Online Adaptation Framework Enables Personalization of Exoskeleton Assistance During Locomotion in Patients Affected by Stroke

Abstract

Robotic exoskeletons can transform mobility for individuals with lower-limb disabilities. However, their widespread adoption is limited by controller degradation caused by varying gait dynamics across different users and environments. Here, we propose an online adaptation framework that leverages real-time data streams to continuously update the user state estimator model. This approach allows the exoskeleton to learn the user-specific gait patterns, effectively customizing the model for each new user. Additionally, we demonstrate a sensor signal transformation technique that enables model transfer across different exoskeleton hardware (from a research-grade exoskeleton to a commercial device). With less than one minute of adaptation, our framework improved gait phase estimation, which directly affects assistance timing, by 40.9% for able-bodied subjects and 65.9% for stroke survivors (p<0.05), and reduced torque profile error by 32.7% compared to the baseline model (p<0.05). Furthermore, in a pilot test, we applied our adaptation framework with human-in-the-loop optimization for control tuning. In a single stroke survivor, this approach led to a 21.8% increase in walking speed and a 6.5% reduction in metabolic cost compared to walking without exoskeleton. While preliminary, these results suggest the potential for personalized exoskeleton assistance in clinical populations.

I. Introduction

IT is the roaring twenties in the field of robotic exoskeletons [1], [2]. The rate of advancement in exoskeleton technology is rapidly increasing each year, expanding the field to bring a societal impact in different sectors ranging from human augmentation enhancing physical capabilities [3]-[7] to assisting individuals with mobility impairments [8]-[12]. The past two decades of research in the field have focused on the fundamental understanding of a robust mechanical design [13]-[18] and an effective control strategy [19]-[22] for targeted applications. Recent work is culminating these findings and developing exoskeleton systems that function well outside of the lab [23]. However, translating such technology still faces immense challenges: limited guarantees in exoskeleton performance across new users, environments, and even hardware, and these challenges greatly diminish the confidence in deploying exoskeletons at a large scale into the everyday fabric of society.

Conventional exoskeleton control strategies generate joint torques that are in line with the user’s underlying biological joint demand during locomotion [24]-[29]. This control approach offloads a user’s required joint mechanical power such as an extension movement at the hip joint. Accurate and effective exoskeleton assistance often leads to meaningful improvements in human outcome measures such as reducing metabolic cost [30]-[32] and increasing walking speed [33]. However, achieving decent human-exoskeleton performance is non-trivial as an ‘accurate’ assistance level is governed by different control parameters, such as timing [34], [35] and magnitude [32], [36], which are sensitive to both the user (e.g., able-bodied vs. patient populations) and task demands (e.g., different speeds of walking or ramps/stairs locomotion). Consistently in the literature, exoskeleton studies have illustrated that there is an optimal level of assistance (especially assistance timing) that maximizes human-exoskeleton performance [19], [37], indicating a need for a robust control framework in capturing the user’s state information in real-time.

Recent breakthroughs in deep learning [38] have enabled improved estimation of the user’s state information in real-time, enabling robust control of these exoskeleton systems during multi-speed locomotion [39]. A robust deep learning-based user state estimator allows for a concrete hierarchical controller where the exoskeleton assistance can dynamically adjust depending on the estimated user’s state (e.g., locomotion mode), potentially even within a single step. One vital user state variable that directly relates to assistance timing is the user’s gait phase [40], [41]. Gait phase corresponds to the user’s joint configuration throughout the gait cycle, typically represented as a linearly increasing function from 0% to 100% throughout one gait cycle (e.g., from one heel strike to the next). Gait phase-based exoskeleton control allows for a time-independent assistance paradigm, enabling a robust assistance strategy independent of the user’s dynamic changes in cadence [41], [42].

Previously, several groups have explored various techniques to develop a robust gait phase estimator [43]-[45]. The most simple approach is time-based estimation, which extracts time since the most recent gait event (e.g., heel contact using a foot switch) divided by the average stride duration from the previous number of strides [27]. Another notable method in estimating gait phase is using an adaptive oscillator, which leverages the sinusoidal nature of hip joint angle during locomotion [43]. Similarly, a holonomic phase variable can be used to represent gait phase by utilizing the user’s hip joint or thigh-limb kinematics [44]. While widely used, these heuristic methods face challenges in real-world deployment, including limited adaptability to multi-mode locomotion, poor performance during abrupt transitions, and the need for manual tuning for new users, an issue less significant in deep learning models [46]. Regardless, it is evident from the increase of literature studies regarding this topic that a robust, multi-mode, user-independent gait phase estimator is needed.

Generally, to train a deep learning-based model that is intended to be deployed to a new user, a dataset consisting of many exoskeleton users (generally from able-bodied individuals) in different environmental contexts is required [39], [47]. While this trained model may have adequate performance in a lab setting, this training pipeline has a fundamental limitation of assuming that the pre-trained model will retain its performance in a new setting (both user and environment), despite the shift in the data distribution. Furthermore, most results from existing literature on deep learning-based user state estimators are heavily skewed towards young and able-bodied individuals, which do not fully represent potential exoskeleton users, such as stroke survivors. To mitigate this limitation, we hypothesized that adapting the deep learning-based model to the user- and condition-specific gait dynamics will further reduce the user state estimation error by personalizing the underlying model parameters. Here, we present a unified online adaptation framework that actively learns the new user’s inherent gait dynamics in real time. In this framework, the pre-trained deep learning-based gait phase estimator automatically adjusts model parameters by leveraging real-time streaming sensor data. Three key innovative features make our framework unique: the ability to translate across different exoskeleton hardware, locomotion settings, and user populations (i.e., stroke gait).

Our work contributes to the field of robotic exoskeletons by improving gait phase estimation and developing an adaptation framework for stroke gait. Specifically, we present: 1) a deep learning-based gait phase estimation system, 2) an online adaptation framework tailored to stroke gait, and 3) a parallel processing architecture for real-time inference. Lastly, we present a proof-of-concept pilot stroke experiment demonstrating how our adaptation framework may impact stroke biomechanical outcomes. Our study highlights the framework’s ability to adjust to stroke gait characteristics, supporting its feasibility within this population. By integrating this adaptation mechanism, our system lays the groundwork for future research on personalized assistive technology in clinical applications.

II. Robotic Hip Exoskeleton

A. Robotic Hip Exoskeleton Design

An autonomous bilateral robotic hip exoskeleton was developed to assist the user’s hip flexion/extension in the sagittal plane (Fig. 1A). An electromechanical actuator (AK80-9, T-motor, China) consisting of a brushless DC motor and a single-stage 9:1 planetary gearhead was integrated at each hip joint to provide joint torque. The low gear ratio allowed the actuator to operate in a quasi-direct drive mode, providing a mechanical feature with two key advantages: high back-drivability and efficiency in torque transmission. The integrated actuator was able to provide up to 18 Nm peak torque, which was approximately 40% of the peak biological hip joint moment during level-ground walking for an average adult male in the United States [48]. The actuator’s peak angular velocity was 25.6 rad/s, which captured the entire range of the hip joint angular velocity across various locomotion modes [48]. The exoskeleton allowed for a wide range of hip joint movement (37° extension and 127° flexion), covering the full range of the hip joint angle during level-ground walking. Additionally, we installed a passive hinge joint, located under the actuator output to allow for hip abduction/adduction. For the user interface, a custom width-adjustable thermoplastic lumbosacral orthosis was used. Two actuator units were attached to this orthosis bilaterally where a height-adjustable aluminum strut was used to align the actuator to the user’s biological hip joint. Lastly, custom carbon fiber thigh interfaces were designed to reorient the actuator output torque into a perpendicular force on the anterior part of the user’s thigh segment. The overall mass of our exoskeleton system was 4.5 kg including electronics (1.4 kg).

Fig. 1.

(A) Autonomous robotic hip exoskeleton designed to assist the user’s hip flexion and extension during locomotion. (B) System architecture of the robotic hip exoskeleton. A microprocessor communicates with sensors and actuators, while an additional coprocessor executes real-time inference and online adaptation. (C) Implementation of the biological torque controller. The torque command profile is generated using gait phase input through a PCHIP spline, with predefined timing and magnitude parameters.

B. Electronic Components and System Architecture

A control unit was mounted posteriorly to the user’s pelvis interface and consisted of a 24V 3600mAh lithium polymer battery (Venom Power, USA) which powered the entire system including actuators and sensors. An onboard microprocessor (myRIO 1900, National Instruments, USA) served as the main host unit (Fig. 1B). An additional Jetson Nano co-processor was integrated for real-time inference and online adaptation of our model using data from onboard sensors. To design the exoskeleton sensor suite, we conducted a literature review of several state-of-the-art robotic hip exoskeletons that were currently available in the field [13], [15], [17], [18], [20], [31]. These systems commonly feature joint encoders and inertial measurement units (IMUs), which focus on capturing the user’s joint and limb kinematics during locomotion. We integrated a single 6-axis IMU (MPU-9250, InvenSense, USA) near the user’s sacrum as a proxy to capture the user’s pelvis kinematics. In addition, although not used in this study, the exoskeleton had two additional IMUs located at each thigh interface. A 12-bit incremental rotary encoder was also installed on each actuator to capture the user’s hip joint angle and angular velocity. The host microprocessor was equipped with a field programmable gate array (FPGA), which allowed the system to communicate with different sensors without overloading the computational power. IMUs were communicated directly via SPI communication. For the actuator, an SPI-CAN converter (MCP 2515), which included a CAN transceiver (TJA 1050), was used to convert data packets to the motor driver. The sensor data acquisition and torque command were updated at 200 Hz in the main control loop while the motor driver executed the targeted torque at 1 kHz. The real-time sensor data were streamed to the high-level coprocessor at 200 Hz via a standard ethernet cable using a TCP/IP protocol for real-time inference and adaptation. The overall exoskeleton controller had a graphical interface on a host laptop that communicated with the microprocessor via Wi-Fi.

