AI-driven universal lower-limb exoskeleton system for community ambulation

Abstract

Exoskeletons offer promising solutions for improving human mobility, but a key challenge is ensuring the controller adapts to changing walking conditions. We present an artificial intelligence (AI)–driven universal exoskeleton system that dynamically switches assistance types between walking modes, modulates assistance levels corresponding to the ground slope, and delivers assistance timely based on the current gait phase in real-time. During treadmill validation, AI-based assistance reduced metabolic cost by 6.5% compared to 3.5% for conventional assistance. We expanded testing the controller in real-world walking, where AI-based assistance showed effective modulation and higher user preference compared to conventional assistance. Leveraging the AI-based approach and a comprehensive dataset, the controller achieved superior performance in environment- and user-state estimations. This approach does not require a separate mode classifier and operates on a user-independent basis, enabling immediate deployment across diverse conditions. This study highlights the potential of AI-driven exoskeletons in facilitating human locomotion in real-world ambulation.

Data-driven controller enables real-time assistance modulation by accurately estimating both the user and the environment states.

INTRODUCTION

Remarkable advancements have been made in the field of lower-limb exoskeletons (1, 2), showcasing their potential to greatly enhance and support human locomotion in various ubiquitous daily activities, such as walking on level-ground (3, 4), sloped surfaces (5), as well as running (6–8). Robotic exoskeletons also bring the hope of solving important societal mobility issues, presenting a potential solution to several long-standing problems related to mobility. They can serve as an assistive device for the elderly or patients with limited walking capability, offering newfound hope for improving their mobility and independence (9–11), and simultaneously as a training tool for patients with gait pathologies, aiding their rehabilitation and recovery (12, 13).

Despite these encouraging developments, there are still challenges to overcome before they can operate synchronously with the user in the real-world environment. Both hardware and controller are critical for the successful deployment of exoskeleton technology, but creating a universal controller that is highly functional across daily activities remains an active, paramount engineering challenge. Specifically, real-world community ambulation involves highly varying walking variables, such as walking speed, walking mode, and ground slope level. Previous studies demonstrated that assistance timing (14–16) and magnitude (17, 18) are critical factors influencing the metabolic benefit of exoskeleton assistance, which is considered the gold standard metric for exoskeleton benefit. In particular, the knee joint increases positive power generation or negative power absorption as the ground slope level becomes steeper during incline walking and decline walking, respectively (19). A recent exoskeleton study highlighted the importance of human-in-the-loop optimization for the user’s metabolic cost, revealing that the peak assistance magnitude should be tailored to the ground slope level (17). This dir.ca.gov/title8/3231.htmlincreasing peak extension moment during the early stance phase as the ground slope increases for both incline and decline walking (19, 20). These suggest that an intelligent and versatile controller that dynamically adjusts assistance levels based on critical user and environmental states may warrant superior benefits of the exoskeleton assistance. However, without accurate real-time estimations, assistance modulation to the varying user and environmental states remains impossible.

To develop more responsive controllers to varying user- and environmental-state variables, such as gait phase and ground slope, heuristic-based estimation methods for wearable robotic applications have been developed (21–25). Alternatively, data-driven approaches have recently emerged as a promising engineering approach for exoskeleton controllers, demonstrating exciting results in recognizing key environmental and user-state information (3, 16, 26, 27), making it smarter than ever. Data-driven approach-based exoskeleton controllers allow for the reduced burden of hand-tuning procedures. These characteristics enable exoskeletons to acquire critical information about the user’s gait in real time—such as biological moment (28, 29), gait phase (16, 26, 30), locomotion mode (27), and ground slope (30, 31)—to control the exoskeleton assistance. Recently, data-driven real-time vision-based mode classification methods for exoskeleton control during walking were validated (27, 32). However, these approaches often require an external camera attached to the user, adding additional equipment and heavy computational processing, which is not the most ideal solution for real-time applications. Alternatively, using mechanical sensors native to the wearable device may simplify the device structure. To resolve this, we propose a purely slope-based mode classification approach using the information from onboard mechanical sensors that is much simpler and less complex than the existing vision-based methods. This is based on representative building construction guidelines such as the Americans with Disabilities Act (ADA) guideline (33) and the California Code of Regulations (34), in which a characteristic of different common modes during community ambulation (stairs, ramps, and level-ground) is the slope level in degrees. This guideline permits a minimum of 19.6° for stairs, and realistically during community ambulation, we very rarely encounter a ramp slope near the minimum stair slope. Therefore, an accurate real-time slope estimation can potentially provide the ground slope level and inform the current user’s walking mode without a separate designated model classifier which often involves an additional external equipment attached to the user and additional computational burden for real-time processing.

Other key variable estimations for exoskeleton control targeted gait phase, a continuous variable representing the phase in a cycle of a movement, to accurately deliver timed assistance, and ground slope to scale the assistance. A number of real-time data-driven gait phase estimators demonstrated high potentials; however, the tested conditions were often limited to a variation in only one or two dimensions (walking speed or mode) (16, 26, 27, 35). It is questionable whether the performance of these gait phase estimators would still be robust when variations in other multiple dimensions are introduced during real-world community ambulation, such as continuously changing slopes, speeds, or walking modes. Our previous deep-learning–based incline slope estimator showed promising results (31); however, it still presented its limited functionality: discrete slope estimation, updated only once per gait cycle, and capable of handling only the upslope setting, not the downslope setting.

Here, we introduce an artificial intelligence (AI)–based universal exoskeleton controller that addresses the presented technical challenges in current exoskeleton controllers, which limit the practicality of exoskeleton deployment in community walking. This controller holistically captures the major variations encountered during community walking in real time: walking mode, ground slope, and gait phase (Fig. 1). Leveraging accurate real-time universal slope and bilateral gait phase estimators, our exoskeleton system achieves the following key functionalities: (i) dynamic switching of assistance types between major walking modes (stairs, ramps, and level-ground) based on the ground slope level, (ii) modulation of assistance levels corresponding to the ground slope, and (iii) timely delivery of assistance to the user based on the current gait phase. This controller seamlessly adapts to the primary variables inherent in community walking, including mode, ground slope, and speed, with little to no delay. We used a unique approach to mode classification by using continuous ground slope estimation for active modulation of real-time exoskeleton control. The modes were sectioned by applying simple thresholds within the continuous slope domain: Slopes greater than 18.5° were classified as stair ascent, between 18.5° and 3.5° as incline, between 3.5° and −3.5° as level ground, between −3.5° and −18.5° as decline, and slopes smaller than −18.5° as stair descent.

Fig. 1. Overall AI-driven universal exoskeleton controller framework.

(A) While the user wearing the robotic exoskeleton walks in community, (B) the data from the onboard mechanical sensors of the exoskeleton, bilateral knee joint encoders and inertial measurement units (IMU), in time series are inputted into the deep-learning–based, user-independent bilateral gait phase estimator and slope estimator, 300- and 800-ms long, respectively. The real-time slope estimate is also used to determine the user’s walking mode. On the basis of the estimations, the exoskeleton assistance is continuously modulated corresponding to the user and environment states (walking mode, ground slope, and bilateral gait phase) in real time.

