Recent advancements in data-driven methods have enabled real-time estimation of biomechanical states for exoskeleton control. While biological joint moments can be directly used to scale exoskeleton assistance, this approach is often suboptimal. An optimized mapping between biological joint moments and exoskeleton assistance could enhance end-to-end controllers based on the user’s physiological state. We introduce a flexible parametrization of biological moment-based control using delay, scaling, and shaping terms to transform joint moment estimates into commanded torque. We performed human-in-the-loop optimization, using metabolic cost to evaluate each iteration’s controller parameters, for 9 subjects across three ambulation modes: level walking at 1.1 m/s, 1.5 m/s, and 5° inclined walking. We evaluated three methods of exoskeleton control: 1. Personalized/Task Dependent, 2. Theoretical Task Dependent/Non-personalized, and 3. Theoretical Multi-Task/Non-personalized. On average, our personalized approach provided the greatest benefit of 18.3% reduction in metabolic cost compared to walking without the exoskeleton, with the task-dependent and multi-task controllers producing similar reductions of 8.6% and 8.4%, respectively. Our results show that fully personalized exoskeleton control parameters can yield large metabolic reductions for cyclic ambulatory tasks when compared to a theoretical generalizable, multi-task controller, highlighting the value of personalizing exoskeleton assistance to users across ambulatory tasks.

Objective: Accurate, non-invasive methods for estimating joint and muscle physiological states have the potential to greatly enhance control of wearable devices during real-world ambulation. Traditional modeling approaches and current estimation methods used to predict muscle dynamics often rely on complex equipment or computationally intensive simulations and have difficulty estimating across a broad spectrum of tasks or subjects. Methods: Our approach used deep learning (DL) models trained on kinematic inputs to estimate internal physiological states at the knee, including moment, power, velocity, and force. We assessed each model's performance against ground truth labels from both a commonly used, standard OpenSim musculoskeletal model without EMG (static optimization) and an EMG-informed method (CEINMS), across 28 different cyclic and noncyclic tasks. Results: EMG provided no benefit for joint moment/power estimation (e.g., biological moment), but was critical for estimating muscle states. Models trained with EMG-informed labels but without EMG as an input to the DL system significantly outperformed models trained without EMG (e.g., 33.7% improvement for muscle moment estimation) (p < 0.05). Models that included EMG-informed labels and EMG as a model input demonstrated even higher performance (49.7% improvement for muscle moment estimation) (p < 0.05), but require the availability of EMG during model deployment, which may be impractical. Conclusion/Significance: While EMG information is not necessary for estimating joint level states, there is a clear benefit during muscle level state estimation. Our results demonstrate excellent tracking of these states with EMG included only during training, highlighting the practicality of real-time deployment of this approach.

Motor module analysis is an important tool in the study of movement, particularly in people with impaired neural control. The most common method for computing motor modules is non-negative matrix factorization (NMF), which identifies a matrix of motor modules and their corresponding time-series activity from electromyography data. NMF has several limitations, including dependence of the muscle weightings on the number of modules selected. Approaches for selecting the number of modules vary between studies, making it difficult to compare and reproduce results. Some metrics of motor control complexity use the variance accounted for when extracting a single motor module (VAF 1 ), yet that module’s structure offers little biomechanical interpretability. In this work, we present a method for computing motor modules using an autoencoder, a neural network architecture that can find latent representations of data. Using a single layer autoencoder, we extracted motor modules from data in able-bodied and individuals post-stroke. The structure of autoencoder-computed modules were significantly less sensitive to selected module number. With the autoencoder-computed modules, increasing the number of modules added new information, instead of splitting previous modules. Autoencoder-computed modules, especially at low module counts, had more distinct and interpretable biomechanical functions. Lastly, the autoencoder-computed modules are consistent with previous NMF studies in persons with stroke, which found fewer modules needed to explain the muscle activity of paretic limbs. Our autoencoder-based method offers a new approach for computing motor modules, with advantages of better stability in module structure across module counts, and a more biomechanically relevant interpretation of VAF 1 .