C. Exoskeleton Control Strategy

A three-tier hierarchical control architecture was implemented to control the entire exoskeleton system. The high-level layer was implemented to infer the user and walking states (in this study, gait phase) that understands the user’s movement and surrounding environments. The mid-level layer governed the device’s dynamic performance using physical control laws, generating assistance envelopes throughout the gait cycle. The low-level layer ensured that the reference torque trajectory was matched using conventional closed-loop current feedback control.

We utilized a biological torque controller in the mid-level layer in the exoskeleton that commands an assistance profile as a function of the gait phase (Fig. 1C). This controller generated smooth and continuous hip joint torque envelopes during the phase in line with the human biological joint demand. To generate the assistance profile, we used a Piecewise Cubic Hermite Interpolating Polynomial spline function dependent on gait phase where the nodes of this curve represented different control parameters, such as assistance timing and magnitude. For the timing parameters, we used values shown to have the largest metabolic cost benefit in able-bodied subjects from the literature [27] with additional fine-tuning from a separate pilot experiment. During real-time exoskeleton control, the biological torque controller employed the estimated gait phase to generate a target torque, and safety measures have been implemented to ensure that the provided assistance was always in the intended direction of the user’s movement. Specifically, to prevent estimation reversals that could lead to incorrect torque generation, we ensured that the gait phase estimate always progressed in a monotonically increasing manner, maintaining the previous timestep’s value until a new estimate in the correct direction was obtained.

III. Gait Phase Estimation

A. Deep Learning-Based Gait Phase Estimator

We designed a deep learning-based user-independent gait phase estimator based on our previous work [46]. This gait phase model was trained using a dataset from 10 able-bodied subjects (a different group from this study) navigating through 5 different locomotion modes (level-ground, ramp and stair ascent/descent). The fully trained model estimated the user’s current gait phase based on kinematic input from onboard sensors. Input sensor data included joint angle and angular velocity from bilateral hip encoders and 3-axis linear acceleration and gyroscopic data from the IMU. While a standard representation of gait phase utilizes heel contact (0%) as a deterministic gait event, we used the maximum hip extension position as the start of the gait cycle, which roughly corresponds to a toe-off event. This approach was specifically selected as we wanted to only leverage sensors native to the hip exoskeleton system (e.g., heel contact detection requiring a foot switch). Using a local peak detection method, we linearly interpolated local hip extension points from 0% to 100%. Recorded sensor data were processed post hoc to reconstruct ground truth (i.e., gait phase percentage) using a local peak detection algorithm on the user’s hip joint angle. Due to the piecewise nature of gait phase (i.e., 100% gait phase is equivalent to 0%), direct usage of gait phase in percentage representation can often lead to poor performance in neural network training. To eliminate this inherent discontinuity in the signal, we projected the processed gait phase onto a unit circle in Cartesian coordinates [41].

B. Deep Convolutional Neural Network

A convolutional neural network is commonly used in wearable robotics to estimate user states. Instead of manual feature engineering, the neural network automatically extracts features from onboard sensor data streamed as time series. The convolutional layers stride over defined data windows to identify relevant patterns, which are then used by the rest of the network to map input features to output states, such as gait phase. We systematically optimized the network architecture and hyperparameters using 10-fold leave-one-subject-out validation. Each trained neural network was evaluated across multiple settings (e.g., accuracy across modes) to ensure that the overall performance was not skewed towards a specific locomotion mode. While the originally presented gait phase model in [46] was a single network with two outputs (left and right gait phase), we adopted a unilateral approach due to the complications when deploying the model to the online adaptation framework (i.e., backward labeling). However, each model leveraged full bilateral sensor data as a fusion of the contralateral leg’s sensor information can further improve the overall gait phase estimation performance [49].

To train the network, we used the mean absolute error of the angular similarity metric (since each gait phase percentage was converted to x and y coordinates on a unit circle) over the gait cycle as the baseline loss function. The range of candidate values for each hyperparameter was selected based on a preliminary experiment evaluating the sensitivity of each hyperparameter on the overall network performance. The input window size, number of convolutional layers, number of filters, kernel size, dropout probability, activation function for convolution and dense layers, optimizer, and learning rate were optimized during this process. With the given hyperparameter candidates, the gait phase models were trained with a batch size of 128, iterating over a maximum of 200 epochs with early stopping and patience of 10. The final optimized network had an input window size of 80 (i.e., 400 ms) and consisted of two 1D convolutional, each with 10 filters and a kernel size of 3, followed by 20% dropout and sigmoid activation. The final dense layer reduced the flattened output to 21 nodes before producing 2 units (x, y coordinates) with tanh activation. The final model resulted in an average gait phase estimation root mean squared error (RMSE) of 7.45±1.84% across all subjects and locomotion modes.

IV. Online Adaptation Framework

A. Online Adaptation Algorithm Development

Accurate user state estimation, such as gait phase, is paramount for maintaining optimal assistance levels for robotic exoskeletons. Deep learning approaches have been widely adopted in wearable robotics for their robustness and the ability to learn the latent representation to map raw sensor data to the desired gait phase. However, the deep learning method has the drawback of requiring a significant amount of human subject data which can be expensive, time consuming, and exhausting (e.g., collecting data in patient populations). Moreover, there is a limitation in the training dataset’s (typically ranging from 10 to 20 subjects) inability to capture the entire data distribution in gait dynamics across all potential exoskeleton users. To mitigate this, our previous study explored a strategic method to improve the deep learning model performance with minimal additional data using transfer learning [41]. Extension of this work led to the development of the online adaptation framework where we were able to fine-tune the baseline user-independent model in an online setting, enabling the exoskeleton system to learn the user- and condition-specific gait dynamics from a new user.

The crux of our proposed online adaptation framework is active and online learning using live-streamed exoskeleton sensor data during locomotion (Fig. 2B). A pre-trained user-independent gait phase estimator was deployed to the Nvidia Jetson coprocessor. This baseline model ran real-time inference at 200 Hz, equivalent to the frame rate that assistance torque profiles were generated. In parallel, the live-streamed sensor data were queued into a fixed-sized data buffer. Every 5 seconds, a backward labeler relabeled the ground truth gait phase using the exoskeleton’s hip joint angle. Afterwards, labeled sensor data were used to update the gait phase model weights via one epoch backpropagation training. To ensure that there was no discontinuity in real-time gait phase estimation during the online learning process, the real-time inference and online adaptation modules operated in parallel using built-in multiprocessing features available in the Nvidia Jetson coprocessor. Finally, at the end of each adaptation cycle, the updated gait phase model was passed from the adaptation module to the inference module. Using this framework, the system dynamically learned the user’s gait pattern and retained high estimation performance agnostic of the user’s walking conditions (e.g., changes in the walking speed).

Fig. 2.

Overall strategy of applying an online adaptation framework on robotic exoskeletons. (A) The exoskeleton operated under a three-tier control scheme where the high-level layer incorporated a deep learning-based gait phase estimator. Here, an online adaptation framework was integrated where the model weights were fine-tuned to the user in real time. The mid-level layer generated torque commands based on gait phase. In this layer, control parameters can further be optimized for the user using human-in-the-loop optimization. The low-level layer ensured that the desired joint torque was matched via standard feedback control. (B) The online adaptation framework utilized a baseline user-independent gait phase estimator pre-trained in an offline setting. During locomotion, an inference module estimated the user’s gait phase in real-time. Asynchronously, an adaptation module used a batch of real-time sensor data and retroactively relabeled the ground truth gait phase to update weights with one epoch training. (C) The pre-trained model can be transferred to different exoskeleton hardware using the sensor data transformation technique. This single instance update process leveraged a small amount of sensor dataset from two devices to transform the new sensor data to the previous exoskeleton’s data form.

B. Backwards Ground Truth Labeling

During real-time inference, a new set of sensor data from onboard sensors was streamed into a data queue buffer (buffer size greater than the model’s input window size). During this process, a potential signal loss can generate anomalies in data, which can cause biased learning, leading to model overfitting. To ensure zero data drop, we employed a universal asynchronous receiver-transmitter communication protocol via the internet protocol suite over a standard ethernet cable between the high-level coprocessor and the exoskeleton microprocessor. From here, we systematically swept the data buffer size that guarantees reliable model training while not causing significant overhead in memory storage.

When the data buffer became full, we retroactively generated ground truth labels using an autonomous backward labeler. We utilized a local peak detector on the user’s hip joint angle for ground truth labeling. To ensure robust peak extraction, we heuristically tuned the detector’s parameters, such as prominence, based on the data distribution exhibited in our locomotion dataset. However, detecting local peaks in a patient population can be challenging, as there may be multiple fluctuations during a single gait cycle. To mitigate potential errors, we restricted the range in which we search for local maximum values on the paretic side by analyzing the segmented data of the contralateral leg, as the peak detection performance of the non-paretic leg is comparable to that of an able-bodied individual (Fig. 3). By utilizing this segmented gait cycle data, the search for local peaks on the patient’s paretic side became significantly more reliable. Furthermore, in the case of stroke gait, these unreliable peaks tend to occur during the swing phase of the gait cycle. Therefore, we found that locating the first local peak on the paretic side following the non-paretic side’s toe-off event was a robust approach for this population. A key phenomenon to note during ground truth labeling was the limitation in labeling gait phase on data from the last gait event to the end of the buffer (cannot interpolate properly). To minimize data loss, this segment was loaded to the following buffer and utilized in the next adaptation cycle. The original gait phase model from our previous work [46] was a single network with four outputs (left and right gait phase in x and y Cartesian coordinates). To implement the online adaptation framework to this architecture, a bilateral ground truth label was required. However, due to the required labeling technique in the last part of the buffer, bilateral labeling was impractical due to the asynchronous nature of gait events across both legs causing significant loss in data (truncation to match the overall input data shape). To bypass this problem, we applied a unilateral gait phase model for each leg with minimal performance degradation compared to the model with bilateral architecture.