We examined the performance of the controller and its effects on the user’s metabolic cost and preference through two major real-time validations. The first part of the study (validation I) was primarily designated to assess the hypothesis that modulating the exoskeleton assistance based on the user’s walking mode, slope, and speed would yield a larger reduction in metabolic cost compared to a conventional controller during varying speed and incline walking (N = 10). The conventional controller remained as the level-ground controller throughout walking and used the conventional time-based gait phase estimation (TBE) method. The TBE represents the user’s current gait phase as the time passed from the start of the most recent stride of each leg divided by the estimated duration of the stride, which is computed by averaging the stride durations of the most recent five strides. We evaluated the real-time performance of the controller during incline and decline treadmill walking with varying speed and slope. In this experiment, the metabolic cost of the user under three different incline treadmill walking conditions was measured: AI-based assistance, conventional assistance, and the unpowered condition to evaluate this hypothesis. Under the AI-based assistance controller, the exoskeleton updated the assistance parameters in real-time based on the walking mode, slope, and gait phase estimates, whereas assistance parameters remained static at the level-ground setting throughout the walking period under the conventional assistance condition. The rationale for the hypothesis is that (i) the biological knee extension moment increases as a function of ground slope level during incline walking (19) and that (ii) the metabolically optimized assistance profile also exhibited a scaled assistance profile with respect to the inclination level (17). The second part of the study (validation II) expanded the assessment to indoor/outdoor real-world community walking conditions, including stair ascent/descent, incline/decline, and level-ground walking, involving the same participant cohort from validation I (N = 10). Each subject walked the test circuit twice: once with conventional assistance and once with AI-based assistance. User preference, on a scale of 0 to 10, was asked whether walking with AI-based assistance was more preferred than walking with the conventional controller.

This unique study presents a versatile exoskeleton controller for community walking enabled by the power of AI. Using a universal slope estimator and a gait phase estimator, the system dynamically modulates exoskeleton assistance to address errors in both indoor and outdoor walking settings, covering extensive scenarios of possible modes, slopes, and walking speeds that have never been achieved by previous approaches. The sensor information for our approach relied on the mechanical sensors native to the device, without any additional equipment. Without the burden of subject-specific tuning or attaching additional equipment to the user’s body, this exoskeleton system improves practicality and usability for users. The validations revealed that our AI-based assistance was more metabolically effective and preferred assistance over conventional assistance. Addressing current technical limitations in the exoskeleton technology, this controller enables us to move closer to creating an adaptive and effective assistive technology that seamlessly integrates into the daily lives of individuals, promoting enhanced mobility and overall well-being. This progress brings us one step closer to using the full potential of exoskeleton technology in transforming the way we navigate and interact with the world around us during community walking. The training data used for the models and the exoskeleton hardware are also available as the Supplementary Materials for researchers to leverage these resources for their exoskeleton developments.

RESULTS

Indoor treadmill validation (validation I)

During indoor treadmill validation (Fig. 2), our AI-based gait phase estimator performed significantly better during both incline and decline treadmill walking conditions with 2.48 ± 0.43% of root mean square error (RMSE) on average (see Figs. 3B and 4) compared to the conventional TBE with 3.62 ± 0.51% of RMSE (P < 0.01). The difference in error between the two methods corresponds to an approximately 31.5% error reduction from the TBE. The TBE showed a larger RMSE during incline walking than during decline walking (4.09 ± 0.87% during incline walking and 3.14 ± 0.49% during decline walking), while the AI-based estimator successfully achieved consistent performance across modes and subjects (2.50 ± 0.44% for incline and 2.47 ± 0.49% for decline). TBE performed poorly especially near the end of strides compared to the AI-based estimator (Fig. 3B). The average RMSE right before 0% of the gait cycle, predicting the upcoming heel-contact, for both incline and decline treadmill walking was significantly lower with the AI-based estimator (P < 0.05), 2.11 ± 0.41%, compared to TBE, 6.23 ± 0.85%.

Fig. 2. Experimental setup.

During validation I (indoor treadmill walking), the subject completed two repetitions for each of the three incline walking conditions (AI-based assistance, conventional assistance, and unpowered) in a randomly assigned order with the reverse order during the second half (ABCCBA). In addition, the subject walked the decline condition with AI-based assistance once. The slope was modulated from 0° to 15° during the incline condition and 0° to −15° during the decline condition. Speed modulation followed a sinusoidal wave pattern with varying frequency and amplitude, ranging from 0.7 to 1.3 m/s, centered at 1.0 m/s. Each walking condition lasted 6 min. Metabolic cost was measured during incline walking conditions. Validation II (real-world community walking) occurred at an indoor/outdoor walking circuit at Georgia Tech main campus. The subject started walking from outside, walked into a building, navigated through multiple floors inside the building, exited out of the building, and returned to the starting point. The walking circuit involved various modes (LG, level ground; I, incline; D, decline; SA, stair ascent; and SD, stair descent) and ground slope levels. The step counts (stp) were fixed only for stairs. User preference was assessed after walking with conventional assistance and with AI-based assistance.

Fig. 3. Validation I results for indoor treadmill walking.

(A to C) Exoskeleton controller performance of a representative subject. The shade region indicates ±1 SD. (A) Real-time slope estimator performance and the true slope during 6-min indoor incline (left) and decline (right) treadmill walking tests. Continuous slope estimation (top) was executed at 100 Hz with an RMSE of 1.30° on average between subjects and walking modes. The exoskeleton toggled the continuous slope estimation once per gait cycle at the start of a new gait cycle detected by real-time gait phase estimation of the ipsilateral leg. (B) Gait phase estimation comparison between our AI-based and the conventional time-based methods during the incline (left) and decline (right) treadmill walking tests. Our AI-based gait phase estimation method outperformed in accuracy compared to the time-based estimation method (P < 0.01). (C) Average torque profiles (top) and the torque profiles in time-series (bottom) during incline treadmill walking under the AI-based assistance and the conventional assistance. (D) Metabolic cost comparison between incline walking conditions: AI-based assistance, conventional assistance, and unpowered. Average net metabolic cost (left) shows that the AI-based assistance led to significantly lower metabolic cost than the conventional assistance and the unpowered conditions. This trend was very commonly present across individual subjects (S) except for S4 (right). The asterisks (*) indicate significant differences between conditions. The error bar indicates ±1 SEM.

Fig. 4. Average estimation performance in RMSE computed between the estimates and the ground-truth for validation I (white region) and validation II (yellow region).

The AI-based gait phase estimation was significantly outperformed compared to the time-based method only during the indoor validation (P < 0.05). The RMSE in slope estimates for exoskeleton control, only per gait cycle, was consistent with the continuous slope estimates but was significantly lower than the continuous estimates during real-world community validation (P < 0.05). This was attributed to much smaller RMSE for the slope estimates for exoskeleton control during the transition steps between modes. The slope-based mode classification was successful with low RMSE for both validations. The asterisks (*) indicate significant differences between conditions. The error bar indicates ±1 SEM.