Objective: Real-time determination of human kinematics and kinetics could advance biomechanics research and enable valuable applications of biofeedback and generalizable exoskeleton control. This work aims to investigate a taskindependent, user-independent method for obtaining precise realtime joint state estimation across lower-body joints during a wide variety of tasks. Methods: We developed a generalizable sensing approach using a suit comprised of inertial measurement units (IMUs) and pressure insoles. With the suit, we collected a dataset of 33 tasks commonly performed during construction and hazardous waste cleanup (N = 10). We then trained deep learning user-independent, task-agnostic models to estimate joint lowerbody kinematics and dynamics using only worn sensor data. We likewise computed joint kinematics and dynamics analytically from sensor data to serve as a comparison tool for model results. Results: Our models achieved overall angle estimation root-meansquared-errors (RMSE) of 6.56±.92°, 8.60±1.01°, 7.58±.89°, and 6.00±.73° compared to 13.9±.1.3°, 15.31±1.0°, 10.76±.70°, and 7.56±.48° via analytical methods at the lower back, hip, knee, and ankle, respectively. Likewise, our models achieved overall normalized moment estimation RMSEs of .207±.069 Nm/kg, .242±.044 Nm/kg, .202±.038 Nm/kg, and .193±.034 Nm/kg compared to .306±.036 Nm/kg, .407±.021 Nm/kg, 1.18 ±.022 Nm/kg, and 1.73±.071 Nm/kg via analytical methods at the lower back, hip, knee, and ankle, respectively. Conclusion: These results are comparable to other state-of-the-art wearable sensing systems, establishing deep learning as a viable sensing approach that generalizes to new users and tasks. Significance: This work shows promise for enabling accurate real-world biomechanical data collection and enhancement of biofeedback systems and wearable robot control.

Abstract In "Experiment-free exoskeleton assistance via learning in simulation", Luo et al. [1] present an ambitious framework for developing exoskeleton controllers through reinforcement learning exclusively in computer simulation. The authors report that a control policy trained on a small dataset from one subject was directly transferred to physical hardware, reducing human metabolic cost during walking, running, and stair climbing by more than any prior device. If confirmed, this would represent a major breakthrough for the field of wearable robotics and their clinical applications. However, a close examination of the published materials casts doubt on these claims. The reported experimental results violate physiological limits on the relationship between mechanical power and muscle energy use during gait 2,3,4 . The algorithmic claims are surprising and cannot be verified; in contrast with established replicability standards in machine learning 5,6 , executable code has not been made available. We conclude that the goals of this study have not yet been verifiably achieved and make recommendations for avoiding publication errors of this type in the future.

Personalizing exoskeleton assistance requires user-specific gait data across many locomotor tasks, yet collecting this data demands repeated motion capture sessions that are costly, time-intensive, and especially burdensome for clinical populations. This challenge is most acute across walking speeds, where gait changes substantially and deviates further in clinical gait. This work introduces a subject-conditioned residual diffusion framework that generates personalized lower-limb kinematics at unseen walking speeds from a subject's gait sequence at a single seen speed. Given sagittal-plane hip, knee, and ankle trajectories at a seen speed and a desired unseen speed, the model generates a residual that transforms the seen trajectory into the unseen one, using a transformer denoiser conditioned on the subject's gait and the two speeds through feature-wise linear modulation. Trained only on able-bodied data, the model achieved a mean absolute error (MAE) of 3.4° on held-out able-bodied subjects. Without any stroke-specific fine-tuning, it achieved a 6.0° MAE on out-of-training-distribution stroke subjects, retaining subject identity for clinical gait. The framework reduced the MAE by over 70

Lower-limb exoskeletons hold tremendous potential for augmenting human mobility across diverse applications, from assisting individuals with ageing-related mobility decline and disabilities to preventing workplace injuries in industrial settings. While human-in-the-loop optimization has driven many successful task-specific demonstrations, traditional control paradigms relying on discrete task classification face fundamental limitations: human movement is inherently continuous and infinitely variable, making task-specific approaches intractable for real-world deployment. This Perspective examines the shift towards end-to-end, artificial-intelligence-driven control systems that leverage real-time estimates of physiological state—specifically, biological joint moments. We discuss both recent advances and the key remaining challenges, including the need to optimize these systems, incorporate safety mechanisms and reduce the burden of collecting a suitable training dataset. Addressing these challenges while maintaining the flexibility and robustness of data-driven approaches will determine whether next-generation exoskeleton systems can meaningfully augment human mobility across diverse populations, environments and devices. Shepherd et al. propose end-to-end AI control of lower-limb exoskeletons based on real-time estimates of physiological signals.