Fig. 3.

Time series graph illustrating the application of a peak detection algorithm in stroke gait. To ensure a robust local peak detection for the paretic side (green), a local maximum peak search range was constrained within the range of the contralateral leg’s gait cycle (represented with two grey bars). This approach minimized potential errors and increased the reliability of peak detection on the user’s paretic leg.

C. Single Cycle Model Training

At every adaptation cycle (for this study, 5 seconds), we online adapted the baseline gait phase model via single epoch training using this ground truth labeled data (Algorithm 1). To ensure that the model can adapt reliably to a small amount of data, we hyperparameter swept the learning rate and the optimizer based on a preliminary offline experiment using exoskeleton data. One important feature to consider during this online adaptation is to freeze any existing batch normalization layers in the architecture. Generally, adding a normalization layer to the network improves the overall performance as the layer can stabilize the input data via re-centering and rescaling. However, if the layer is not locked during adaptation, the normalization statistics can be changed based on the buffer data, which can lead to performance degradation. After the adaptation cycle, the adapted model was transferred to the real-time inference module.

Algorithm 1 Online Adaptation for User State Estimation

$$\begin{array}{cc}\hfill & 1:\mathrm{Given}:\text{model}\pi ,\text{parameters}{\theta }_0,\text{window}T,\mathtt{adapt}\hfill \\ \hfill & 2:\mathrm{Init}:\mathtt{buf}=\mathtt{zeros}(\mathtt{ch},\mathtt{T})\hfill \\ \hfill & 3:\mathbf{while}\mathtt{True}\mathbf{do}\hfill \\ \hfill & 4:x_t=\mathtt{recv}()▷\text{Robot data}\hfill \\ \hfill & 5:\mathtt{buf}[\mathtt{end}+1]=x_t▷\text{Enqueue data}\hfill \\ \hfill & 6:\mathtt{buf}=\mathtt{buf}[1:\mathtt{end}]\hfill \\ \hfill & 7:y_t={\pi }_{{\theta }_i}(x_t)▷\text{State estimate}\hfill \\ \hfill & 8:\mathtt{send}(y_t)▷\text{Update robot}\hfill \\ \hfill & 9:\mathbf{if}{x}_t\text{has heel strike and}\mathsf{adapt}\mathbf{then}\hfill \\ \hfill & 10:\mathtt{lables}=\mathtt{label}(\mathtt{buf}[t_p:\mathsf{t}])▷\text{Label stride}\hfill \\ \hfill & 11:\mathtt{data}=\mathtt{buf}[t_p:\mathtt{t}]▷\text{Segment stride}\hfill \\ \hfill & 12:{\theta }_{i+1}=\mathtt{train}(\mathtt{data},\mathtt{labels})▷\text{Update model}\hfill \\ \hfill & 13:t_p=t\hfill \\ \hfill & 14:\mathtt{model}.\mathtt{save}()\hfill \\ \hfill & 15:\mathbf{end}\mathbf{if}\hfill \\ \hfill & 16:\mathbf{end}\mathbf{while}\hfill \end{array}$$

D. Adaptation and Inference Parallelization

Speed of inference was a crucial characteristic targeted during this algorithm design. This is necessary because delayed state estimation can result in phase lag in assistance provided by the exoskeleton and thus would result in very poor performance during ambulatory tasks. To combat this, we employed parallel processing to reduce latency as much as possible. We used 3 processes, one for data input and output, one for model inference, and the last for model adaptation. The data input and output thread were singularly responsible for communication with the exoskeleton and would block waiting for the new sensor data which was sent from the robot at a synchronous rate of 200 Hz. The newest data was passed to the inference process which immediately queried the model for the state estimate, taking less than 0.005s (capable of 200 Hz operation) and the estimate was sent back to the main processor. Only when gait events were detected would the third process run, training on the newest strides of data, but since this process was separated from the inference process the real-time inference speed was maintained. Training time took approximately 0.6-0.8 seconds per stride, meaning training time was faster than the necessary 0.2 Hz.

V. Device-to-Device Model Transfer

A. Device Sensor Data Transformation

Different exoskeleton designers often locate and orient sensors differently, which is a common limiting factor for deep learning-based models in the existing paradigm, making them susceptible to shifts in sensor data [47]. Theoretically, sensor data could be transformed from the coordinate system of one device to another based solely on the geometric relationship of sensor orientations and locations between exoskeletons. However in practice, this approach can be inaccurate or infeasible given three primary considerations: 1) the exact sensor placements may be unknown, especially when working with data from commercial devices; 2) the placement of the sensors may be known but unreliable (e.g., if using data from wearable sensors attached directly to the body, which are often placed based on rough estimates of anatomical landmarks); and 3) user fit often changes across devices from imperfect orthotic interfaces, causing geometric measurements of sensor placements relative to an assumed ground frame to have inaccuracies when fit to the user.

Previously, we introduced an algorithm to optimize a transformation between sensors of different exoskeleton that does not need any assumptions about sensor orientation or position a priori [50]. In this study, we leverage this method and demonstrate a zero-shot exoskeleton-to-exoskeleton model transfer (i.e., no retraining of the underlying model) (Fig. 2C). Our approach used the kinematic trajectories of the exoskeleton sensor data collected during two 10-second-long walking trials (one trial per exoskeleton) to optimize a transformation between the sensors from each device. This allowed the sensor calibration from one device to another to be completed in one-shot (i.e., only done one time for a new exoskeleton) and did not require the subject to assume any specific calibration poses aside from their normal gait, maximizing the similarities in the sensor data between the two trials. By applying this transformation to the incoming data stream, the gait phase estimator can be transferred across devices without the need to retrain the model to accommodate the sensor locations of the new device (i.e., without needing to collect a device-specific labeled dataset).

B. IMU Transformation Optimization

To optimize the transformation of IMU data from a new exoskeleton (EXO2) into the coordinate system of an old exoskeleton (EXO1), our optimization approach used data sampled over 10 consecutive strides from each exoskeleton. The resulting EXO1 and EXO2 data were aligned in time post hoc, such that they were of equal length $N$. The accelerometer and gyroscope data measured from the IMUs of EXO1 and EXO2 (i.e., $A_{E1}$, $A_{E2}$, $G_{E1}$, $G_{E2}$) were then extracted from the data as,

$$A=\left[\begin{array}{cccc}\hfill a_{1,1}\hfill & \hfill a_{1,2}\hfill & \hfill \cdots \hfill & \hfill a_{1,N}\hfill \\ \hfill a_{2,1}\hfill & \hfill a_{2,2}\hfill & \hfill \cdots \hfill & \hfill a_{2,N}\hfill \\ \hfill a_{3,1}\hfill & \hfill a_{3,2}\hfill & \hfill \cdots \hfill & \hfill a_{3,N}\hfill \end{array}\right],$$ (1)

$$G=\left[\begin{array}{cccc}\hfill g_{1,1}\hfill & \hfill g_{1,2}\hfill & \hfill \cdots \hfill & \hfill g_{1,N}\hfill \\ \hfill g_{2,1}\hfill & \hfill g_{2,2}\hfill & \hfill \cdots \hfill & \hfill g_{2,N}\hfill \\ \hfill g_{3,1}\hfill & \hfill g_{3,2}\hfill & \hfill \cdots \hfill & \hfill g_{3,N}\hfill \end{array}\right],$$ (2)

such that the rows of $A$ and $G$ represented the $x\text{-}$, $y\text{-}$, and $z$-axes of the sensors, respectively. Since the relative change in orientation and position between the IMUs of EXO1 and EXO2 were unknown due to differences in user fit between the devices, we optimized a rotation matrix ${}^{E1}{\mathit{R}}_{E2}^{\ast }$ and position vector ${}^{E1}{\mathit{p}}_{E2E1}^{\ast }$, further denoted as ${\mathit{R}}^{\ast }$ and ${\mathit{p}}^{\ast }$, to transform the IMU data measured by EXO2 ($A_{E2}$ and $G_{E2}$) into the EXO1 coordinate system (${\widehat{A}}_{E1}$ and ${\widehat{G}}_{E1}$). Specifically, we defined the vector $\mathit{Z}\in {\mathbb{R}}^{6\times 1}$, comprised of three Euler angles and the position vector ${\mathit{p}}^{\ast }$,

$$\mathit{Z}=\left[\begin{array}{c}\hfill \widehat{\theta }\hfill \\ \hfill \widehat{\varphi }\hfill \\ \hfill \widehat{\psi }\hfill \\ \hfill \widehat{\mathit{p}}\hfill \end{array}\right],$$ (3)

such that the rotation matrix is

$$\widehat{\mathit{R}}(\widehat{\theta },\widehat{\varphi },\widehat{\psi })={\mathit{R}}_z(\widehat{\psi }){\mathit{R}}_y(\widehat{\varphi }){\mathit{R}}_x(\widehat{\theta }),$$ (4)

