Physiology-informed layered sensing for intelligent human-exoskeleton interaction

arXiv:2508.12157 · eess.SY, cs.SY · Submitted 2025-08-16 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "Physiology-informed layered sensing for intelligent human-exoskeleton interaction".

Dev: Here's the gist of the paper. Wearable exoskeletons hold transformative promise for restoring mobility across diverse users with muscular weakness or other impairments <ref:2508.12157#pg2> Here, we present a soft, lightweight smart leg sleeve that achieves anatomically aligned, layered multimodal sensing by integrating textile-based surface electromyography (sEMG) electrodes, ultrasensitive textile strain sensors, and inertial measurement units (IMUs) <ref:2508.12157#pg2> Together, these sensors provide real-time perception to support three core objectives: controlling personalized assistance, optimizing user effort, and safeguarding against injury risks <ref:2508.12157#pg2> This work establishes a physiology-aligned sensing architecture that reframes exoskeleton perception from motion tracking to real-time physiological decoding, offering a pathway towards intelligent, adaptive, and personalized wearable robotics <ref:2508.12157#pg2>.

Rosa: Pruning redundant reasoning tokens The core problem addressed is the structural imbalance where exoskeleton intelligence advances on the actuation side while a disconnect remains on the perceptive functions linked to the wearer physiological parameters <ref:2508.12157#pg3> Existing systems typically address these elements in isolation, and what is lacking is a compact, deployable, physiology-aligned sensing architecture capable of perceiving across the body’s layers and closing appropriate feedback loops <ref:2508.12157#pg4>.

Dev: The proposed system functions as a form of physiological tomography that decodes the body's state layer by layer through a smart textile leg sleeve <ref:2508.12157#pg4>.

Rosa: This layered sensing platform enables three fundamental tasks in the control of assistive wearable exoskeleton devices supporting real-time adaptation in personalized mobility assistance, reduction of the overall metabolic energy consumption, and safeguarding the users against unwanted actions or excessive strain at the human-exoskeleton interface <ref:2508.12157#pg4>.

Paper summary: Rosa: We talked about the core thesis of this paper—that the gap in wearable robotics is the lack of perception linked to wearer physiology—and how they propose a layered sensing architecture to solve that. This paper, "Physiology-informed layered sensing for intelligent human-exoskeleton interaction," argues that exoskeletons need to move beyond just mechanical tracking and start decoding biological parameters directly from the skin.

Dev: They claim this is achieved by integrating textile-based surface electromyography electrodes, ultrasensitive textile strain sensors, and inertial measurement units into a single sleeve. This setup allows for anatomically aligned, layered multimodal sensing that provides real-time perception to support three main goals: controlling personalized assistance, optimizing user effort, and safeguarding against injury risks.

Taro: So the methodology they used involves a three-phase experimental protocol to validate these decoders. Task one focused on joint moment estimation in a biomechanics lab under zero-torque mode, training the decoder with both electromyography signals and inertial signals.

Rosa: Then Task two took them into real-world conditions, monitoring metabolic trends during overground walking with active assistance, where they defined the ground truth label as the relative change in metabolic rate between two specific time windows.

Dev: And finally, Task three targeted injury risk detection by injecting random low-magnitude torque pulses and classifying risk states using only the unilateral strain signals.

Taro: What I find interesting is how they approach that data processing for each task. For joint moment estimation, they used a temporal convolutional network with causal and dilated layers for the IMU input.

Rosa: And for metabolic trend classification, they used bilateral electromyography and inertial signals with kernel sizes adjusted to reflect the slower dynamics of metabolic changes in that type of measurement.

Dev: And for risk detection, they employed a one-dimensional ResNet encoder trained with focal loss to really emphasize those rare emergency events, which allowed them to detect risks within one hundred milliseconds <ref:2508.12157#pg3>.

Taro: What the numbers show is that when you combine these approaches—the IMU and EMG for moment estimation, and the different signals for metabolic tracking—you get a very low ankle moment root mean square error of zero point one three Newton meters per kilogram in the leave-one-subject-out protocol <ref:2508.12157#pg3,leave-one-subject-out protocol>.

