments in this invention (FIGs. 2A-E).

Metabolic optimization relied on metabolic measurements to fine-tune exoskeleton assistance parameters for each participant. Metabolic optimization uses a sample-efficient optimization approach to identifying the exoskeleton control parameters that minimize the metabolic cost of walking for a specific person. To perform metabolic optimization, a participant walked with exoskeleton assistance on a treadmill for two minutes while respirometry measurements were recorded. A minimum of two minutes of respirometry data were required to provide an estimate of the steady-state metabolic cost of walking. Once a fixed number of assistance conditions, referred to as one generation of optimization, were completed, an optimizer ranked the conditions in order of metabolic cost, updated the optimization parameters, and selected a new set of promising assistance conditions to evaluate. We used the same metabolic optimization approach from exoskeleton experiments which used Covariance Matrix Adaptation Evolutionary Strategy as the optimization framework.

Data-driven optimization used the exact same optimization framework as metabolic optimization, but relied on a data-driven model to individualize assistance parameters for each participant. The data-driven classification model used portable sensor data to estimate which assistance conditions were the most beneficial to the participant, allowing the assistance conditions to be ranked and used to update the optimization instead of metabolic estimates (FIGs. 1-H).

Data-driven Optimization

Data-driven optimization personalized assistance by using a data-driven classification model to determine which assistance conditions provided the largest benefits for each person. Data from untethered sensors in the exoskeleton were passed into a data-driven classification model. The data-driven classification model estimated the likelihood that one assistance condition provided a larger metabolic reduction than another assistance condition (FIGs. 1-H). The data-driven classification model was a logistic regression model.

The model inputs includes carefully processed portable sensor data including four torque parameters, as well as ankle angle and ankle velocity measurements. The torque parameters prescribe the ankle assistance profile and consist of four values: peak torque, peak timing, rise timing, and fall timing. The angle and velocity measurements were sampled from a rotary encoder in the ankle joint of the exoskeleton worn on the left leg. The portable sensor data was processed by segmenting the ankle angle and velocity measurements by gait cycle, whenever a heel strike was detected by the pressure sensor insoles in the exoskeletons. The first six gait cycles of data were discarded. The remaining gait cycles of data were discretized by averaging the measurements into 30 discrete bins and then averaged across all gait cycles for that condition. The processed data was reshaped into a single vector with 64 values, including the torque parameters, 30 binned values for the ankle angle across the gait cycle, and 30 binned values for the ankle velocity across the gait cycle. The model input also consisted of 64 values, the vector of sensor data from one assistance condition subtracted from the vector of sensor data from a different assistance condition. This difference in the sensor measurements provided the model with information about how the torque and person’s movements varied between the two conditions. Previous data-driven models accurately estimated energy expenditure from wearable sensor data, in part by formatting data by gait cycle.

The data-driven classification model was trained to compare two assistance conditions at a time, determining which condition provided a larger reduction in the metabolic cost of walking. Training the data-driven classifier input data from portable sensors and ground truth labels from metabolic measurements during many exoskeleton assistance conditions. The sensor data was taken as input into the model to estimate the likelihood that the first of the compared assistance conditions reduced the metabolic cost of walking more than the second condition. This probability was a continuous value from 0 to 1, with 1 indicating the highest likelihood that the first condition reduced the metabolic cost of walking more than the second condition. The ground truth labels were computed by subtracting the metabolic costs, estimated with two- minutes of respirometry data, for two conditions. A negative valued label indicated the first condition was more beneficial, reducing the metabolic cost of walking more than the second condition. A positive valued label indicated the second condition was more beneficial, reducing the metabolic cost of walking more than the first condition. The model was trained with data from previous metabolic optimization experiments, where ten participants walked under approximately 3600 exoskeleton assistance conditions. We also used regularization, a technique that encourages simpler models and avoids overfitting to training data in order to improve model estimates for new data points. The model fitting included a lasso regularization term, penalizing the absolute value of the model weights multiplied by a regularization parameter with a value of 1. The data-driven classification model was trained to capture a relationship between leg movement and the metabolic cost of walking with assistance. The data-driven model classified pairs of assistance conditions using a linear set of weights and the logistic function. These weights can be visualized, but are difficult to interpret (FIGs. 1-H). We believe that an assistance condition with larger, smooth ankle motion and later peak timing of torque are more likely to be classified as providing a larger metabolic reduction. The model weights indicate that increasing dorsiflexion at 40% of the gait cycle and plantarflexion near toe-off may help reduce the metabolic cost of walking. A smooth ankle motion, with reduced ankle velocity during stance, also appears to be associated with a lower metabolic cost of walking. The flight phase was not informative, contributing 0.2% to the total classification probability. The model weights for the torque parameter values indicated that higher torque magnitudes, later timing, and a longer duration of assistance were associated with lower metabolic cost. Previous experiments found people have varied preferences for peak torque and rise time, while peak and fall timing were consistent. This may explain why the largest model weight was associated with peak timing.