The optimized vector ${\mathit{Z}}^{\ast }$ was computed as

$${\mathit{Z}}^{\ast }=\left[\begin{array}{c}\hfill {\widehat{\theta }}^{\ast }\hfill \\ \hfill {\widehat{\varphi }}^{\ast }\hfill \\ \hfill {\widehat{\psi }}^{\ast }\hfill \\ \hfill {\widehat{\mathit{p}}}^{\ast }\hfill \end{array}\right]=\text{arg}\underset{\widehat{\theta },\widehat{\varphi },\widehat{\psi },\widehat{\mathit{p}}}{\text{min}}\left(\sum \tilde{G}(\widehat{\theta },\widehat{\varphi },\widehat{\psi })+\sum \tilde{A}(\widehat{\theta },\widehat{\varphi },\widehat{\psi },\widehat{\mathit{p}})\right),$$ (5)

where $\tilde{G}$ and $\tilde{A}$ represent the mean squared error between the accelerometer and gyroscope data of EXO1 and the transformed data of EXO2. Thus,

$$\tilde{G}=\left[\begin{array}{c}\hfill {\tilde{g}}_1\hfill \\ \hfill {\tilde{g}}_2\hfill \\ \hfill {\tilde{g}}_3\hfill \end{array}\right],$$ (6)

where each element ${\tilde{g}}_i$ of $\tilde{G}$ was computed as

$${\tilde{g}}_1=\frac{1}{N}\sum _{j=1}^N{(g_{i,j}-{\widehat{g}}_{i,j})}^2\forall g_{i,j}\in G_{E1,i},{\widehat{g}}_{i,j}\in {\widehat{G}}_{E1,i},$$ (7)

and

$$\tilde{A}=\left[\begin{array}{c}\hfill {\tilde{a}}_1\hfill \\ \hfill {\tilde{a}}_2\hfill \\ \hfill {\tilde{a}}_3\hfill \end{array}\right],$$ (8)

where each element ${\tilde{a}}_i$ of $\tilde{A}$ was computed as

$${\tilde{a}}_i=\frac{1}{N}\sum _{j=1}^N{(a_{i,j}-{\widehat{a}}_{i,j})}^2\forall a_{i,j}\in A_{E1,i},{\widehat{a}}_{i,j}\in {\widehat{A}}_{E1,i}.$$ (9)

We assumed that EXO1 and EXO2 were fixed to the user’s pelvis, allowing ${\widehat{G}}_{E1}$ and ${\widehat{A}}_{E1}$ to be computed from rigid body kinematics as

$${\widehat{G}}_{E1}=\widehat{\mathit{R}}{G}_{E2},$$ (10)

$${\widehat{A}}_{E1}=\widehat{\mathit{R}}{A}_{E2}+({\widehat{\alpha }}_{E1}\times \widehat{\mathit{p}})+({\widehat{G}}_{E1}\times ({\widehat{G}}_{E1}\times \widehat{\mathit{p}})),$$ (11)

where ${\widehat{\alpha }}_{E1}$ is the angular acceleration of the EXO2 IMU computed in the EXO1 coordinate system as

$${\widehat{\alpha }}_{E1}=\frac{d}{dt}{\widehat{G}}_{E1}.$$ (12)

Thus, after optimizing ${\widehat{\mathit{R}}}^{\ast }$ and ${\widehat{\mathit{p}}}^{\ast }$, incoming IMU data from EXO2 could be quickly transformed into the EXO1 coordinate system using (10) and (11) before being input to the neural network to estimate the user’s gait phase.

C. IMU Transformation Accuracy Evaluation

To evaluate the accuracy of our IMU transformation, two able-bodied subjects (2 males, height of 178 and 183 cm, body mass of 70 and 78 kg, and age of 30 and 28 years) participated in a validation trial after providing informed consent to the protocol approved by the Georgia Institute of Technology Institutional Review Board. Each subject walked on a treadmill at 1.3 m/s while wearing two different exoskeletons, Gait Enhancing and Motivating System (GEMS, Samsung Electronics, South Korea) hip exoskeleton (EXO1) and our custom robotic hip exoskeleton (EXO2). To promote similar lower-limb kinematics between the two conditions, the exoskeletons did not provide assistance during the trials and the subjects’ step frequency was constrained to 110 bpm by using a metronome. During each walking bout, the exoskeleton recorded right and left hip joint angles and velocities using actuator-mounted encoders and accelerometer and gyroscope data from a pelvis-mounted IMU at 200 Hz. Using the data from each subject, 10 consecutive strides of data collected from EXO1 and EXO2 were aligned in time based on the right leg’s peak hip extension joint position. The data were also truncated such that the 10 strides of data from EXO1 and EXO2 were of equal length. Finally, the resulting pelvis-mounted IMU data were filtered using a zero-lag 5th order lowpass filter with a 20 Hz cutoff frequency to remove differences in high frequency noise between the two sensors.

Using the time-aligned data from each subject, a unique pelvis IMU transformation (i.e., ${\widehat{R}}^{\ast }$ and ${\widehat{p}}^{\ast }$) was optimized. We ran the optimization using the constrained optimization function, fmincon (Matlab 2019b, Mathworks, MA). Using the resulting transformation from Subject 1, the RMSE between the IMU data of EXO1 and EXO2 was computed under three conditions: 1) without transforming the EXO2 data, 2) by transforming the EXO2 data into the EXO1 coordinate system using the optimized transformation from the same subject, and 3) by transforming the EXO2 data into the EXO1 coordinate system using the optimized transformation from the alternative subject. This analysis was then repeated using the transformation computed from Subject 2’s data.

The stride-averaged results of the pelvis IMU transformation for Subjects 1 and 2 (Fig. 4) indicated that implementing the pelvis IMU transformation reduced the reconstruction error for the accelerometer and gyroscope EXO2 data by 6.53 m/s2 (82.6%) and 0.15 rad/s (40.6%), respectively, with respect to the EXO1 data (Fig. 5). Additionally, the reconstruction error of the transformed IMU data when optimized on a different subject was similar to that of optimizing the transform on the same subject, with an average increase in the accelerometer reconstruction error of 0.17 m/s2 and a decrease in the gyroscope reconstruction error of 2×10-4 rad/s, respectively. In general, optimizing the rotation matrix and position vector to transform the pelvis IMU from EXO2 into the EXO1 coordinate system substantially improved reconstruction error for the accelerometer and gyroscope data. Further, we found that applying an optimized transformation to data collected on a new subject had little impact on the overall reconstruction error.

Fig. 4.

Stride-averaged pelvis IMU data from two subjects. (Red) EXO2 data before transforming the data. EXO2 data after realignment using the optimized transformation from each subject’s data (blue). EXO2 data after realignment using the optimized transformation from the alternate subject’s data (purple). Pelvis-mounted IMU data from the subject wearing EXO1 (green). The shaded regions represent±1 standard deviation.

Fig. 5.

Accelerometer and gyroscope reconstruction RMSE. The resulting RMSE for the (A) accelerometer and (B) gyroscope data is shown relative to the EXO1 data without a transformation (red), with a transformation optimized using the same data (blue), and with a transformation optimized using the data from the alternate subject (purple).

D. Online Gait Phase Estimation Using Sensor Transformation

To analyze our strategy for transferring the gait phase estimator to a new hip exoskeleton, three additional subjects (2 males and 1 female, mean height of 177.8±4.4 cm, mean body mass of 67.3±10.8 kg, and mean age of 20.3±1.2 years) walked in EXO2 on the treadmill at 1.2 m/s for two minutes after providing informed consent. The exoskeleton provided sagittal hip assistance based on a predefined spline profile, which computed the desired exoskeleton torque as a function of instantaneous gait phase with a peak assistance of 7 Nm (flexion and extension). The user’s gait phase was computed online using the baseline gait phase estimator trained using data from EXO1 implemented on a separate onboard Nvidia Jetson coprocessor; however, since the neural network was trained on EXO1 data but deployed on EXO2 data, the incoming stream of pelvis accelerometer and gyroscope data was first transformed into the EXO1 coordinate system before being input to the neural network using (10) and (11) with the rotation matrix and position vector optimized from Subject 1.

After the experimental protocol was completed, the gait phase estimator was then implemented offline on the recorded data without applying the optimized transformation to the IMU data, which served as a baseline condition to compare the importance of the optimized transformation. We did not test the gait phase estimator online without the transformation since it would have provided nonsensical torque assistance, which could be uncomfortable or dangerous for the user and would likely propagate gait phase estimator error far beyond the purely mathematical error resulting from an incorrect transform. The error resulting from the online, transformed and offline, untransformed gait phase estimator implementations was then computed as the RMSE between the ground truth and estimated gait phase over the two-minute trial.

The resulting RMSE of the online gait phase estimator when implemented on a new exoskeleton was 2.65% (Fig. 6). This resulted in a 76.8% reduction in gait phase RMSE relative to implementing the neural network on EXO2 without transforming the IMU data. We found that the gait phase estimation error, when deployed on the new exoskeleton (EXO2), was comparable to the results of our previous study [46], which deployed the gait phase estimator on the same device from which it was trained (EXO1). Thus, our results demonstrated that by optimizing the IMU transformation using only 10 strides of unlabeled data from each exoskeleton, the convolutional neural network was able to maintain model accuracy, even when implemented on a new never-before-seen exoskeleton with a different fit, IMU model, and IMU placement compared to the initial device. This work presented a first in deep learning-enabled exoskeleton control, demonstrating the possibility to transfer models across devices without the need to recollect any labeled training data for the model, greatly decreasing the barrier-to-entry for enabling deep learning-based control on new exoskeleton devices.

Fig. 6.

Resulting gait phase estimation RMSE. The gait phase estimation RMSE is shown when implementing the gait phase estimator on the new exoskeleton. Online gait phase estimation was implemented by transforming the pelvis-mounted IMU in-the-loop before inputting the data into the gait phase estimator. An offline post hoc analysis was also conducted to compute the resulting RMSE without implementing the IMU transformation. The error bars represent ±1 standard deviation.