The continuous slope estimator achieved accurate performance during the treadmill walking with 1.30° ± 0.42° of RMSE, on average, during both incline and decline treadmill walking (see Figs. 3A and 4). During treadmill walking, the continuous slope estimates during the decline walking condition, 1.09° ± 0.23° of RMSE, performed significantly better than during the incline condition, 1.52° ± 0.67° of RMSE (P < 0.05). The variation in the performance of the continuous slope estimator across subjects was nearly three times larger during incline walking compared to decline walking. Notably, the performance of the slope estimator during the updates on the exoskeleton controller, 0% of the gait cycle of each leg, showed a robust, consistent accuracy across both modes and subjects (RMSE: 1.19° ± 0.38° averaged across modes).

The slope-based mode classifier successfully detected walking modes in real time using the slope estimate, achieving an average accuracy of 2.97 ± 1.42% during treadmill walking across modes. This mode classification in the high-level controller was translated to toggling the corresponding mid-level controller (incline, decline, or level-ground mode) based on the detected mode. This slope-based mode classification allowed for switching the exoskeleton assistance mode corresponding to the estimated mode and modulated the assistance type (stiffness or damping assistance mode), and the assistance parameter was modulated based on the magnitude of the estimated ground slope. This allowed for a larger magnitude of assistance under the AI-based assistance than conventional assistance walking conditions (Fig. 3C).

Metabolic comparison between walking conditions

The assistance modulation using our AI-based controller on the incline treadmill varying slope and speed walking condition significantly reduced the user’s metabolic cost compared to the unpowered condition by 6.2 ± 2.8% (P < 0.05), on average, across subjects (Fig. 3D). The AI-based controller was also significantly more effective in offloading user’s energetic demand than using the conventional controller, although the conventional controller also significantly reduced the user’s metabolic cost compared to the unpowered condition by 3.5 ± 2.8% (P < 0.05). This trend in metabolic cost between walking conditions was consistent across subjects except for one (Fig. 3D).

Real-world community validation (validation II)

The real-time slope estimation RMSE during exoskeleton control updates (once per gait cycle at heel contact of each leg) was 1.86° ± 0.24° (Figs. 4 and 5). The average slope estimation RMSE for exoskeleton controller updates during outside level-ground mode was 2.62° ± 0.35° for steady state and 3.73° ± 1.16° for transition steps only. The slope-based mode classifier successfully classified the mode of walking during the real-world community walking experiment with 2.19 ± 0.43% of RMSE, involving 757.7 ± 30.8 steps, on average, to complete the real-world test circuit during the AI-based assistance condition. The average mode classification demonstrated lower RMSE during steady-state locomotion (1.88 ± 0.28%) compared to transition steps (10.7 ± 5.8%). This discrepancy arises from the nature of our slope-based mode classification method, where the walking mode is updated once per gait cycle at each heel contact. For the outdoor testing circuit, changes in slope and mode for ground-truth labeling were determined at the first instance when the user was fully in the new slope or mode (i.e., both feet were in the new mode). During transition steps, the continuous slope estimate began to change in response to the user’s movement. However, if the slope estimate approached the ground truth slope of the new mode either too early or too late during the transition, it led to a misclassification where the mode either lagged or led the ground truth (see subplots in Fig. 5). The AI-based gait phase estimation resulted in 2.22 ± 0.24% of RMSE, which was slightly, not statistically significantly, lower compared to the TBE method. The AI-based gait phase estimation also achieved significantly lower RMSE at the end of the gait cycle (2.00 ± 0.42%) compared to the TBE method (4.23 ± 1.10%).

Fig. 5. The performance of the real-time AI-based continuous slope estimation and mode classification and corresponding torque assistance of a representative subject during the real-world testing.

Six subsections of the whole period of circuit completion (middle) are zoomed in and represented in the subplots with the correspondingly colored dots. The subjects walked in the real-world testing circuit at their preferred walking speed. The circuit involved a wide range of modes and ground slope conditions, −33° to 33°.

The average preference scale was 7.2 ± 1.3 across subjects, indicating that the AI-based assistance condition was more preferred (5, neutral) than the conventional assistance condition for real-world testing. Nine subjects scored the preference scale above neutral, 5, except for a single subject scoring 4. The reason why the subject preferred conventional assistance over AI-based assistance was that the conventional assistance felt better during the stair ascent mode. Some notable comments from the subjects who preferred the AI-based assistance over the conventional assistance were that the AI-based assistance was more natural, mode transition and assistance timing were more intuitive, and the assistance timing and magnitude during decline and stair descent modes (stiffness controller during the convention assistance versus damping controller during the AI-based assistance) feel better compared to the conventional assistance.

DISCUSSION

The AI-based approach allows for accurate real-time user and environmental state estimations without user-specific calibration procedure

The real-time performance of our environmental state (mode and ground slope) and the user state (gait phase) estimation demonstrated a compelling result for real-time exoskeleton control compared to other recent works. First, the slope estimation performed substantially better compared to our previous study (1.5° of RMSE for 0° to 15° treadmill walking) (31), and this current work is capable of handling both incline and decline walking condition with a wider range of slope using a single unified estimator. Compared to many other previous gait phase estimators tested in real-time (16, 26, 35, 36), our AI-based gait phase estimator demonstrated superior performance in terms of RMSE in a more diverse and challenging test condition with high variation in different dimensions (mode, speed, and slope). Notably, the tested condition for this study covered a wider range of both slope and speed than one of the most recent data-driven gait phase and slope estimators for a powered ankle exoskeleton (gait phase RMSE: 4.8 ± 2.4% and slope RMSE: 2.4 ± 1.3° for the treadmill walking condition with ± 10° inclination and a speed of 0.8 to 1.2 m/s) (36). We attribute these changes to two key improvements: A vastly expanded dataset collected across numerous conditions and users and an optimized deep-learning system. Another important aspect of our controller is that the variability of the performance in slope and phase estimation across subjects was much smaller, while the average RMSE maintained a better performance than the data-driven method by Leo et al. (36). In addition, our approach does not require additional sensors than the sensors already existing on the knee exoskeleton and without any sensors attached to the foot or ankle for real-time operation, which are often used to detect ground slope but unideal for knee exoskeletons, to achieve accurate performance. Furthermore, our system was engineered on a user-independent basis. In a real-world setting, this requires no prior user-specific data, adjustment, or tuning for the system except for initializing the knee angle to 0° to function robustly and consistently. With these key characteristics, the device can be immediately deployed across different users, days, and uses, and it requires substantially less time and effort to implement than other systems that are user dependent, which necessitate complex calibration steps for specific users. This key feature also increases the practicality and usability of such an exoskeleton system during daily life.