The data-driven classification model used in these experiments also took as inputs the subject’s height and weight, but further analysis revealed this information was not relevant to the model classifications. Models trained in exactly the same way, with or without the subject’s height or weight, classified 96% of the exoskeleton assistance pairs evaluated in the data-driven optimization experiments the same. Thus, a person’s height and weight were not informative when determining the effectiveness of assistance conditions. To perform optimization, a generation of conditions were ranked by using the probability values estimated by the data-driven classifier. Each pair of assistance conditions passed into the data- driven classification model resulted in one probability value of whether the first or second assistance condition provided a larger benefit. All possible pairs of conditions were classified with the data-driven model to receive a set of probability values. Each assistance condition was scored summing the probabilities from all pairs that included that assistance condition. If a condition was the second condition in the pair, the negative probability was added to the total score for that condition. The assistance conditions were ranked by the magnitude of scores, with a larger value indicating the condition was more likely to provide a larger reduction in the metabolic cost of walking (FIGs. 1-H). These ranked conditions replaced the need for metabolic measurements with respirometry, allowing the optimizer to update its internal parameters and generate a promising set of new assistance conditions to evaluate.

Tethered optimization experiments

Tethered exoskeleton experiments were performed in an indoor laboratory setting to determine how effective data-driven optimization was compared to a range of other assistance conditions. Participants wore tethered bilateral ankle exoskeleton emulators. The exoskeleton assistance was governed by a torque pattern parameterized by four parameters: peak magnitude, peak time, rise time, and fall time ⁷ . The control loop ran at 1000 Hz on a real-time computer (Speedgoat). Exoskeleton sensor measurements were recorded at a rate of 2000 Hz, including pressure values from shoe insoles, commanded torque parameters, applied torque, ankle angle, and ankle velocity. These measurements were used to estimate gait cycle percent using a time-based control method to perform torque tracking with a combination of classical controls and iterative learning, accounting for systematic errors between the desired and applied torque. Two tethered exoskeleton experiments were used to evaluate the effectiveness of a range of assistance conditions. The first experiment compared assistance conditions while participants walked at 1.25 m/s, a normal walking speed previously used for metabolic optimization experiments ^7,9 . Healthy young adults (n = 9, 5 men and 4 women; age = 24.8 ± 1.8 yr; body mass = 65.3 ± 8.0 kg; height = 1.73 ± 0.07 m) completed a two-day experimental protocol. The first day, participants performed experiments to personalize assistance parameters with metabolic optimization and data-driven optimization, in a randomized order. Participants completed eight generations of optimization for each approach. Each generation consisted of eight assistance conditions. The optimizations were initialized with the generic assistance parameters, corresponding to the average optimized parameters of a previous group of expert participants. The optimizations were initialized with a covariance size that included a 20% range of the normalized assistance parameters, corresponding to a sigma value of 0.1. The metabolic optimization conditions lasted two minutes, the minimum time needed to estimate the steadystate metabolic cost with respirometry, requiring a total time of 128 min of walking. The data- driven optimization conditions lasted only 30 seconds because people quickly converge to steady state motion, requiring a total time of 32 min of walking. For each participant, the parameters identified using data-driven and metabolic approaches were similar. For example, peak torque values were well correlated (R ² = 0.76, p = 1.4xl0' ⁴ , n = 9). The second day, participants performed a standing rest condition followed by assistance conditions including walking in normal shoes and walking with the exoskeletons under a zero-torque mode, generic assistance profile, metabolic optimized parameters, and data-driven optimized parameters. The assistance conditions for these validation tests were randomized and presented in a doublereversal ABCDDCBA order to mitigate the effects of noise in the metabolics measurements and trial order. Each condition lasted for six minutes and included metabolic measurements.