VI. Human Subject Experiment

A. Experimental Protocol

We recruited eight able-bodied individuals (5 males and 3 females) with a mean age of 27.2±4.1 years, body mass of 70.46±7.45 kg, and height of 173.6±7.0 cm. We additionally recruited six individuals (4 males and 2 females) with chronic stroke (¿ 6 months post-stroke) with a mean age of 56.5±14.3 years, body mass of 74.7±13.0 kg, and height of 172.5±9.8 cm (inclusion/exclusion criteria and subject’s baseline physical capabilities reported in the Appendix). The study was approved by the Georgia Institute of Technology Institutional Review Board and informed written consent was obtained for all participants. The study included three experimental protocols to validate the performance of our online adaptation framework: 1) eight able-bodied individuals walking on a treadmill, 2) three able-bodied individuals walking on a split-belt treadmill, and 3) six stroke survivors walking on a treadmill. The primary objectives of these protocols are as follows: Experiment 1 evaluates the adaptation results in an able-bodied population, Experiment 2 examines the adaptation framework in able-bodied individuals with simulated asymmetric gait (replicating stroke gait), and Experiment 3 assesses adaptation in actual stroke gait. These three experiments form the main validation process, following a logical progression from able-bodied to stroke subjects incrementally.

For Experiment 1, each subject was asked to walk on a standard treadmill (Tuff Tread, TX) at a walking speed of 1.2 m/s while wearing our robotic hip exoskeleton (Fig. 7A). The exoskeleton controller utilized the baseline model for gait phase estimation (detailed information on the data used to train the baseline model, as well as its performance, is presented in the Appendix). During walking, the exoskeleton provided bilateral hip flexion and extension assistance using our biological torque controller. As subjects acclimated to the exoskeleton, we gradually increased the assistance magnitude, aiming for the maximum level without causing discomfort (e.g., mechanical play around the user interface causing the actuator to misalign with the user’s hip joint). We set a minimum magnitude threshold at 6 Nm to ensure sufficient assistance for each user, necessary to produce a meaningful biomechanical effect (for Experiment 1, peak hip torque ranged from 6 to 9 Nm). Immediately following this acclimation phase, the online adaptation framework was applied by toggling a keyboard input from a separate experimental computer. During the adaptation phase, the gait phase model weights were updated every 5 seconds. The adaptation was applied for a minimum of 10 iterations with a stopping criterion of 2% convergence tolerance (gait phase error not changing more than 2%). Following the adaptation trial, two one-minute validation walking trials were conducted: one using the baseline gait phase model and the other using the adapted model. During this experiment, we recorded the estimated gait phase, hip joint angle, and applied joint torque.

Fig. 7.

Experimental protocol to evaluate the performance of the online adaptation framework on robotic hip exoskeletons. The framework was evaluated on different users and walking environments to test its generalizability. (A) A group of able-bodied subjects walked on a standard treadmill while wearing a robotic hip exoskeleton. (B) Three able-bodied subjects walked on a split-belt treadmill inducing an asymmetric gait environment. (C) A group of stroke survivors walked on a treadmill while the framework adapted the baseline gait phase model.

Experiment 2 had a similar walking protocol, except all subjects walked on a split-belt treadmill (Bertec, OH), operating with different belt speeds to simulate an asymmetric gait (Fig. 7B). The primary purpose was to induce an asymmetric gait in an able-bodied population to pilot test the algorithm before using it with stroke subjects. For all three subjects, the treadmill’s left and right belts were set to 0.8 m/s and 1.6 m/s, respectively. Initially, both belts started at 1.6 m/s, after which one belt was reduced to 0.8 m/s to induce asymmetric gait, simulating the slower paretic-side gait commonly observed in stroke populations. Despite this adjustment, the setup maintained the same effective average speed as in Experiment 1, which was conducted under tied-belt conditions. All other protocol details, including the acclimation and adaptation phases, remained the same as in Experiment 1.

For Experiment 3, six stroke survivors participated in a similar walking protocol as Experiment 1 (Fig. 7C). For the stroke group, since each subject had a different self-selected walking speed, we used an overground walkway (Zeno, ProtoKinetics, PA) to compute the subject’s baseline walking speed of not wearing the exoskeleton. The treadmill speed was set to 80% of each subject’s self-selected overground walking speed. Similar to the previous experiments, we used the biological torque controller for the exoskeleton providing bilateral hip assistance. The torque magnitude was consistently preset to 6 Nm across subjects to account for gait variations in stroke subjects. Higher values could cause mechanical device ‘play’ due to using the same timing controller parameters across subjects.

During the acclimation phase, we increased the assistance magnitude incrementally where the assistance level for the paretic and non-paretic legs was set asymmetrically based on the subject’s comfort and on-site clinician’s observation of the overall gait quality. Gait quality was assessed based on the following criteria: 1) alignment of the hip joint between the subject and the actuator, 2) presence of any new inappropriate gait deviations, 3) reduction of existing gait deviations, and 4) improvement in symmetry and swing phase clearance. Following the acclimation phase, we performed an adaptation trial where the minimum number of iterations and stopping criteria were set equal to Experiment 1.

B. Data Analysis

For the validation trial, gait phase RMSE was calculated from one minute of walking data. RMSE is a popular evaluation metric in the field [51]-[53], compared to other metrics such as mean absolute error, as it is more sensitive to outliers and allows for a greater penalty on larger errors than smaller estimation errors. Additionally, the use of RMSE as a commonly used evaluation metric for gait phase estimation allows for easy comparison of our model performance across many studies analyzing gait phase estimation error with varying techniques. For each walking condition (baseline and adapted), we determined the ground truth using the hip joint angle for each leg, evaluated all error points from the last 10 gait cycles, and calculated the RMSE. Similarly, we calculated the joint torque RMSE for the validation trials. For both gait phase and joint torque results, we also calculated the relative RMSE improvement. The relative improvement was determined by the error difference between the adapted model and the baseline model, divided by the baseline model error.

C. Statistical Analysis

For the adaptation and validation trials, we conducted a statistical analysis of the gait phase estimation error for both baseline and adapted models. We performed a one-way (for stroke subjects a two-way, where the second independent variable was the leg) repeated measures analysis of variance (ANOVA) by setting a significance level of 0.05 (Matlab 2019b, Mathworks, MA). For the stroke experiment, additional post-hoc analysis with a Bonferroni correction was applied to evaluate a statistical difference between each model.

D. Pilot Testing of the Online Adaptation Framework for Improving Stroke Gait During Level-Ground Locomotion

We wanted to expand our framework to assess the feasibility of our system in a realistic setting and explore additional features that could enhance its biomechanical benefits. This served as a proof of concept for future studies, demonstrating the potential of our system to improve gait biomechanics in stroke survivors. For this pilot Experiment 4, a single stroke survivor walked while wearing the GEMS hip exoskeleton (Fig. 8A). Our first objective was to examine how gait phase estimation error influences performance in a real-world scenario, such as overground walking, where step-to-step speed variability leads to greater fluctuations in error. We compared the final adapted model to the baseline model to assess its robustness under these conditions. Following the overground trial, we conducted a treadmill walking trial to enable a more formal analysis of biomechanical outcome measures. The treadmill was controlled with a self-paced mode using the real-time position of a motion capture marker (Nexus 2.12, Vicon Motion Systems, UK). To ensure the safety of the subject, the speed range of the self-paced treadmill was set to ±50% of the subject’s baseline self-selected walking speed from the overground walkway.

Fig. 8.

(A) Pilot testing of the online adaptation with additional control parameter optimization in improving the stroke survivor’s outcome measures (walking speed and energetic cost). (B) Bayesian-based human-in-the-loop optimization of exoskeleton control parameters using the user’s self-selected walking speed.

Along with our online adaptation, we wanted to optimize the underlying exoskeleton control parameters to generate a personalized assistance profile for the stroke subject. To further optimize parameters, we implemented a Bayesian human-in-the-loop optimization (Fig. 8B) based on previous literature [37]. Different from a conventional approach, we used the user’s self-selected walking speed as the evaluating cost function [33]. The four control parameters for optimization were bilateral peak flexion and extension assistance timing. To ensure that the high-level gait phase adaptation and mid-level controller parameter optimization do not counteract each other, we conducted the adaptation/optimization experiment sequentially. We first performed our online adaptation of the stroke survivor’s gait phase using a generic assistance profile (same as the one for able-bodied subjects). Once the adaptation was completed, we locked the neural network model weights. Afterward, we carried out the parameter optimization process using this adapted gait phase model. For the first iteration, the general timing parameters (optimized based on able-bodied subjects) were evaluated. At the end of each iteration (one minute each), the subject’s walking speed was evaluated by taking an average speed from the last 15 seconds of data. During this evaluation phase, the optimization updated the cost landscape (axis representing each control parameter) using Gaussian processes [54]. Based on the updated landscape, the next sampling parameters were chosen by maximizing the expected improvement. After a total of 24 iterations (break was enforced after every six iterations), we stopped the optimization process and selected the final optimized control parameters. All processes (including the self-paced treadmill) were operated on a host computer (Matlab 2019b, MathWorks, MA) adjacent to the treadmill.