There are a few main potential reasons why our methods led to a better performance in estimation compared to previous works. The first is the difference in the sensor information features. While all approaches used mechanical sensor information, inertial measurement units (IMU) or encoders, as the input into the data-driven models, our approach used IMUs attached to the thigh and the shank, bilaterally capturing the dynamics of the whole leg movement above the ankle joint. Kang et al. (35) used a six-axis IMU attached to the trunk of the exoskeleton and bilateral hip joint encoders, capturing the movement of the user’s trunk and hip angle and velocity. Real-time gait phase estimation for an ankle exoskeleton used the mechanical sensor information only surrounding the ankle joint: a shank IMU and an ankle encoder for Shepherd et al. (16) and shank and foot IMUs for Medrano et al. (36). Therefore, our study collected more holistic information about the dynamic movement of the whole leg by incorporating both thigh and shank IMUs, which may have facilitated the model in learning key kinematic features for fine distinctions along the gait phase regression. The second reason is the difference in the data-driven approaches (comparing the extended Kalman filter approach by Medrano et al. and our deep-learning approach) and between deep-learning approaches in terms of model structure (such as the number of features, kernel sizes, and many others). The third reason is the differences in the training datasets. Medrano et al. (36) used a training dataset based on the simulated IMU information based on the human walking dataset, not IMU information directly from the exoskeleton. This method poses a potential disadvantage in the case if a discrepancy between the exoskeleton and the user’s body, which can occur if the user’s orthotic exists at the intersection between the exoskeleton and the user’s limbs. Our training data were collected from the mechanical sensors attached to the exoskeleton, not to the user. Therefore, the training data are potentially better accommodating of the variations and noise that may have been introduced during the real-time validation testing. The fourth reason is the variation within the training dataset. Our training data covered a wide variation of dimensions including walking setting (treadmill and overground), modes (stairs, ramps, level-ground), speeds, slopes, assistance, and users and were one of the most comprehensive to date for gait phase estimation. The raw data collected to train the model are also available along with this publication as the Supplementary Materials for any researchers interested in leveraging the data for their innovation.

The error in the real-time slope-based mode classification approach of the study demonstrated a more competent result across five different walking modes (stair ascent/descent, ramp incline/decline, and level ground) compared to other real-time mode classification methods for exoskeletons which often require much heavier computational processing and complex architecture (27, 32, 37). Our approach also achieved robust performance by solely using onboard sensors on the exoskeleton, whereas similar vision-based works often used an external camera attached to the user to recognize the modes (27, 32). Therefore, compared to other designated data-driven mode classifiers, our approach simplifies the equipment and overall system complexity by simply differentiating the walking mode based on the accurate slope estimate by thresholds with comparable or better accuracy than state-of-the-art methods.

Our controller, which captures both external and internal states for real-time control, can be compared to the recent development of the biological torque controller for a powered hip exoskeleton, widely regarded as one of the most advanced exoskeleton controllers to date. The biological torque controller solely relies on the estimated joint moment in real time, determining the assistance magnitude by multiplying the estimate by a constant scale factor and applying a time delay. Essentially, it maps the user’s biological needs at the joint to modulate assistance in real time. In contrast, our controller captures both environmental states (e.g., walking mode and ground slope) and user states (e.g., gait phase), which are critical for walking. Meanwhile, the biological torque controller only has access to the user’s internal state (biological hip moment). When tailoring assistance parameters to specific walking modes—such as maximizing assistance benefits or prioritizing safety during stair descent or decline walking—our controller would be a more suitable approach, as it adjusts torque assistance based on both user and environmental states. In comparison, the biological torque controller provides torque assistance purely based on biological torque using a global mid-level controller without accounting for environmental changes. Recent findings from human-in-the-loop optimized assistance (38) and Molinaro et al.’s work (29) suggest that torque assistance resembling the biological torque profile is not the most optimal form of assistance. This indicates that even with a biological torque controller, assistance parameters still need to be tuned (e.g., through low-pass filtering, scaling, and delaying) to achieve the most effective user outcomes. In addition, it remains unknown whether constant assistance parameters would be globally optimal across different locomotor tasks. Therefore, mode-specific assistance parameterization would likely be required to optimize performance across different modes of locomotion. In addition, it is currently unknown whether the biological torque controller would serve as an optimal assistance controller across different modes at the knee joint without experimental validation. The primary functional role of the hip joint is power generation during community ambulation, whereas the knee joint’s role varies substantially—acting as a power absorber during descent modes and a power generator during ascent modes. For this reason, the knee joint functions similar to a spring during upward steps and similar to a damper during downward steps (39). This functional variability supports the use of our pseudo-impedance controller, which appropriately switches between stiffness control and damping control depending on the recognized mode. We believe that these two assistance controllers offer distinct approaches to exoskeleton control, and merging them could further enhance exoskeleton capabilities for assisting community ambulation. For example, the biological torque controller’s assistance parameters—such as a scale term (0 to 100%) and a delay term (in milliseconds) applied to the estimated biological moment—may not be globally optimal across different walking conditions. Incorporating our high-level controller, which recognizes walking conditions, could help optimize these parameters for each walking mode, thus maximizing the benefit of exoskeleton assistance. In addition, the mode classification and ground slope estimator of our work may offer an opportunity for designing an intelligent controller for individuals with knee osteoarthritis that provides more optimized assistance strategy to the activities that they experience a larger challenge due to pain such as stair walking (40, 41).

Assistance modulation based on the user and environmental state variables led to a larger metabolic reduction

Our AI-based intelligent exoskeleton controller allowed the assistance to be adjusted to the environmental states and better synchronized to the user’s motion by timely delivering the assistance in real time during both indoor and outdoor walking conditions. During incline treadmill walking, the assistance modulation using the AI-based controller achieved significantly larger metabolic reduction than the assistance using the conventional controller (P < 0.05). This indicates that although both the AI-based and the conventional assistance significantly reduced the metabolic cost of the users compared to the unpowered condition, our AI-based assistance approach was significantly more effective in reducing the user’s metabolic cost than the conventional assistance approach (P < 0.05). This result supports the acceptance of our study’s central hypothesis that the AI-based assistance condition would lead to a more significant metabolic reduction for the user than the conventional assistance during locomotion that included varying inclines and walking speeds. This metabolic trend was consistent across subjects in validation I, except for one subject, AB04. For AB04, the AI-based assistance was functionally modulated in response to changing inclination levels during validation I, resulting in a substantial difference in average peak assistance magnitudes between the two assistance conditions: 11.0 Nm under the AI-based assistance and 6.4 Nm under the conventional assistance. However, beyond this, because of the lack of additional secondary outcome measures (e.g., biomechanics or EMG), it is unclear why the metabolic cost was not reduced in this subject.

The timing of the assistance in the gait cycle and the assistance parameter modulation based on the mode and the magnitude of the slope are the key components differentiating the AI-based assistance compared to the conventional controller. Under the stiffness controller (for level-ground and upslope modes), the peak knee flexion angle during the early stance phase alone leads to scaled assistance to the ground slope level due to a larger angular deviation from the equilibrium angle (0°). However, the metabolic result from the treadmill testing indicates that the combination of scaled assistance parameters in conjunction with the scaled knee flexion angle with the slope level, enabled by the AI-based controller, has been more effective in offloading the user’s physical effort. The timing of the assistance could have been another main factor that resulted in the metabolic difference under the two assistance conditions. However, although the RMSE of our AI-based gait phase estimation was significantly lower than the TBE-based (P < 0.01), the average difference in the RMSE between the two methods, 1.6%, was not considerably large during incline treadmill walking. On the other hand, the peak assistance magnitude showed a more significant difference between the two assistance conditions. Therefore, we speculate that the scaled assistance to ground slope level achieved by our ground slope estimator may have been a larger influencing factor in resulting in a more metabolically effective assistance strategy for dynamically varying walking conditions compared to the conventional controller. This indicates that exoskeletons that dynamically adjust their assistance to accommodate varying ground slopes in a biomimetic manner may have advantages over conventional controllers that largely use the level-walking mode to accommodate inclines (5) (a very common simplification in control design). However, there is still potential for further refinement of the assistance strategy to adapt more precisely to varying walking conditions in future studies. The current scaling method for assistance parameters is based on a simplified linear increase relative to the ground slope for each walking mode. While this linear approach effectively captures the general trend within each mode, it does not fully account for the more nonlinear relationship between ground slope and peak biological moment. Future developments could enhance exoskeleton assistance by integrating our high-level controller, which accurately estimates walking mode and ground slope, with an advanced parameter mapping method that more closely aligns with the intricate relationship between environmental conditions and biological demands.