Rosa: And the metabolic classification accuracy is quite high at ninety-seven point one percent under those real-world conditions, showing that the system isn't just working in a controlled lab setting but holds up when things get messy <ref:2508.12157#pg3>.

Dev: They also showed that the risk detection model achieved a recall of zero point nine six under leave-one-subject-out validation for the injury prediction task, which is significant for any safety-critical application <ref:2508.12157#pg1>.

Taro: The paper’s main claim is that by structuring the reasoning between these layers—joint moment estimation, metabolic trend classification, and risk detection—they've reframed exoskeleton perception from simple motion tracking to real-time physiological decoding.

Rosa: So the fundamental reason this matters is that it establishes a physiology-aligned sensing architecture that offers a pathway toward intelligent, adaptive, and personalized wearable robotics by making the wearer's internal state central to the control logic.

Conclusion: Dev: So wrapping up this discussion on "Physiology-informed layered sensing for intelligent human-exoskeleton interaction," the authors are essentially presenting a system that moves exoskeletons from being purely motion-aware devices to being physiologically aware ones.

Rosa: They show that by combining different sensor modalities—muscle activity, strain, and inertia—and structuring the reasoning between them into task-specific decoders, you can achieve a level of concurrent physiological loop closure that is feasible and beneficial across different testing conditions.

Taro: The authors highlight that this unified architecture allows for mapping internal human states like skeletal motion, muscular activation, and skin deformation onto dedicated sensing modalities.

Dev: This means they've created a coherent framework that unifies the physiological inputs and the control logic into one system designed to enhance autonomy, safety, and personalization across various commercial and research-grade wearable robotics.

Rosa: The real implication for us is that this sensing logic is transferable; it provides a foundation for future technologies like upper-limb prosthetic devices with effort-aware control or even general-purpose somatosensory skins for robotic systems needing human perception.

Taro: That means the boundary between biological and artificial systems becomes more porous, requiring compact, responsive, and physiologically grounded sensing platforms to work together effectively.

Department of Engineering, University of Cambridge · Department of Electronics Systems, Aalborg University · Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong · School of Instrumentation and Optoelectronic Engineering, Beihang University · Department of Electrical and Electronic Engineering, The University of Hong Kong SAR · Department of Data and System Engineering, The University of Hong Kong · Department of Materials and Production, Aalborg University

eess.SY, cs.SY

Submitted: 2025-08-16

Updated: 2025-10-16

Comments: 21 pages, 5 figures, 43 references

DOI: 10.1002/sys3.70049

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 98/100

The gist: Wearable exoskeletons hold transformative promise for restoring mobility across diverse users with muscular weakness or other impairments <ref:2508.12157#pg2> Here, we present a soft…

Key concepts

Layered Multimodal Sensing
This involves combining different types of sensors—like muscle activity (sEMG), skin deformation (strain sensors), and motion data (IMUs)—into a single sleeve. This allows the system to perceive the body's state across multiple physical layers simultaneously, providing a comprehensive view of human physiology rather than just tracking movement.
Physiology-Aligned Sensing Architecture
This is a design approach where the sensing hardware and software are structured around specific physiological needs: joint moment control, energy optimization, and safety. Instead of just measuring motion, the system aims to decode internal human states like muscle force and metabolic load to intelligently guide the exoskeleton's assistance.
Physiological Tomography
The system functions as a form of tomography that maps the body's internal state layer by layer using wearable sensors. It decodes complex bodily conditions—such as skeletal motion, muscular activation patterns, and skin deformation—into actionable data streams for real-time control and personalized assistance.
Real-Time Physiological Decoding
This is the core innovation where raw sensor signals are processed by neural decoders to translate them into meaningful physiological metrics. For instance, sEMG signals are used to estimate muscle force, and strain sensors detect skin deformation, allowing the system to perceive the user's internal effort and state instantly.