Once the optimization was completed, we evaluated the personalized exoskeleton (adapted gait phase model + optimized controller parameters). Two proof-of-concept validation trials were conducted: 1) self-selected walking speed and 2) metabolic cost of walking. These trials were performed separately because walking speed-based optimized parameters do not necessarily guarantee significant metabolic cost benefits (e.g., increasing walking speed resulting in higher energy expenditure). For the walking speed trial, the subject walked on the same self-paced treadmill (two minutes each) with and without the exoskeleton. For the metabolic cost trial, the subject walked on a treadmill at a constant speed, set to their baseline walking speed, for six minutes with and without exoskeleton (a short break was provided between conditions to prevent fatigue). During each condition, the subject’s metabolic cost was measured using an indirect calorimetry system (K5, COSMED, Italy). The metabolic cost trial consisted of three conditions where we recorded the user’s respiratory data: 1) quiet standing, 2) walking without the exoskeleton, and 3) walking with the exoskeleton. For each condition, we calculated the user’s metabolic cost from the last 2 minutes of oxygen and carbon dioxide rates using a modified Brockway equation [55]. The reported metabolic cost from the experiment was the user’s net metabolic cost where we subtracted the quiet standing metabolic cost from each corresponding condition.

VII. Results

A. Experiment 1: Real-Time Inference and Online Adaptation of Gait Phase on Able-Bodied Subjects

The online adaptation framework was evaluated on able-bodied subjects to observe its performance in a population with minimal variations in gait dynamics (Fig. 7A). On average, the online adaptation framework converged after 3 iterations (corresponds to 15 seconds) to a final adapted model (Fig. 9A). After the adaptation phase, the final adapted model reduced the relative gait phase estimation root mean squared error (RMSE) by 40.9±11.5% (mean±standard deviation) from the initial baseline model to a final absolute gait phase RMSE of 1.8±0.2% (paired t-test, p = 0.0002). This difference in the estimation performance was also consistent in an additional online validation trial, verifying the model’s ability to maintain accuracy outside of the initial adaptation trial (Fig. 9B). During the validation trial, the adapted model had an absolute gait phase RMSE of 2.0±0.4%, resulting in a 31.8±16.1% lower relative gait phase RMSE than the baseline model (paired t-test, p = 0.002). The influence of the gait phase estimation error propagated to the error in the applied exoskeleton joint torque. Torque error was quantified as the RMSE between the ideal gait-phase-based ground truth and the exoskeleton’s estimated torque, capturing magnitude and timing misalignment. The coefficient of determination (R2) of 0.95 between gait phase errors and joint torque errors indicated that there was a strong linear correlation between gait phase and exoskeleton assistance as expected. For exoskeleton torque, the adapted model had a torque magnitude RMSE (% target assistance level) of 9.9±3.2%, resulting in a 32.7±16.6% lower torque RMSE than the baseline model (paired t-test, p = 0.003, Fig. 9C).

Fig. 9.

Online adaptation framework performance validation on able-bodied subjects. The performance of the adaptation framework was evaluated on able-bodied subjects walking on a treadmill at 1.2 m/s. (A) The online adaptation framework converged after 3 iterations to a final adapted model. (B) The adapted model maintained performance during a separate validation trial. (C) The gait phase estimation error directly influenced the error in the applied exoskeleton torque. The shaded regions and error bars represent ±1 standard error of the mean (SEM) and the asterisks indicate statistical differences (p<0.05).

B. Experiment 2: Online Adaptation Framework During Simulated Asymmetric Gait

The online adaptation framework was evaluated on three able-bodied subjects walking on a split-belt treadmill, inducing an asymmetric gait (Fig. 7B). Inducing this simulated gait walking condition resulted in a baseline absolute gait phase RMSE of 6.8±4.3% for the speed-changed leg (slow speed, Fig. 10A). During the adaptation phase, both legs converged to a steady gait phase estimation after 6 iterations. The final adapted model reduced the relative gait phase RMSE by 44.4±9.8% from the baseline model to a final absolute RMSE of 2.0±0.5% for the unchanged leg and reduced the relative gait phase RMSE by 58.5±36.0% to a final absolute RMSE of 2.8±0.7% for the speed-changed leg (Fig. 10B).

Fig. 10.

Online adaptation framework performance validation on able-bodied subjects in a simulated asymmetric gait environment. Three able-bodied subjects walked on a split-belt treadmill (belt speed of 0.8 and 1.6 m/s) to evaluate the performance of our adaptation framework in a simulated asymmetric gait environment. (A) The online adaptation framework converged reliably on both legs after 6 iterations to a final adapted model. (B) The final adapted model reduced the gait phase estimation error reliably from the baseline model for both legs. The shaded regions and error bars represent ±1 SEM.

C. Experiment 3: Generalizability of the Online Adaptation Framework to Stroke Populations

The online adaptation framework was evaluated on a group of stroke survivors to validate its consistency in performance on a population with high variations in inter-limb coordination (Fig. 7C). Similar to the simulated asymmetric gait result, the baseline model performance was different between the paretic and non-paretic leg with an absolute gait phase RMSE of 10.5±4.8% and 6.2±1.8%, respectively (Fig. 11A). During the adaptation phase, the paretic and non-paretic legs converged to a steady gait phase estimation after 7 and 4 iterations, respectively. The final adapted model reduced the relative gait phase RMSE by 69.6±9.2% from the baseline model to a final absolute gait phase RMSE of 2.9±0.6% for the paretic leg and reduced the relative gait phase RMSE by 62.2±10.7% on average to a final absolute RMSE of 2.2±0.3% for the non-paretic leg (paired t-test, p=0.001 and p=0.004, Fig. 11B).

Fig. 11.

Online adaptation framework performance validation on stroke survivors. The performance of the adaptation framework was evaluated on stroke survivors walking on a treadmill at their self-selected walking speed. (A) The online adaptation framework converged reliably after 7 and 4 iterations for the paretic and non-paretic legs to a final adapted model, respectively. (B) The final adapted model reduced the gait phase estimation error reliably from the baseline model for both the paretic and non-paretic legs. The shaded regions and error bars represent ±1 SEM and the asterisks indicate statistical differences (p<0.05).

For one of the stroke survivors, we evaluated the final adapted model in an outdoor walking condition. During the validation trial, the adapted model had an absolute gait phase RMSE of 6.5% and 4.7% for the paretic and non-paretic leg, resulting in a 70.4% and 44.2% lower relative gait phase RMSE than the baseline model, respectively (Fig. 12A). As shown with the representative subject’s hip joint angle (Fig. 12B), poor gait phase estimation is mostly influenced by high variations in hip kinematics of the paretic leg, specifically when the subject dynamically changed the walking speed. This deviation in the subject’s gait pattern caused the baseline model to false detect certain gait events, often generating a suboptimal assistance profile (Fig. 12D).

Fig. 12.

Representative stroke survivor’s adaptation performance in an overground validation trial. (A) The adapted model maintained high performance for both the paretic and non-paretic legs during a separate outdoor overground validation trial. (B) the subject’s hip joint kinematics exhibiting a reduced hip range of motion in the paretic leg. (C) The subject’s variations in the gait pattern resulted in a substantial performance degradation in gait phase estimation for the baseline model. (D) The baseline model’s poor gait phase estimation resulted in a suboptimal hip assistance profile.

D. Experiment 4: Pilot Testing of the Online Adaptation Framework for Improving Stroke Gait During Level-Ground Locomotion

The final adapted model reduced the relative gait phase RMSE by 79.6% from the baseline model to a final absolute gait phase RMSE of 2.9% for the paretic leg and reduced the relative gait phase RMSE by 55.4% to a final absolute gait phase RMSE of 3.5% for the non-paretic leg (Fig. 13A). Our online validation result showed that the fully optimized system (both gait phase estimation and assistance timing parameters) increased the stroke survivor’s self-selected walking speed by 21.8% and reduced the metabolic cost of walking by 6.5% compared to not wearing the exoskeleton (Fig. 13C).

Fig. 13.

Integration of gait phase adaptation and control parameter optimization in improving stroke gait. We tested the integrated personalization framework on stroke gait via online gait phase adaptation and control parameter optimization using the Samsung exoskeleton. (A) The online adaptation framework performed reliably, and the final adapted model reduced the gait phase estimation RMSE by 79.6% and 55.4% from the baseline model for the paretic and non-paretic leg, respectively. (B) Using human-in-the-loop optimization, we optimized the exoskeleton’s flexion and extension assistance timing based on the subject’s self-selected walking speed. (C) Using the fully integrated system, the exoskeleton increased the stroke survivor’s self-selected walking speed by 21.8% and reduced the metabolic cost of walking by 6.5% compared to not wearing the exoskeleton.

VIII. Discussion

Literature has shown that assistance timing is a key control parameter that governs the overall human-exoskeleton performance, such as reducing the metabolic cost of walking [29], [56]. Gait phase, a percentage representation of the user’s joint configuration during a gait cycle, allows exoskeleton assistance timing to operate in a time-independent manner (e.g., adaptable to changes in the user’s walking speed). Our linear correlation result supports this notion as gait phase estimation error directly influenced the error in the applied joint torque (Fig. 12D), which is as expected. This immediate effect of gait phase accuracy on the assistance level is substantial as it can potentially reduce the overall human-exoskeleton performance (e.g., metabolic cost). For example, a 6% gait phase shift in an ankle exoskeleton assistance timing can alter the overall metabolic cost benefit by 3.5% [56], indicating the need for accurate gait phase estimation for exoskeleton control.

The current direction in this field is exciting because recent literature has emphasized the importance of accurate assistance timing in maximizing net human-exoskeleton performance [19], [27], [33], [37]. In this regard, several recent studies have shown promising results in developing high-performing gait phase estimators [45], [57]-[63]. However, most of these studies have focused on developing either a user-independent model or a model that generalizes to other environmental settings, with a lack of understanding of how these models perform across user populations, especially in users who exhibit a drastic change in gait pattern. As a result, many high-performing gait phase models were often evaluated on a population with similar aggregate joint and limb trajectories to the gait dynamics of a new user. Although some studies have been conducted on populations other than able-bodied individuals [59], [61], [63]-[65], except for Livolsi et al., these studies have primarily tested exoskeleton systems on elderly populations, where the gait dynamics are similar to those of the able-bodied population, but with slight changes in kinematics, such as a reduced range of motion. Consequently, a user-independent model can be developed for this population.