The second experiment was used to evaluate the same set of assistance conditions at additional speeds. Healthy young adults (n = 3, 3 men; age = 24.0 ± 2.0 yr; body mass = 66.0 ± 8.0 kg; height = 1.76 ± 0.05 m) completed the experiment. Participants completed the same experimental protocol used in the first tethered exoskeleton experiment that optimized and evaluated assistance conditions for 1.25 m/s, but performed three additional walking conditions of 0.75 m/s, 1.75 m/s, and a 10° incline at 1.25 m/s.

Speed-adaptive controller and validation experiment

We developed a speed-adaptive controller that adjusted exoskeleton assistance based on walking speed. During real-world walking, people naturally vary their speed. We hypothesized that adjusting exoskeleton assistance based on walking speed would provide larger metabolic reductions than a constant pattern of assistance. We estimated the walking speed of each step using a linear model, relating measured stride durations to measured walking (FIGs. 1-H). Walking speed estimates from each step were used to interpolate exoskeleton assistance between the parameters optimized at different fixed walking speeds (FIG. 4D). The expected stance duration was also adjusted with the speed estimates.

The third tethered exoskeleton experiment evaluated if adapting assistance to variations in walking speed could provide larger metabolic reductions than a fixed generic assistance profile.

Healthy young adults (w = 3, 3 men; age = 24.0 ± 2.0 yr; body mass = 66.0 ± 8.0 kg; height = 1.76 ± 0.05 m) completed the experiment. These participants had previously completed the first two tethered exoskeleton experiments, providing data-driven optimized parameters for walking speeds of 0.75, 1.25 and 1.75 m/s. Participants walked on a treadmill while the speed varied sinusoidally from 0.75 to 1.75 m/s with a period of 30 seconds. Participants completed assistance conditions including walking in normal shoes and walking with the exoskeletons under a zero-torque mode, generic assistance profile, and data-driven optimized parameters using the speed-adaptive controller to adjust assistance with each step. The validation tests were randomized and presented in a double-reversal ABCDDCBA order to mitigate the effects of noise in the metabolics measurements and trial order.

Untethered exoskeleton design

The untethered exoskeleton was designed to provide the optimized assistance parameters from the tethered exoskeleton experiments during extended real-world use. The optimized parameters required assistance with a peak torque of 54 Nm while walking 1.5 m/s. The exoskeleton was designed to provide this level of assistance without the motor overheating. A portable battery was selected to allow 30 minutes of continuous walking on a single charge. The device weight was minimized to reduce the metabolic effort required to carry the exoskeleton.

The untethered exoskeleton weighed 1.2 kg on each ankle and consisted of the same frame, shoe, and pressure sensor insole as the tethered exoskeleton, with the addition of a portable motor, drum transmission, electronics, and a battery. The brushless motor (AK80-9, CubeMars) contained a single stage 9: 1 gear ratio and internal motor driver electronics. The motor weighed approximately 0.5 kg and was selected to apply the optimized torques and ankle velocities from the tethered exoskeleton experiments by using a 5:1 drum transmission. The custom drum transmission was machined from 7075 aluminum. A cable connected the heelspur to the motor drum. When the motor applied torque to the drum, the cable transmitted this force to create torque about the ankle joint of the exoskeleton. The drum and cable transmission had added benefits of being backdrivable. The cable could also be driven to a slack state to allow the person to move freely when desired. The untethered exoskeleton used a Raspberry Pi 4b microcontroller to read sensor data and perform real-time control and optimization at a rate of 200 Hz. A breakout board enabled sensors to interface with the microcontroller. A step-down voltage converter enabled the electronics to be safely powered by a portable battery. The total weight of electronics was 0.15 kg. A lithium polymer battery had a nominal voltage of 24 V, capacity of 1300 mAh, and weight of 0.3 kg.