Terminology

Summary

Wearable exoskeletons hold transformative promise for restoring mobility across diverse users with muscular weakness or other impairments <ref:2508. p. 0

Where the paper says it

Occhipinti (lgo23@cam. uk) Abstract Wearable exoskeletons hold transformative promise for restoring mobility across diverse users with muscular weakness or other impairments.

12157#pg2> Here, we present a soft, lightweight smart leg sleeve that achieves anatomically aligned, layered multimodal sensing by integrating textile-based surface electromyography (sEMG) electrodes, ultrasensitive textile strain sensors, and inertial measurement units (IMUs) <ref:2508. p. 0
Where the paper says it

However, their translation beyond laboratory environments remains limited by sensing systems that capture movement but not underlying physiology. Here, we present a soft, lightweight smart leg sleeve that achieves anatomically aligned, layered multimodal sensing by integrating textile-based surface electromyography (sEMG) electrodes, ultrasensitive textile strain sensors, and inertial measurement units (IMUs).

12157#pg2> Together, these sensors provide real-time perception to support three core objectives: controlling personalized assistance, optimizing user effort, and safeguarding against injury risks <ref:2508. p. 0
Where the paper says it

Each sensing modality targets a distinct physiological layer: IMUs track joint kinematics at the skeletal level, sEMG monitors muscle activation at the muscular level, and strain sensors detect skin deformation at the cutaneous level. Together, these sensors provide real-time perception to support three core objectives: controlling personalized assistance, optimizing user effort, and safeguarding against injury risks.

12157#pg2> This work establishes a physiology-aligned sensing architecture that reframes exoskeleton perception from motion tracking to real-time physiological decoding, offering a pathway towards intelligent, adaptive, and personalized wearable robotics <ref:2508. p. 0
Where the paper says it

96), all validated on unseen users using a leave-one-subject-out protocol. This work establishes a physiology-aligned sensing architecture that reframes exoskeleton perception from motion tracking to real-time physiological decoding, offering a pathway towards intelligent, adaptive, and personalized wearable robotics.

12157#pg2>.

Pruning redundant reasoning tokens

The core problem addressed is the structural imbalance where exoskeleton intelligence advances on the actuation side while a disconnect remains on the perceptive functions linked to the wearer physiological parameters <ref:2508.12157#pg3> Existing systems typically address these elements in isolation, and what is lacking is a compact, deployable, physiology-aligned sensing architecture capable of perceiving across the body’s layers and closing appropriate feedback loops <ref:2508. p. 1

Where the paper says it

Joint moment estimation is often hidden within non-interpretable mappings, metabolic monitoring depends on cumbersome respiratory setups, and injury detection relies on fixed force thresholds that ignore individual context. What is lacking is a compact, deployable, physiology-aligned sensing architecture capable of perceiving across the body’s layers and closing appropriate feedback loops.

12157#pg4>. The proposed system functions as a form of physiological tomography that decodes the body's state layer by layer through a smart textile leg sleeve <ref:2508.12157#pg4>. This layered sensing platform enables three fundamental tasks in the control of assistive wearable exoskeleton devices supporting real-time adaptation in personalized mobility assistance, reduction of the overall metabolic energy consumption, and safeguarding the users against unwanted actions or excessive strain at the human-exoskeleton interface <ref:2508. p. 1
Where the paper says it

Here we introduce a unified, physiology-layered perception architecture that functions as a form of physiological tomography that decodes the body's state layer by layer, implemented through a smart textile leg sleeve. The leg sleeve-integrated sensors suite serves as an essential companion device seamlessly integrated in smart garment through a scalable printing technique, enabling 3 fundamental tasks in the control of assistive wearable exoskeleton devices supporting: a) real-time adaptation in personalized mobility assistance, b) reduction of the overall metabolic energy consumption, and c) safeguarding the users against unwanted actions or excessive strain at the human-exoskeleton interface.

12157#pg4>.

How the training signal is built

The system employs a three-phase experimental protocol to support and validate real-time neural decoders for three core perceptual objectives <ref:2508.12157#pg5>. Task I focuses on joint moment estimation in a biomechanics laboratory under zero-torque mode, where the decoder is trained using both sEMG and IMU signals <ref:2508. p. 3

Where the paper says it

Task I: Joint moment estimation in a biomechanics laboratory To develop a decoder for estimating ankle joint moments in real time, we first conducted experiments under controlled laboratory conditions (Fig. S4) with the exoskeleton operating in a zero-torque mode.