Individuals post-stroke with hemiparetic gait often experience impaired mobility, which can be restored using powered exoskeletons. Specifically, hip exoskeletons can deliver positive work to the leg to help initiate leg swing; however, it is unclear which limb assistance strategy (e.g., unilateral or bilateral assistance) best improves walking performance due to the asymmetric characteristics of hemiparetic gait. Therefore, the purpose of this study was to investigate how different hip exoskeleton limb assistance strategies affect lower-limb joint biomechanics during walking for individuals post-stroke. We hypothesized that bilateral assistance would best improve late stance paretic leg orientation and swing initiation, unilateral paretic limb assistance would improve interlimb asymmetries and unilateral nonparetic limb assistance would indirectly improve late stance kinematics. We found bilateral assistance significantly increased paretic hip flexion angle, providing an improvement in leg swing initiation, and improved ankle joint work symmetry, suggesting an improvement in propulsion mechanics. Unilateral paretic limb assistance significantly increased paretic knee flexion during swing, which is a common rehabilitative target to improve stiff-knee gait for individuals post-stroke, and significantly improved joint work symmetry between limbs, which is beneficial for individuals who have increased reliance on the nonparetic limb. Unilateral nonparetic assistance did not improve late stance kinematics. These results provide insight into joint-level responses of individuals post-stroke and can help tailor hip exoskeleton assistance for individuals post-stroke based on rehabilitation targets.

Natural selection has produced diverse vision systems, from simple patches of photoreceptors to complex camera eyes, representing just one set of evolutionary outcomes. Computational evolution offers a way to systematically test hypotheses, isolate individual factors, and ask the "why" questions behind vision. We recreate vision evolution by coevolving eyes and behaviors in embodied agents and use this to illuminate principles shaping vision across different levels of the Marr's hierarchy. This leads to three key findings: First, we provide computational evidence that task-specific selection drives bifurcation in eye evolution. Second, we reveal how optical innovations naturally emerge to resolve fundamental trade-offs between light collection and spatial precision. Third, we uncover scaling laws between visual acuity and neural processing that provide insights into long-standing hypothesis behind eye and brain size. Our work introduces a paradigm that uses embodied artificial intelligence (AI) as hypothesis-testing machines that can help accelerate discoveries in vision science.

Robotic and biofeedback-assisted interventions are promising alternatives to surgical intervention and supplements for traditional physical therapy for children with gait impairments. This work utilizes a human-in-the-loop optimization strategy to adaptively modulate parameters for a lightweight robotic knee exoskeleton and biofeedback video game to maximize learning potential following the challenge point framework. We tested our approach on three able-bodied participants and one pediatric patient with genu recurvatum, a common walking pattern in children with neurological injuries. We implement a Covariance Matrix Adaptation-Evolutionary Strategy (CMA-ES) optimizer to enforce a target success rate of 70 % by continuously adjusting visual biofeedback and exoskeleton assistance parameters. Our experimental results demonstrate the system's ability to maintain the target challenge level for the pediatric participant. Stance hyperextension decreased significantly from pre- to post-training trials on day $2\left(9.2^{\circ}\right)$ and $3\left(3.2^{\circ}\right)$ of the case study. Swing flexion approached the clinical target of 65° by the end of the third day. The promising optimizer performance and changes in gait kinematics validate the feasibility of autonomous parameter tuning to maximize learning potential in pediatric gait rehabilitation.

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.

Deploying machine vision for wearable robot control faces challenges in terms of usability, reliability, privacy, and costs.

Ankle exoskeletons have the potential to augment mobility, but control strategies have largely failed to seamlessly adapt to changes in the locomotion task. Here, we introduce a multi-headed network that predicts gait speed, ground incline, stance/swing transitions, and percent stance. These predictions are mapped to exoskeleton torque using typical biological torques as a guide. The model was trained on 9 subjects walking/jogging for 12 minutes across a range of speeds and inclines. The controller was validated on 4 subjects, and achieved stance phase prediction error of 3.4% across a range of speeds and inclines, both inside and outside the training set distribution. A secondary analysis showed similar accuracy could have been obtained with only 10% of the collected data, suggesting researchers may need fewer total strides of training data, provided the data is sufficiently diverse across users and tasks. Metabolic cost was improved during running compared to wearing the exoskeleton powered off, but was beneficial for only one subject during level walking and ramp ascent when compared to no exoskeleton. Overall, our controller smoothly adapted to time-varying inclines and walking/jogging speeds, and achieved high accuracy with a reduced training dataset, though larger torque magnitudes may be required to see metabolic benefit.