Historically, the number of autonomous exoskeleton research studies investigating the effect of assistance on the user’s metabolic cost with a powered knee exoskeleton has been relatively sparse compared to studies with a hip or ankle exoskeleton (2, 42). The metabolic reduction with the conventional controller during incline treadmill walking (3.5% on average) is comparable to the metabolic reduction achieved by the powered assistance using bilateral knee exoskeleton (3.0%) during 8.5° incline walking at 1.1 m/s, as conducted by our study group in the past (43). In this study, our AI-based assistance, modulating the assistance to the variations in the walking condition, took a step further in becoming a more viable assistance strategy for metabolic reduction with a 6.2% metabolic reduction from the unpowered condition on average, around twice of benefit compared to the conventional controller. On the basis of the author’s knowledge, MacLean and Ferris’s knee exoskeleton (5) is the only fully autonomous powered knee exoskeleton that showed metabolic reduction with knee assistance compared to walking without the device (4.2%) during 15° incline walking while wearing a heavy backpack while their assistance led to an increase in metabolic cost during level-ground and incline walking without the backpack. However, the controller used in their study is not fully disclosed, which limits future studies to replicate the controller for comparison. Our study discloses the technical details of the controller design and the hardware design, making it possible for researchers to build upon our development for future studies.

A limitation of this study was that the unpowered condition was used as the baseline of the comparison, rather than walking without the exoskeleton. In a small pilot testing (N = 2) on treadmill with the slope changing both incline and decline walking and the speed changing similar to validation I, the effect of added mass from wearing the device resulted in about 11% increase on average (refer to Supplemental Materials for more details). Although our machine-learning–based assistance condition was more effective in offloading the user’s energetic consumption compared to the conventional assistance condition, we did not evaluate which component in the difference between the two conditions, the more accurate gait phase or the assistance modulation, contributed more to the metabolic difference. Future studies may explore whether the metabolic benefit is more sensitive to the timing or the assistance magnitude.

In summary, this work showcases cutting-edge advancements in real-time control and assistance modulation for exoskeletons during community ambulation. Notably, this high performance in critical environmental-state (ground slope and walking mode) and user-state estimations (bilateral gait phase) are achieved using the mechanical sensors native to the device without the need for additional sensors, ensuring seamless deployment across diverse users and conditions without any user-specific data or calibration. This performance result is attributed to its holistic approach, capturing the dynamic movement of the entire legs through bilateral thigh and shank IMUs, capturing a wide spectrum of variables with a comprehensive training dataset, and leveraging AI techniques. Furthermore, our slope-based mode classification enabled equivalent or better performance than previous methods that often required an external vision camera, simplifying equipment and computational requirements. Our real-time AI-based assistance allowed for dynamic assistance modulation to varying walking mode, ground slope, and bilateral gait phase with high accuracy exceeding previous engineering approaches during both indoor and real-world community walking settings. Compared to conventional assistance, our AI-based assistance that modulated assistance according to environment and user states was more metabolically effective when tested under the varying speed and incline walking condition and was significantly more favored during the real-world community walking condition. Real-time environment and user state estimations using a single-mode agnostic slope estimator and gait phase estimator achieved state-of-the-art performance, and this allowed for more favored, accurate, and timely delivered exoskeleton assistance corresponding to the variations in modes, slopes, speeds, and users. Our AI-based system demonstrates an exciting stride close toward the deployment of an intelligent exoskeleton that assists human locomotion in real-world community ambulation.

MATERIALS AND METHODS

Exoskeleton system

Experiments were conducted using our bilateral, lightweight, autonomous robotic knee exoskeleton (Fig. 6) capable of assisting the user’s knee flexion/extension (9 and 15 Nm as the max continuous and the peak torque, respectively). A planetary gearhead (9:1 gear-ratio, quasi-direct drive) integrated powered actuator (AK-80-9 KV100, T-motor, China) serves as the source of torque for the exoskeleton. The waist orthotic (Orthomerica, USA) securely holds onto the user’s pelvic area. Each side of the exoskeleton consists of a two degree-of-freedom (DOF) free-pivoting joint, which bifurcates the rigid bar, allowing the user free movement of the hip. In addition, to accommodate any misalignment of the exoskeleton and the user’s leg along the axis of the rigid bar, a one DOF locking joint is incorporated in series with the two DOF free-pivoting joint. The exoskeleton is composed of multiple onboard mechanical sensors. Two six-axis IMUs (MPU9250, InvenSense, USA) are mounted on each leg of the exoskeleton- one on the thigh bar and another on the shank bar. A 12-bit rotary encoder installed in the motor driver measures both the actuator angle and angular velocity. The main onboard microprocessor (NI myRIO, National Instruments, USA) facilitates communication between electronics. The real-time machine-learning model runs on a comicroprocessor (Jetson Nano, NVIDIA, USA), and the data are exchanged between the main exoskeleton microprocessor and the comicroprocessor via Transmission Control Protocol/Internet Protocol (TCP/IP) communication. All communication and exoskeleton control updates are executed at 100 Hz. The total mass of the robotic exoskeleton system is 4.4 kg including all electronics, weighing 1.4 kg. The CAD files of the exoskeleton hardware are available for researchers who are interested in replicating the exoskeleton system, and further related information is noted in the Supplementary Materials.

Fig. 6. The bilateral powered knee exoskeleton.

The device is consisted of multiple bilateral onboard mechanical sensors (thigh and shank IMUs, knee joint encoders, and force-sensitive resistors).

Participants

For training data collection, the total of 16 able-bodied adults (nine males/seven females, mean ± SD, 23.6 ± 4.1 years old, 173.3 ± 5.6 cm, 70.0 ± 9.3 kg) were recruited to complete training data collection. The subjects participated in overlapping tasks: 10 subjects on treadmill, 10 subjects on overground ramps, and 9 subjects overground stairs. For indoor and real-world validation experiments, 10 able-bodied adults (eight males/two females, mean ± SD, 24.2 ± 1.5 years old, 173.9 ± 1.4 cm, 72.9 ± 3.4 kg) were recruited. Five subjects for the validations previously participated in the training data collection. All subjects provided written informed consent to participate in this experiment approved by the Georgia Institute of Technology Institutional Review Board (protocol number: H19167).