However, for patient populations, such as stroke survivors with gait asymmetry, relying on user-independent deep learning models for exoskeletons may not be effective. So far, for user-independent estimation and control in patient populations, adaptive oscillators have been the most effective gait phase estimation method, demonstrating success in stroke gait, amputee, and brain injury [63]-[65]. Conversely, while deep learning models excel in handling dynamic transient movements, an area where adaptive oscillators fall short, they suffer performance degradation in populations with significant gait deviations. However, our work addressed this limitation by integrating an adaptation pipeline, successfully estimating gait phase in stroke survivors without manual tuning or experimenter intervention.

Our adaptation results have demonstrated the robustness of our framework in multiple settings, corroborating our hypothesis that the deep learning model can further improve performance by learning user-specific data and personalizing the underlying model parameters to the individual’s gait patterns. Although the absolute gait phase error of 1% for able-bodied subjects may appear insignificant, it is crucial to consider the relative improvement in comparison to the baseline model. In this study, the baseline model was already a state-of-the-art user-independent model that worked effectively for various able-bodied individuals. Despite this, our adaptation results demonstrated that the system can still adapt, even with a high-performing baseline model. We think that this improvement would be even more substantial for researchers using our framework with a mediocre-performing gait phase estimator. This is often the case due to limited access to state-of-the-art models.

Moreover, our adaptation framework demonstrated robustness in reliably adapting to the user’s gait dynamics in various settings, such as different locomotion modes and walking speeds (as demonstrated in the Appendix), highlighting the system’s capability to adapt to different walking conditions. This generalizability was not limited to able-bodied subjects but also extended to stroke populations. This was exemplified in the representative stroke survivor’s overground data (Fig. 12C), where our adapted model consistently maintained high estimation performance during an outdoor overground trial, demonstrating its resilience to changes in the walking condition (i.e., speed changes).

Lastly, our approach was effective in stroke gait adaptation. Kinematic variations in stroke gait (Fig. 12B), such as reduced range of motion in the paretic leg, greatly degrade the baseline model’s performance as such patterns were not captured within the original able-bodied training dataset. However, the online adaptation framework not only leveraged the baseline model’s gait representations but also quickly learned the new user’s inherent gait dynamics in less than a minute of walking. In both the able-bodied (asymmetric gait) and stroke groups, our framework reliably converged to an adapted model and consistently reduced the gait phase estimation error rate on average by 58.6% from the baseline model (p<0.05). This adaptability and flexibility make our method a promising approach for translating exoskeleton technology to stroke survivors, where collecting extensive user-specific data is often impractical. We consider the absolute improvement of 8% in gait phase estimation (corresponding to 80% relative improvement) on the paretic side (as shown in Fig. 11) to be substantial. To provide a comprehensive explanation, a typical stride duration of an average stroke survivor is roughly 2 seconds [66] and this 8% gait phase estimation error corresponds to 160 ms. Previous literature has emphasized the importance of assistance timing in exoskeleton control [19] and a constant 160 ms error (either leading or lagging) in timing can have a significant impact on the overall human outcome measures, such as an increase in the user’s metabolic cost.

Literature has indicated that optimal assistance strategy (e.g., magnitude and timing) varies across users [19], [37]. For able-bodied subjects, this exoskeleton performance benefit can range up to, in the context of the metabolic cost of walking, 18% if the selected assistance profile is suboptimal [67]. Recently, the field has started to adopt this optimization scheme in exploring optimal exoskeleton assistance levels for stroke gait [12], [68]. Human-in-the-loop optimization is a popular, yet powerful, method that directly integrates the user’s feedback into the control parameter optimization process [69]. While it was a single-subject proof-of-concept validation, we want to highlight that completing the personalization exoskeleton framework would require mid-level control parameter optimization. Our exciting results of improving the stroke survivor’s mobility by increasing the walking speed (21.8%) and reducing the metabolic cost of walking (6.5%) paved a promising direction. In the final part of our study, we aimed to test the feasibility of integrating mid-level control parameter optimization in our online adaptation framework with patient populations. Specifically, we conducted pilot testing on a stroke survivor, as there was a lack of understanding in the current field regarding the application of human-in-the-loop optimization in patient populations. Although an N=1 pilot experiment, we wanted to demonstrate the feasibility of synergizing our adaptation approach with a standard human-in-the-loop optimization which has produced promising results. However, future work should fully validate this integrated system by conducting large-scale clinical studies. Some potential research directions relating to this would include determining the objective function (metabolic cost vs. walking speed) and optimizing other control parameters (assistance magnitude vs. timing). We think that our high-level online adaptation and mid-level controller optimization contribute to the same goal of personalizing the exoskeleton system to a new user. These two layers, interlinked and dependent on one another, created a fully stacked personalization framework, significantly contributing to maximizing the user’s gait performance as the system organically adapts and learns the user’s gait pattern over time.

The current study has several limitations that need to be addressed in future work. One key limitation is the small number of stroke survivors tested. To fully validate the framework’s ability to enhance gait biomechanics, a larger pool of subjects is needed. Additionally, our system adapts only to moderate assistance levels that avoid spastic responses in stroke gait. Furthermore, our recruited stroke survivors were relatively high-functioning, which may result in discrepancies in gait phase adaptation performance for individuals with slower walking speeds. Future biomechanical studies should explore varying assistance magnitudes and adaptation frameworks that maintain effectiveness across different impairment levels while minimizing adverse effects, enabling safer and more personalized high-assistance exoskeleton control. Another limitation is the potential for skewed learning (i.e., catastrophic forgetting), where the model may overfit to incoming data, compromising robustness across multiple users and environments. For example, providing large torque assistance can cause mechanical ‘play,’ leading to false peak detection and a feedback loop of degraded performance. Additionally, the adaptation framework relies on a backward labeler for ground truth labels, which is straightforward for gait phase but more complex for other state variables like physiological state. Lastly, online adaptation requires cyclical movement patterns, and sudden stops can disrupt the model training due to inaccurate labeling.

A great advantage of our framework is its generalizability to other applications. This personalization framework is transferrable to other high-level user state variables beyond gait phase. As long as the variable can be reliably reconstructed post-hoc, then an adaptation framework such as this one can be used. For example, we have demonstrated this works for walking speed and locomotion mode (detailed results presented in the Appendix) but could be extended to other state variables such as ground slope estimation [70], extending the framework to other ambulation contexts beyond level-ground walking. Importantly, our user-independent gait phase system is agnostic to locomotion modes such as stairs and ramps due to our unique training strategy [46] allowing it to seamlessly integrate with these other strategies in the future. Since backward labeling in these other modes is feasible, the framework can maintain reliable adaptation performance in these settings.

Another important feature of our framework is its ability to transfer the model across physical hardware. In this study, we demonstrated the ability to transfer the gait phase model across devices by simply applying a transformation to the incoming data stream. However, this sensor data transformation was applied to exoskeletons targeting the same joint. To expand this framework to a new joint, a new set of baseline training data needs to be collected one time, but then similar transformations can be created to transfer between hardware. Regarding other joints, one scenario that needs be considered is the use case of multi-joint exoskeleton control (e.g., hip and knee joints). In this case, the performance of our model will not be affected unless the assistance provided at the other joint significantly alters the user’s hip kinematics. However, such cases are rare, as indicated by our previous study which showed that excessive assistance that modifies the user’s nominal joint kinematics can have a negative impact on the overall human-exoskeleton performance [25]. Since the primary objective of exoskeleton control is to provide assistance that enhances human outcome measure, excessive assistance is unlikely to occur. This indicates that our baseline gait phase model performance should be maintained even if an additional degree-of-freedom is added to the exoskeleton structure.

IX. CONCLUSION

Our online adaptation framework demonstrated the feasibility of translating exoskeleton technology to real-world applications. Specifically, we showed that the exoskeleton sensor suite can be adapted to new hardware, enabling seamless model transfer from research-grade devices to commercial exoskeletons. Furthermore, we successfully adapted the baseline gait phase model for both able-bodied individuals and stroke survivors, achieving significant performance improvements with less than a minute of adaptation. Lastly, while it was a proof-of-concept, integrating our framework with human-in-the-loop optimization improved a stroke survivor’s gait function in both energy and walking speed. Future studies should build on our work by rigorously validating how this adaptation strategy can improve biomechanical outcomes. We believe that the next phase of research should focus on testing and validating our framework in outdoor and real-world settings. The past decade of exoskeleton research has focused on understanding effective assistance strategies during locomotion within controlled environments. While these efforts have significantly advanced the state of exoskeleton systems [71], more researchers need to move outside of the laboratory environment and into the real world for meaningful advancement in the field to continue. Our research findings provide significant scientific and technological contributions to the field, starting a broader movement of taking exoskeletons into the real world for improving the mobility of diverse populations.

Biographies

Inseung Kang received the B.S., M.S., and Ph.D. degrees, all in Mechanical Engineering, from the Georgia Institute of Technology, Atlanta, GA, USA, in 2016, 2018, and 2021, respectively. From 2022 to 2023, he was a Postdoctoral Associate in the Department of Brain and Cognitive Sciences at the Massachusetts Institute of Technology, Cambridge, MA, USA. He is currently an Assistant Professor in the Department of Mechanical Engineering at Carnegie Mellon University, Pittsburgh, PA, USA, where he has directed the MetaMobility Lab. His research focuses on the intersection of wearable technology, robotics, and artificial intelligence to improve lower-limb mobility for individuals with motor impairments.