The design of the untethered exoskeleton required several trade-offs. The highest design priority was providing the 54 Nm peak torque, specified from previous optimization experiments, with the lightest motor and transmission. We considered several factors to ensure that the motor would provide 54 Nm during operation. We simulated the torque needed to provide desired assistance, overcome transmission inefficiencies, and accelerate the mass or the motor rotor and drum to the ankle velocity during walking at 1.5 m/s. The motor had to operate at a safe steady - state temperature, to prevent damage to the windings. A brushless motor was selected for its high efficiency and large peak torque capability. While this untethered exoskeleton was designed for the optimized parameters of our experimental participant group, other participants may require a different device with another torque-weight trade-off to provide the same metabolic reductions.

Another design trade-off was selecting whether to place the motor and electronics near the assisted joint or closer to the hips. While the energy cost of carrying mass at distal joints is high, locating motors and electronics near the assisted joint results in efficient power transmission, better control bandwidth, and less total weight. We selected a drum and cable transmission located on the shank of the leg. The motor wound the cable around the drum transmission, creating a force on the heel spur and resulting in a torque about the ankle. Mounting the motor and electronics at the hip would have required a cable and sheath to transmit forces to assist the ankle joint. This cable and sheath transmission introduce complex, non-linear friction, making device control more challenging, reducing control bandwidth, and decreasing efficiency. The cables and sheath and additional electrical wires also add additional weight to the system.

Opportunistic optimization approach

Opportunistic optimization overcame the challenges of optimizing assistance during short bouts of real-world walking by accumulating data across many bouts. The opportunistic optimization approach used the same data-driven classification model and optimization method that was validated in the tethered experiments, but only updated the optimization whenever 22 consecutive strides were collected for one control law (FIGs. 3A-C). A fixed number of 22 strides was selected to capture approximately the same number of strides as the 30 second condition duration used in the tethered data-driven optimization experiments. Once this sufficient number of strides were collected, the next assistance condition was applied to the user (FIGs. 3A-C). Similar to the tethered experiments, the first six strides of data were discarded due to the person changing speeds at the beginning of the bout. The exact same data-driven classification model evaluated in the tethered exoskeleton experiments was used for the real- world optimization experiments. The torque parameters for peak timing and fall timing were fixed to the average values of the data-driven optimized parameters from the first tethered exoskeleton experiment. The optimized values of peak timing and fall timing changed little with variations in speed, indicating a fixed value was likely sufficient. The peak torque and rise timing values were optimized, requiring only six assistance conditions to be collected in one generation of optimization.

Once data for all control laws of a generation were collected, the data-driven classification model ranked the control laws. This control law ranking updated the optimizer. Separate optimizations were performed for three bins of walking speed of less than 1.22 m/s, between 1.22 and 1.38 m/s, and greater than 1.38 m/s. These speeds were chosen from the 33rd and 66th percentile of real-world walking distributions, providing equal likelihood for the participant to walk in each bin. Speed-adaptive control interpolated assistance based on the speed of each individual step (FIGs. 2A-E). When a condition of data was collected the estimated walking speeds for all steps during that condition were averaged to select which speed bin to store the data for the optimization process. When a generation of conditions were collected for a speed bin the optimization parameters for that bin were updated. The current parameter mean value of the other speed bins were also updated. The mean parameter value was updated proportionally to the ratio of sigma, representing the convergence of the optimization (FIGs. 3A-C). This allowed parameters in all speed bins to update more quickly at the beginning of the optimization, with a decreased influence as the optimizations within each speed bin converged.