Experimental protocol

This study consisted of three major experiments: training data collection for deep-learning models, indoor validation on treadmill (validation I), and real-world validation in a real-world community ambulation circuit (validation II). Separate models were tested between validations: indoor models for indoor treadmill validation and real-world models for real-world community ambulation validation. Both indoor and real-world models share the same training data from treadmill and overground ramps, but only real-world models additionally included the training data from overground stairs to accommodate the presence of stair modes in the walking circuit for the real-world community testing. Note that to maintain the tested models being user independent, if a subject also participated in the validations after participating the training data collection, then the models tested for the subject were trained with the training data excluding data from the particular subject.

Training data collection protocol

The training data collection was designed to capture a wide variety of variables: mode, slope, walking speed, and exoskeleton assistance (Fig. 7). The cohort participating in training data collection varied and partially overlapped between treadmill (N = 10), overground ramps (N = 10), and overground stairs (N = 9) from the total of 16 participants. The indoor validation was intended to evaluate the capability of AI-based slope and gait phase models during level-ground, incline, and decline treadmill walking. We first collected training data on both overground ramps and treadmill walking settings with the subjects walking with the exoskeleton. Force-sensitive resistors (FSRs), attached to the user’s heel, were only used for labeling the ground-truth heel contacts for gait phase segmentation (for validation), not as a part of inputs into models for real-time estimations. The data from the knee joint encoder, shank IMU, and thigh IMU located on both legs were recorded during walking. This information was later used to train the indoor gait phase and slope models, which were tested during validation I. Before starting data collection, the bilateral knee joint encoders were initialized to 0° while the subject was standing straight. This step is essential because the exoskeleton uses incremental encoders rather than absolute encoders.

Fig. 7. The walking conditions of the training data collection for the indoor model and the real-world model.

The training data for the indoor model included the user walking (i) on the treadmill at the various slope, speed, and exoskeleton assistance parameters and (ii) on the overground ramps at various slopes. The real-world model incorporated the training data collection on overground stairs at various slopes in addition to the identical training data used for the indoor model.

For treadmill data collection, the subjects walked on a treadmill (CAREN, MoTek, Netherlands) at varying inclinations ranging from −14° to 14°, changing by 2° increments every 2 min. During walking, the walking speed was set to three levels (0.6, 1.0, and 1.4 m/s) for 20 s each for the first minute, and this speed setting was repeated again for the second minute. Two different exoskeleton assistance types were also applied to the user during the trial. During the first minute, the exoskeleton provided assistance using the predetermined parameters corresponding to the mode and slope. During the second minute, the exoskeleton provided assistance using the stiffness or damping assistance corresponding to the mode (incline, decline, or level ground) with random assistance parameters updated once per gait cycle to introduce variation in the input data to corresponding labels. The two stiffness terms under the stiffness controller, one for extension assistance during early stance and another for flexion assistance near swing flexion, were assigned to random values between zero to maximum stiffness values during incline and level-ground walking, 1.2 Nm·deg.⁻¹ for extension assistance and 0.6 Nm·deg.⁻¹ for flexion assistance. Similarly, the damping coefficient under the damping controller during decline walking was also assigned to a random value between zero to maximum damping coefficient value during decline walking, 0.12 Nm·s·deg.⁻¹

On the overground ramp setting, the subjects started walking at their preferred pace on level ground before entering a 5-m-long ramp at various inclination levels (± 5.2°, 7.8°, 9.2°, 11.0°, 12.4°, 15.5°, and 18.0°). Exoskeleton assistance was applied with corresponding assistance to the mode and slope. The overground ramp ascent/descent walking was repeated with the goal of 10 passes per slope condition. Last, the subject walked with the goal of 10 passes on an 8-m-long level-ground overground path. For the training data, transitions between the modes were labeled as a step change at the 75% gait cycle of the leg, with the first step landing into the new mode at the end of the gait cycle.

Last, to further expand our examination of the comprehensive capability of the AI-based slope and gait phase models under more extreme conditions for real-world community testing, we additionally collected training data of nine able-bodied adults overground walking at various stair levels (±19.6°, 24.6°, 30.0°, and 35.3°) with the goal of 20 passes at six steps per pass. The subjects walked at their preferred pace, and exoskeleton assistance corresponded to the mode and ground slope. The ground truth labels for gait phase were generated by linearly scaling the time between heel contacts, detected by the FSR data, to represent 0 to 100% of the gait cycle. The entire raw data used as the training dataset are available as the Supplementary Materials.

Validation I—Indoor treadmill

Participants were asked to walk with the exoskeleton on a treadmill at dynamically changing slope and speed, according to predetermined profiles (Fig. 2A). The treadmill interfaced with MATLAB via TCP/IP communication. For the 6-min walking trial, the ground slope was initially set to 0° for the first 25 s, and the maximum slope (15°) was commanded. After the slope reached the maximum, the treadmill remained at the maximum for 20 s before descending back to 0°. Once the slope reached 0°, it remained there for 25 s until the trial finished. The average slope of the slope profile was 8.4°. The walking speed profile also dynamically varied during the trial. The speed started at 1 m/s and remained constant for 15 s. Then, the speed followed a sinusoidal profile, which updated once per second, with a maximum peak set to 1.3 m/s and a minimum peak set to 0.7 m/s throughout the trial (see Fig. 2A). The variation in the speed profile started with 60-s period for 60 s, 30-s period for 60 s, and 22.5-s period for 45 s. This completed the first half of the speed variation, and the order was reversed for the second half of the 6-min walking trial. The goal was to broadly examine the real-time performance of the controller at many speed and slope combinations.

On the first day, the subjects were fitted with the exoskeleton and practiced walking with the exoskeleton providing assistance based on the AI-based controller on the treadmill for 30 min for the acclimation purpose. The 30 min of walking was broken down into 6-min sections with optional breaks in between: a 6-min walking trial on the level ground at 1.0 m/s, a 6-min walking trial on 8.5° inclined surface at 1.0 m/s, and three 6-min walking trials with the testing slope and speed profiles.

The data collection for the indoor validation was performed on the second visit. Three exoskeleton conditions were tested: an unpowered condition, a conventional assistance condition, and the AI-based assistance condition. Under the conventional assistance condition, the timing of the assistance was controlled by the TBE, and the mid-level assistance remained conventional as the level-ground controller. Under the AI-based assistance condition, the assistance was dynamically modulated by the outputs from the AI-based estimators. Before beginning the first trial, the resting metabolic cost of the subject was measured for 4-min, while the subject is standing with the exoskeleton worn. Before beginning data collection for walking, the subject walked on the treadmill for 5 min (evenly distributed between 0° and 8.5° at 1.0 m/s) for the reacclimation purpose. Next, the user walked for six intervals for incline walking trials, two per assistance strategy, for 6-min per interval. To assign each assistance strategy to an interval, we followed an ABCCBA pattern and randomly assigned the walking conditions to A, B, or C for each user. Throughout the incline trials, the user’s metabolic cost was measured using indirect calorimetry to measure the volume of oxygen uptake and carbon dioxide production. Between each interval, subjects were offered an optional 3-min break. After finishing the incline walking trials, the subject lastly walked on the treadmill for a decline walking condition with the exoskeleton providing the AI-based assistance for 6 min.