Dean D. Molinaro received the B.S. degree in Mechanical Engineering from the University of Florida, Gainesville, FL, USA, in 2017 and the Ph.D. degree in Robotics from the Georgia Institute of Technology, Atlanta, GA, USA, in 2023. He is currently an Applied Scientist at the Robotics and AI (RAI) Institute, Cambridge, MA, USA, working on control of agile, autonomous robots. His research interests include autonomous robotics, wearable systems, and learning-based methods for robotic control.

Dongho Park received the bachelor’s degree in Robotics from Kwangwoon University, Seoul, South Korea, in 2018 and the master’s degree in Medicine from Yonsei University, Seoul, South Korea, in 2021. He is currently a Ph.D. student in Robotics at the Georgia Institute of Technology, Atlanta, GA, USA, under the supervision of Prof. Aaron Young. His research interests include lower-limb exoskeletons, mobility rehabilitation, and gait biomechanics.

Dawit Lee received Bachelor’s, Master’s, and Ph.D. degrees in Mechanical Engineering from the Georgia Institute of Technology in 2015, 2018, and 2023, respectively. His dissertation work focused on developing effective lower-limb wearable systems for improving human mobility under the supervision of Prof. Aaron Young. He is currently a Postdoctoral Scholar at Stanford University. His research interests include personalized control strategies for wearable assistive devices, physics- and AI-based modeling of human movement, and the development of translational biomechanics tools for biomechanics assessment.

Pratik Kunapuli received the bachelor’s degree in Computer Engineering and master’s degree in Electrical and Computer Engineering from Georgia Institute of Technology in 2019 and 2020, respectively. He conducted his master’s thesis, under the supervision of Dr. Aaron Young, related to online adaptation of state estimation for exoskeleton devices. He is currently a Ph.D. student at the University of Pennsylvania, advised by Dr. Vijay Kumar and Dr. Dinesh Jayaraman. His research interests lie at the intersection of machine learning and robotics, specifically enabling agile control of aerial vehicles with reinforcement learning.

Kinsey R. Herrin received the B.S. degree in chemistry from the University of Georgia, Athens, GA, USA, in 2008, and the M.S. degree in prosthetics and orthotics from the Georgia Institute of Technology, Atlanta, GA, USA, in 2010. She received the clinical orthotics residency from Children’s Healthcare of Atlanta, Atlanta, GA, USA, in 2011 and the clinical prosthetics residency from the University of Michigan, Ann Arbor, MI, USA, in 2013. She is currently a Principal Research Scientist with the Georgia Institute of Technology, Woodruff School of Mechanical Engineering, Atlanta, GA, USA. Her research interests include wearable robotics for disability, human interface design, and clinical outcomes-driven research.

Aaron J. Young received the B.S. degree in biomedical engineering from Purdue University,West Lafayette, IN, USA, in 2009, and the M.S. and Ph.D. degrees in biomedical engineering from Northwestern University, Evanston, IL, USA, in 2011 and 2014, respectively. He is currently an Associate Professor with the Woodruff School of Mechanical Engineering, the Georgia Institute of Technology, Atlanta, GA, USA, and has directed the Exoskeleton and Prosthetic Intelligent Controls Lab since 2016. His research interests include AI-integrated control systems for robotic prosthetic and exoskeleton systems for improving lower limb human mobility.

Appendix

A. Inclusion and Exclusion Criteria for Stroke Participants

Stroke subjects who participated in our study (Table 1) had to meet our inclusion criteria: they had to be between 18 and 85 years old, have a history of stroke occurring at least 6 months prior, and obtain physician approval for safe participation. They needed a Mini-Mental State Examination (MMSE) score above 17, the ability to sit unsupported for at least 30 seconds, and the capacity to follow a 3-step command. Participants were required to walk unaided (rail use allowed) at a minimum speed of 0.4 m/s, walk for at least 6 minutes, and navigate small slopes (3 degrees) and steps (6 steps). They also had to be willing and able to participate in a 1 to 4-hour experiment with breaks.

Exclusion criteria included loss of sensation in the legs, complete spinal cord injury, recent concussion (last 6 months), severe cardiovascular conditions, severe arthritis or orthopedic issues limiting lower body movement, other neurological disorders (e.g., Parkinson’s, ALS, MS, dementia), history of head trauma, lower extremity amputation, non-healing lower extremity ulcers, renal dialysis or end-stage liver disease, legal blindness or severe visual impairment, pacemaker or metal implants in the head region, use of seizure-threshold-lowering medications, concurrent participation in another clinical study, and any condition deemed by the researcher to affect study outcomes or confound results.

B. Effect of Using Multi-Mode Gait Phase Estimation for Online Adaptation

The aim of this study was to extend the framework’s applicability to different environmental settings, including various ambulation modes. In such cases, a multi-mode gait phase estimator would be required for the baseline model. However, since the main study only evaluated the system’s adaptation performance on level-ground walking, it was crucial to ensure that there was no performance discrepancy between the two baseline models in this particular mode. To address this concern, we performed an offline analysis using treadmill walking data from three able-bodied individuals (all males, mean age of 30.0±3.6 years, height of 174.3±8.3 cm, and body mass of 66.9±6.2 kg), similar to Experiment 1 in the main study.

We evaluated the performance of two gait phase estimator models: 1) a model trained on data from five different locomotion modes (i.e., the baseline model in the main study) and 2) a model trained on data from level-ground walking only. To ensure a fair comparison, we used the same neural network architecture and hyperparameters as in the main study. The results showed that the difference in gait phase estimation performance between the two models during level-ground treadmill walking was negligible (Fig. 14). Furthermore, running a simulated adaptation trial using offline data with these two models confirmed that the adaptation performance converged similarly to the results obtained for able-bodied subjects (Fig. 9 in the main text). This indicates that our adaptation framework was capable of learning the user’s specific gait pattern rather than adapting to the generic ambulation task.

C. Offline Adaptation Performance Validation of Gait Phase and Walking Speed Estimation

Prior to deploying our framework, we conducted offline validation to ensure that the developed algorithm could reliably adapt to new users. Additionally, during this validation period, we tested the framework’s generalizability to other user state variables, such as walking speed. 11 additional subjects (10 males and 1 female, mean height of 1.80±0.05 m, mean body mass of 72.5±4.98 kg, and mean age of 23.1±1.7 years) were recruited for data collection after providing informed consent. All subjects walked on a treadmill at speeds ranging from 0.3 m/s to 1.2 m/s with 0.1 m/s increments (one minute per speed condition) for a total of 10 minutes while wearing a robotic hip exoskeleton presented in our previous work [72]. This varying speed profile was specifically chosen to mimic an overground scenario where the user can dynamically modulate their walking speed, which can influence overall estimation performance (e.g., poor gait phase estimation during slow walking speeds). During all trials, unilateral exoskeleton sensor data (hip joint encoder, trunk, and thigh IMUs) were recorded at 100 Hz. The exoskeleton utilized a biological torque controller to provide bilateral hip assistance during each trial. The controller employed a conventional time-based estimation method using a foot switch to detect gait events.

TABLE I.

Stroke Subject Baseline Information

Subject ID	Gender	Age (years)	Paretic Side	Time Since Stroke (months)	Height (cm)	Body Mass (kg)	Self-Selected Walking Speed (m/s)
ST1	M	38	R	108	180.5	89	0.64
ST2	M	58	L	146	175.5	81	0.86
ST3	M	59	R	150	190	87.2	0.87
ST4	M	32	R	36	165	81.6	1.18
ST5	F	67	L	62	156	57.9	0.98
ST6	F	47	R	82	167.5	59.9	1

Fig. 14.

Gait phase estimator performance during level-ground walking using different baseline models. Two different baseline models were used: 1) a model trained on data from 5 locomotion modes and 2) a model trained on data from only level-ground mode. (A) For both systems, the adaptation framework reduced the baseline estimation error compared to the baseline model. (B) Using a different set of data did not have any effect on the underlying gait phase estimation performance during level ground.

Fig. 15.

Offline performance validation of the adaptation framework. The user’s gait phase and walking speed estimation models were adapted using the framework while the treadmill speed varied. The fluctuations in the estimated gait phase and walking speed were due to the dynamic changes in the user’s walking speed. The shaded regions represent ±1 standard deviation.

D. Generalizing the Adaptation Framework to Other Locomotion Modes

One key objective of this study was to deploy the exoskeleton system to real-world settings. To ensure that our online adaptation framework can reliably adapt to the user’s gait in varying environments, it was important to validate the system’s generalizability in other locomotion modes. We utilized a dataset from our previous human subject experiment [46] during multi-mode locomotion (i.e., ramp ascent and ramp descent with a slope incline of 11°) while wearing a hip exoskeleton, which was the same dataset used in generating the baseline gait phase estimator. To ensure user-independent nature of the baseline model, we utilized training data from nine subjects at a time to run a 10-fold validation for each testing subject. During adaptation, we utilized the same baseline parameters as in the main study (the ramp mode adaptation had fewer iterations than the level-ground, primarily due to the limited number of strides available in the dataset). Our results indicated that the system reliably adapted after five iterations for each given mode (Fig. 16). Post-adaptation, the final adapted model reduced the relative gait phase estimation RMSE by 42.61±13.86% for ramp ascent and 32.23±15.15% for ramp descent compared to the baseline model (p<0.05).

Fig. 16.

Performance of the online adaptation framework during multimodal locomotion. (A) The baseline gait phase estimation model was able to adapt to a new locomotion mode after 5 iterations. (B) After the adaptation phase, adapted model significantly reduced the gait phase estimation error compared to the baseline model. The shaded regions and error bars represent ±1 standard error of the mean and asterisks indicate statistical significance (p<0.05).