A primary challenge in continual learning (CL) for wearable robotics, especially prosthetics, is balancing the need to retain learned knowledge (stability) with the necessity to adapt to new information (plasticity). This balance is crucial for online adaptation, enabling systems to transition between tasks without losing prior knowledge. In this paper, we introduce a novel online optimizer-based framework designed to manage the stability-plasticity balance through strategic datapoint replay and learning-rate adjustments of a deep neural network. We applied this framework to speed estimation systems for transfemoral prostheses (TFA users), conducting offline validation tests using data from 10 individuals with TFA, and online tests with three TFA and six able-bodied (AB) participants. Our results demonstrate statistically significant improvements: in offline settings, our method showed a 39.2% increase in stability and a 35.2% boost in plasticity over traditional CL approaches during leave-one-subject-out validation. Similarly, in real-time trials with AB participants, we observed statistically significant gains in handling both previously encountered and new walking speeds. Finally, trials with individuals with TFA showed that the system improved the plasticity of the baseline model by 67.45% and the stability of the traditional CL approach by 31.36%; reducing overall average walking speed estimation error by 19.47%.

This study introduces a novel continual learning algorithm that incrementally improves the performance of deep-learning-based walking speed estimators during level-ground walking with a powered knee-ankle prosthesis. While user-dependent (DEP) estimators generally outperform user-independent (IND) estimators, they require the pre-collection of DEP training data. In contrast, our real-time algorithm adapts IND estimators to self-labeled DEP data generated during walking, eliminating the need for pre-collected datasets. The algorithm also features a biomimetic scaling mechanism that adjusts prosthetic assistance based on speed estimates. We evaluated our algorithm on novel subjects (N=10) with unilateral above-knee amputations during treadmill and overground walking. For treadmill trials, when adapted with estimated and ground truth labels, estimators achieved mean absolute errors (MAEs) of 0.074 0.023 (mean, standard deviation) and 0.074 0.018 m/s, respectively, reflecting a significant 28% (p < 0.05) reduction in MAE compared to non-adapted estimators. For overground trials, treadmill-adapted estimators demonstrated a significant 18% (p < 0.05) reduction in MAE compared to non-adapted estimators. Our algorithm significantly reduced speed estimation errors within one minute of walking and delivered biomimetic assistance (r = 0.91) across speeds. This approach allows off-the-shelf powered prostheses to seamlessly adapt to new users, delivering biomimetic assistance through precise, real-time walking speed estimation.

Accurate walking speed estimation in lower-limb prostheses is crucial for delivering biomechanically appropriate assistance across varying speeds. However, training robust models requires extensive domain-specific, user-dependent (DEP) data, which is impractical for every new prosthesis user. This study presents a transfer learning framework to simplify and enhance the training process. Convolutional neural networks were pre-trained on publicly available datasets from able-bodied (AB) individuals and transfemoral amputees using the Open Source Leg (OSL) knee-ankle prosthesis, then fine-tuned with data from a transfemoral amputee using the Power Knee (PK) prosthesis. The fine-tuned models, AB-PK and OSL-PK were trained with varying data amounts and evaluated across constant and variable walking speed trials, with performance compared to DEP models trained from scratch on PK data. Training and testing were conducted on a per-subject basis, with performance averaged across subjects (N=7). The lowest post-fine-tuning error was observed in AB-PK, with RMSE values of 0.041 m/s for constant speeds, 0.072 m/s for variable speeds, and 0.088 m/s for novel speeds not included in the original training data. Significant error reductions were observed in both fine-tuned models compared to DEP when fewer than 30 gait cycles per speed of training data were available. Notably, AB datasets appeared highly viable for this application and may even outperform OSL datasets in transfer learning for walking speed estimation, perhaps due to the much larger original training dataset. This approach highlights the potential of transfer learning across different subject populations and devices, offering insights into the data needed to achieve state-of-the-art speed estimation.