Validation II—Real-world validation in a community walking circuit

The last visit was designated for evaluating the real-time system performance in a real-world walking scenario. The outdoor/indoor overground circuit is composed of various levels of stairs, incline/decline surfaces, and level-ground walking at Georgia Institute of Technology campus (see Fig. 2B). The subject completed walking the circuit twice: once with the conventional controller and once with the AI-based controller. After completing both walking trials, the subject was asked whether the AI-based assistance was more preferred on a scale of 0 to 10 (0, strongly disagree; 5, neutral; and 10, strongly agree) with the walking conditions referred to condition 1 and condition 2 to blind them to the subject. The order of the walking conditions was randomized for each subject.

AI-driven slope and gait phase estimators

Model design and optimization

We implemented a deep convolutional neural network (CNN) to estimate the user’s bilateral gait phase and slope using the mechanical sensor data from the exoskeleton. For treadmill testing, the training data collected from the treadmill and overground ramp walking were used to train the model. The training data incorporated in the real-world community models included the training data used for the treadmill experiment and on overground stairs (see Fig. 7). With 28 input channels from mechanical sensors (14 inputs from each leg: two inputs from the knee joint encoder, and six inputs from each of the two IMUs for gyroscope and accelerometer). The raw data from IMUs were used as inputs.

Tto facilitate the engineering approach in a user-independent fashion, if a subject participated in both the training and testing experiments, we eliminated the subject’s training data from the deep-learning model training set (trained with all other subjects) before deployment for that subject. The purpose of this approach is to avoid the models to learn the specific user’s walking during the training process. The process of architecture and hyperparameter optimization process for the models are more detailed in the Supplementary Materials.

The gait phase estimator model architecture consists of six one-dimensional convolutional layers and four fully connected layers. Using the additional four channels of offline gait phase estimator output (x and y polar coordinates representing the user phase for each leg) and 28 input channels from the mechanical sensors, our CNN architecture for the slope estimator consisted of two one-dimensional convolutional layers, two one-dimensional average pooling layers, and two fully connected layers. After each convolutional layer, batch normalization and activation functions were added to optimize the training for the model, and a one-dimensional pooling layer was added to reduce computational cost. The input sequence length to the model was 300 ms for the gait phase estimation model and 800 ms for the slope estimation model. Both models were trained offline using Python v 3.6.9 and the deep-learning package, Pytorch v1.4.0. The gait phase model was trained using the Adam Optimizer, with a learning rate of 0.0005 and batch size of 256, for 150 epochs. Both the convolutional and the fully connected layer of output channel were activated using the Sigmoid function. The loss function was computed as the RMSE of the angular error between the estimated and true gait phase converted into polar coordinates. The slope phase model was trained using the Adam Optimizer, with a learning rate of 0.0001 and batch size of 128, for 150 epochs. The loss function was computed as the RMSE between the estimated and true slope. Both convolutional and fully connected layers of the output channel were activated using the rectified linear unit.

Real-time implementation

For real-time exoskeleton control, the raw bilateral gait phase estimator outputs were used to determine the continuous user’s gait phase for each leg separately. Then, these gait phase estimates were concatenated with 28 channels of mechanical sensor data as the input to toggle the raw continuous slope estimate. The raw continuous slope estimate from the model was averaged with a 150-ms moving window (15 samples for 100 Hz), smoothing the profile, to generate the final continuous slope estimate profile for the exoskeleton control purpose.

Exoskeleton controller

High-level layer

The high-level layer of the exoskeleton controller first determines the mode that the user is currently in. The method is purely based on the estimated ground slope with ±3.5° and ±18.5° as the boundaries (Fig. 8). The mode is set to level ground if the estimated slope is within the ±3.5° boundaries, incline mode if the estimated slope is between 3.5° and 18.5°, and decline mode if the estimated slope is between −3.5° and −18.5°. The mode is classified as stair ascent or stair descent mode if the estimated slope is above 18.5° or below −18.5°, respectively. This is a unique approach in that it does not necessarily require a separate discrete mode classifier for switching the exoskeleton assistance depending on modes, which also reduces additional processing task to the microprocessor during real-time operation. The real-time slope estimation is used to determine the assistance parameters, and a single-mode agonistic gait phase estimator that outputs bilateral estimations is used to control the timing of the assistance in the mid-level control layer.

Fig. 8. Overall AI-based exoskeleton control design diagram.

The onboard mechanical sensor data (IMU and knee joint encoder) are passed into the real-time bilateral gait phase estimator, and the senor data and the bilateral gait phase estimation are combined and input into the real-time slope estimator. The slope estimation is used to classify the walking mode using thresholds in degrees and to determine the assistance parameters. Then, the bilateral gait phase estimation is used to control the timing of the assistance to the user. Last, the determined torque command is executed at the low-level control layer.

Mid-level layer

The mid-level assistance strategy was a pseudo-impedance–based controller (Eq. 1) (44)

$$\mathrm{\tau }=k(\mathrm{\theta }-{\mathrm{\theta }}_{\mathrm{e}})+b\stackrel{·}{\mathrm{\theta }}$$ (1)

modeling the joint as a spring and damper system in which τ is the torque magnitude, k is the stiffness term, b is the damping coefficient, θ is the knee angle, θ_e is the equilibrium angle, and $\stackrel{·}{\mathrm{\theta }}$ is the knee velocity. The knee angle and velocity are obtained from the encoder attached to the actuator of the exoskeleton. The following sign convention was used for the study: Positive torque corresponds to extension torque, and positive knee angle and velocity correspond to flexion. Depending on the determined mode, either k or b is set to zero or approximately zero to mimic a spring or damping controller, respectively. The slope output of the continuous slope estimator was toggled once at the start of every new gait cycle, detected by the gait phase estimator, to update the assistance parameters and the mode of the mid-level controller, which remained unchanged for the rest of the gait cycle. Each leg was controlled and updated independently. The stiffness under the stiffness controller and the damping coefficient under the damping controller were linearly ramped from zero to the target value over the ramp time starting at the onset timing of the assistance. This was to avoid an abrupt torque output change during the transition between an unassisted state to an assistance state. In addition, the assistance parameters in Eq. 1—k for stiffness controller and b, under the corresponding assistance strategy, for stiffness controller or damping controller—were linearly scaled with respect to the ground slope. This scaling approach captures the general trend of the peak biological moment observed with varying ground slopes, ensuring that the assistance provided aligns with the increasing demands of the knee joint in different walking conditions (19).