Real-world optimization experiments

The real-world optimization experiment used the untethered exoskeleton to optimize assistance during naturalistic bouts of walking and then evaluated the optimized assistance profiles. Healthy young adults (n = 10, 6 men and 4 women; age = 24.2 ± 1.8 yr; body mass = 67.0 ± 8.2 kg; height = 1.72 ± 0.07 m) completed a two-day protocol. On the first day participants walked outside in a public setting along a path consisting of concrete, asphalt, and brick sidewalks for approximately one hour while the untethered exoskeleton provided assistance during data-driven optimization (FIG. 5B). To emulate naturalistic motion the participants received audio cues to tell them to start and stop walking bouts. The durations of these bouts were randomized from a preselected distribution (FIG. 5D) that matched naturally occurring bout durations. The participant performed a standing rest between bouts for a randomized duration of five to ten seconds. To encourage a normal range of speeds we provided participants with an audio prompt, such as “Walk as if you were walking to catch a bus”, at the start of each bout based on the self-selected speeds associated with these prompts. The randomized distributions of speeds (FIG. 5C) were sampled to match a naturalistic distribution of walking speeds.

On the second day, participants performed outdoor and indoor validation tests to evaluate the benefits provided by the real-world optimized assistance. For the outdoor validation, participants walked along a 566 path in the same public setting with a fixed ordering of bouts of specific distances and corresponding speed prompt commands that were selected to match real-world distributions. The participants completed this outdoor course once for each condition including data-driven optimized assistance, generic assistance, and normal shoes. Participants performed a standing rest for the three minutes following completion of the path to ensure accurate respirometry measurements to compute the total metabolic cost of completing the course. The order of the conditions was randomized. The duration of walking for each bout was timed with a stopwatch. The indoor validation consisted of a standing rest condition followed by six treadmill conditions, each lasting six minutes. The treadmill conditions consisted of walking at 1.25 m/s, 1.5 m/s, and 1.25 m/s with an incline of 10 degrees. The participant completed each treadmill condition twice, once with real-world optimized assistance and once with normal shoes. The order of the conditions was randomized. One participant completed additional indoor conditions of walking at 1.25 m/s with an incline of 5 degrees, at 1.25 m/s with a backpack load of 20% of their bodyweight, and on a stairmill at 50 steps per minute.

Participant surveys on exoskeleton usability

Participants completed a series of surveys to evaluate the ease of use, comfort, and functionality of the untethered exoskeleton after the completion of all experiments. Participants completed a System Usability Scale survey to determine how easy the system was to operate. Users reported our exoskeleton was relatively easy to use (Table 1) with a score of 72.5, which was in the 65th percentile among 5000 devices previously surveyed. Participants also completed surveys adapted from the Orthotics and Prosthetics Users' Survey, which acts as a self-report instrument for evaluating the outcomes of prosthetics and orthotics services in a clinically useful manner. In terms of comfort, participants found it easiest to endorse that the weight of the device was manageable, easy to put on, and their clothes were free of wear (Table 2). In terms of comfort, participants found it hardest to endorse that the exoskeleton was comfortable throughout the day. In terms of functionality, participants found standing, walking indoors and outdoors, and donning/doffing the exoskeleton to be easy (Table 3). In terms of functionality, participants found picking objects up from the ground and walking up a steep ramp as slightly difficult. Six out of ten participants reported that they preferred using the exoskeleton rather than normal shoes during outdoor walking. Table 1. Usability survey results of exoskeleton participants. The System Usability Scale, which uses a Likert scale, was used to evaluate the usability of the untethered exoskeleton. Participants completed this survey after completing all experiments (n = 10). The untethered exoskeleton was in the 65th percentile of a distribution of 5000 devices evaluated with the System Usability Scale. Table 2. Survey results on the comfort of the untethered exoskeleton. This survey was adapted from the Orthotics and Prosthetics Users' Survey, which acts as a self-report instrument for evaluating clinically useful outcomes of prosthetics and orthotics services. Participants completed the survey after all experiments (n = 10). Table 3. Survey results on the functionality of the untethered exoskeleton. This survey was adapted from the Orthotics and Prosthetics Users' Survey, which acts as a self-report instrument for evaluating the outcomes of prosthetics and orthotics services in a clinically useful manner. Participants completed this survey after all experiments (n = 10).