3D surface reconstruction is essential across applications of virtual reality, robotics, and mobile scanning. However, RGB-based reconstruction often fails in low-texture, low-light, and low-albedo scenes. Handheld LiDARs, now common on mobile devices, aim to address these challenges by capturing depth information from time-of-flight measurements of a coarse grid of projected dots. Yet, these sparse LiDARs struggle with scene coverage on limited input views, leaving large gaps in depth information. In this work, we propose using an alternative class of "blurred" LiDAR that emits a diffuse flash, greatly improving scene coverage but introducing spatial ambiguity from mixed time-of-flight measurements across a wide field of view. To handle these ambiguities, we propose leveraging the complementary strengths of diffuse LiDAR with RGB. We introduce a Gaussian surfel-based rendering framework with a scene-adaptive loss function that dynamically balances RGB and diffuse LiDAR signals. We demonstrate that, surprisingly, diffuse LiDAR can outperform traditional sparse LiDAR, enabling robust 3D scanning with accurate color and geometry estimation in challenging environments.

Fall incidents due to slips are some of the most common causes of injuries for industry workers and older adults, motivating research to assist balance recovery following slips. To assist balance recovery during a slip, a detection algorithm that can work with an assistive device, such as an exoskeleton, needs to be able to detect slips rapidly after onset, which remains a critical gap in the field. Here, we compared the ability of linear discriminant analysis (LDA), extreme gradient boosting (XGBoost), and convolutional neural networks (CNN) to detect slip using only native sensors on a hip exoskeleton. We trained and evaluated user-independent models on early-stance (ES) and late-stance (LS) slips of various magnitudes collected through treadmill-based slips. All models, except LDA with LS slips, detected slips with >90% accuracy. Overall, he best model was XGBoost, with its fastest results achieving average detection times and median accuracies of 155.06 ms at 96.25% for ES slips and 228.88 ms at 93.75% for LS slips, while also achieving 100% sensitivity at 195.64 ms (ES) and 266.24 ms (LS). Our results indicate a promising direction for further research into designing a generalizable model for balance recovery during slip perturbations using robotic hip exoskeletons.

Lower limb exoskeletons show promise to assist human movement, but their utility is limited by controllers designed for discrete, predefined actions in controlled environments, restricting their real-world applicability. We present an uncertainty-aware control framework that enables ankle exoskeletons to operate safely across diverse scenarios by automatically disengaging when encountering unfamiliar movements. Our approach uses an uncertainty estimator to classify movements as similar (in-distribution) or different (out-of-distribution) relative to actions in the training set. We evaluated three architectures (model ensembles, autoencoders, and generative adversarial networks) on an offline dataset and tested the strongest performing architecture (ensemble of gait phase estimators) online. The online test demonstrated the ability of our uncertainty estimator to turn assistance on and off as the user transitioned between in-distribution and out-of-distribution tasks (F1: 89.2). This new framework provides a path for exoskeletons to safely and autonomously support human movement in unstructured, everyday environments.

Repetitive manual labor tasks involving twisting, bending, and lifting commonly lead to lower back and knee injuries in the workplace. To identify tasks with high injury risk, we recruited N = 9 participants to perform industry-relevant, 2-handed lifts with a 11-kg weight. These included symmetrical/asymmetrical, ascending/descending lifts that varied in start-to-end heights (knee-to-waist and waist-to-shoulder). We used a data-driven musculoskeletal model that combined force and motion data with a muscle activation-informed solver (OpenSim, CEINMS) to estimate 3-dimensional internal joint contact forces (JCFs) in the lower back (L5/S1) and knee. Symmetrical lifting resulted in larger peak JCFs than asymmetrical lifting in both the L5/S1 (+20.2% normal [P < .01], +20.3% shear [P = .001], +20.6% total [P < .01]) and the knee (+39.2% shear [P = .001]), and there were no differences in peak JCFs between ascending versus descending motions. Below-the-waist lifting generated significantly greater JCFs in the L5/S1 and knee than above-the-waist lifts (P < .01). We found a positive correlation between knee and L5/S1 peak total JCFs (R2 = .60, P < .01) across the task space, suggesting motor coordination that favors sharing of load distribution across the trunk and legs during lifting.