In the joint power domain, the knee joint is highlighted as a power absorber with a high extension moment during the early stance phase for decline or stair-descent walking, rather than a power generator because a damping controller is often used in powered prosthesis devices for walking modes (45, 46). Thus, the damping controller, where the stiffness is zero and the damping coefficient is nonzero for assisting power absorption, was used for the decline or stair-descent walking modes for this study (Fig. 8). Under the damping controller, extension assistance was activated at heel contact until the extension offset timing with a nonzero damping coefficient and remained 0 Nm for the remainder of the gait cycle. The damping coefficient was linearly scaled from 0 to 0.12 Nm·s·deg.⁻¹ for 0° to −18.5° and 0.12 to 0.16 Nm·s·deg.⁻¹ for −18.5° to −40.0° (Table 1). The assistance parameters of the Eq. 1 were chosen considering the maximum torque capability of the exoskeleton (15 Nm) and user’s feedback in preference from a pilot testing that we performed in the initial stage of the controller development phase.

Table 1. Assistance parameters, onset and offset timing (% gait phase) for extension and flexion assistance for each mode (stair ascent/descent, incline/decline, and level-ground). “~” indicates not applicable.

	Stair-ascent	Incline	Level-ground	Decline	Stair-descent
Extension onset (% gait phase)	3	3	3	0	0
Extension offset (% gait phase)	38	43	35	40	35
Ramp timeext (ms)	150	120	150	120	120
kext,max (Nm·deg.−1)	1.6	1.2	0.3	0	0
θe,ext (°)	0	0	0	0	0
bext,max (Nm·s·deg.−1)	0.003	0.003	0.003	0.12	0.16
Flexion onset (% gait phase)	50	50	50	~	~
Flexion offset (% gait phase)	67	67	67	~	~
Ramp timeflex (ms)	150	120	150	~	~
kflex,max (Nm·deg.−1)	0.7	0.6	0.4	~	~
θe,flex (°)	65	65	65	~	~
bflex,max (Nm·s·deg.−1)	0.003	0.003	0.003	~	~

For the mid-level controllers during stair ascent, incline, and level-ground modes, extension torque assistance during the early stance phase was designed as a stiffness controller, which is capable of generating torque assistance in the form of either positive or negative power depending on the direction of the motion relative to the direction of the exoskeleton torque. Our previous study (43) comparing the biomechanical effects of three assistance strategies (timing-based controller, stiffness controller, and proportional myoelectric controller) showed that the stiffness controller was the most preferred controller by the user, while there was no significant difference in key human biomechanical outcomes (energetics, muscle activity, and joint kinetics) between controllers. This prior work provided another motivation in favor of stiffness control for stair ascent and incline walking. Therefore, we chose the stiffness assistance paradigm for stair ascent and incline for this study, but in theory, any reasonable mid-level controller could have been used with our AI system. During incline and stair ascent modes, the equilibrium angle was set to 0° for the early stance phase, and the stiffness was linearly scaled from 0.3 to 1.2 Nm·deg.⁻¹ for 0° to 18.5° and 1.2 to 1.6 Nm·deg.⁻¹ for 18.5° to 40° (see Table 1). For level-ground mode, the assistance parameters were held constant regardless of the slope estimate. The onset timing of early stance torque assistance was at 3% of the gait cycle, and the offset timing for the extension torque was 35% of the gait cycle for the level-ground mode and 43% of the gait cycle for the incline mode. The onset and offset timings for the level-ground early stance extension assistance were based on the previously determined human-in-the-loop optimized assistance parameter set for the level-ground controller by Frank et al. (17); however, the offset timing for this study was set to a point later than the timing shown in their work based on our pilot work in which users preferred a longer duration of the extension torque assistance during the early stance phase. Knee flexion assistance starts before toe-off and ends during swing flexion under the stiffness controller. The stiffness during flexion assistance is linearly scaled from 0.4 to 0.6 Nm·deg.-1 for 0° to 18.5° and 0.6 Nm·deg.-1 to 0.7 Nm·deg.-1 for 18.5° to 40° (see Table 1). The onset and offset timings for flexion assistance were identical across modes under the stiffness controller. The damping coefficient remained at 0.003 Nm·s·deg.⁻¹ to smooth the torque output when the assistance was active under the stiffness controller.

During the early stance phase assistance, only extension assistance was translated to active torque command, and flexion torque command was forced to be 0 Nm across modes. On the other hand, the extension torque command was set to 0 Nm, and only flexion torque was provided during the swing flexion phase under the stiffness controller. Because of the limitation of the maximum torque output capability of the actuator, the torque commanded was saturated to 15 Nm.

The onset/offset timings of extension and/or flexion assistance were controlled by TBE for the conventional controller and by the AI-based gait phase estimator for the AI-based controller. The TBE determined the user’s current gait phase as the time passed from the most recent heel contact detected by two FSRs, attached to the bottom of the shoes of each leg, divided by the average duration of the most recent five strides stored in a buffer, and the buffer was updated at the completion of each stride.

Data analysis

Model performance

The ground-truth for the walking mode (stair ascent/descent, incline/decline, and level-ground) was determined using the same thresholds used for exoskeleton control (±3.5° and ±18.5°). Accuracy assessment for mode estimation involved both continuous estimates and estimates at instances of exoskeleton mode updates over the entire walking period. The slope estimate was directly compared to the ground-truth slope profile for each testing condition, encompassing both treadmill and outdoor settings. The changes in the slope and mode for the ground-truth labeling between each mode transition for the outdoor testing circuit were done when the user is fully in the new slope or mode (both feet are in the mode). For evaluation, the transition step was defined as the entering and exiting steps into and out of the mode. The ground truth of the gait phase was calculated by linearly scaling the time between the start and the end of each gait cycle detected by the FSRs from 0 to 100%. The gait phase error was first computed in the difference in the angle in the polar coordinate domain between the ground truth and the estimate, and then the difference in angle was converted to a percent scale by dividing the angle by 2π. To compare the gait phase estimation performance within the same walking scenario, the performance of the TBE was computed during the AI-based assistance walking condition to be compared against the AI-based estimator. The RMSE between the estimated value and the ground truth was computed to evaluate the performance of the estimator.

Metabolic cost

To determine the metabolic cost of the user during incline treadmill walking (W/kg), the user’s oxygen consumption and carbon dioxide production were measured using indirect calorimetry (TrueOne 2400Parvo, Parvo Medics, USA). The metabolic cost was estimated on the basis of the modified Brockway equation using measured oxygen consumption and carbon dioxide production (47). The resting metabolic cost, measured while the user is standing and wearing the exoskeleton, was subtracted from the user’s metabolic cost measured during walking conditions to determine the net metabolic cost for each walking condition. The net metabolic cost of the last 3-min during the incline treadmill walking was averaged for each exoskeleton condition for comparison.

Statistical analysis

A paired t test was used to test the statistical significance (P < 0.05) in the difference in the (i) performance of the gait phase estimation between the TBE and our AI-based estimator during validation I and validation II) preference comparison between the two assistance conditions during experiment II. One-way repeated measures analysis of variance (ANOVA) test with post hoc Benferroni correction test (P < 0.05) was used to test the statistical significance in the human outcomes between the walking conditions (unpowered, conventional assistance, and AI-based assistance). All statistical testing was performed in MATLAB (MathWorks, Natick, MA). The quantitative data in the manuscript are presented in means ± SD.

Supplementary Materials

This PDF file includes:

Supplementary Text

Figs. S1 to S3

Tables S1 to S10

Legend for movie S1

References

Other Supplementary Material for this manuscript includes the following:

Movie S1