12157#pg5>. Task II involves real-world metabolic trend monitoring during overground walking with active assistance, where the ground-truth label is computed as the relative change in metabolic rate between two specific time windows <ref:2508. p. 4
Where the paper says it

The corresponding ground-truth label was computed as the relative change in metabolic rate between the 6-second window preceding t (t-6 s to t) and a forward-offset window centered at t + 27 s (t+24 s to t+30 s). This 30-second forward offset accounts for the physiological delay in oxygen uptake kinetics during moderate-intensity walking [25, 26], enabling each label to reflect a meaningful metabolic trend.

12157#pg6>. Task III targets injury risk detection by injecting random low-magnitude torque pulses and classifying risk states from unilateral strain signals <ref:2508. p. 5
Where the paper says it

walked under zero-torque mode while experiencing random, low-magnitude torque pulses (<0. 3 Nm/kg) applied during early (0-20%) and late (80-100%) gait cycle phases (Supplementary Note 3).

12157#pg7>.

Where the gains come from

The performance evaluation demonstrates that concurrent closure of multiple physiological loops is feasible and beneficial, achieving an ankle moment RMSE of 0.13 Nm/kg, a real-time metabolic trend classification accuracy of 97. p. 2

Where the paper says it

Across users and conditions, the system achieves an ankle moment RMSE of 0. 13 Nm/kg, a real-time metabolic trend classification accuracy of 97.

1%, and injury risk detection within 100 ms with a recall of 0. p. 0
Where the paper says it

13 Nm/kg), (2) real-time classification of metabolic trends (accuracy = 97. 1%), and (3) injury risk detection within 100 ms (recall = 0.

96 <ref:2508.12157#pg4>. The joint moment decoder leverages the complementary strengths of IMUs for phase timing and sEMG for amplitude-level variations in muscle force output <ref:2508. p. 3
Where the paper says it

The joint moment decoder was trained using both sEMG and IMU signals as inputs (Fig. 2, top right), leveraging the complementary strengths of the two modalities: IMUs provided phase timing and gait structure, while sEMG refined amplitude-level variations in muscle force output.

12157#pg8>. For metabolic trend classification, the results show that sEMG signals reveal clear metabolic state separability more clearly than IMU features <ref:2508. p. 4
Where the paper says it

The importance of modality-specific sensing is further illustrated in Fig. 3d, where t-SNE embeddings of extracted features show that metabolic state distinctions emerge more clearly from sEMG signals than from IMU, reinforcing the critical role of physiologically grounded inputs in decoding user effort.

12157#pg5>. The risk detection model achieved 96. p. 6
Where the paper says it

This contrasts with joint moment estimation, where IMU carried more predictive power—highlighting task-specific modality relevance. Risk detection model.

4% recall under LOSO validation, demonstrating the safety value provided by skin-deformation monitoring <ref:2508. p. 5
Where the paper says it

3e), demonstrating the system’s ability to detect biomechanical anomalies based on cutaneous-level sensing. This task highlights the additional safety value provided by skin-deformation monitoring, complementing traditional force-threshold-based approaches that may overlook subtle shifts too weak to trigger traditional safety thresholds.

12157#pg7>.

System-level performance and user evaluation

The system's usability was assessed through a user study where responses indicated high perceived accuracy, responsiveness, comfort, and usability of the layered sensing platform <ref:2508.12157#pg9>. The effort level is computed from sEMG signals using classical root-mean-square extraction and normalized to each subject’s maximum observed amplitude during baseline walking trials <ref:2508. p. 6

Where the paper says it

We opted to present effort level as a single intuitive percentage rather than raw torque curves, which may be less interpretable for end users. The effort level is computed from sEMG signal features using classical root-mean-square extraction, not as a definitive physiological metric but as an interpretable proxy for front-end visualization, chosen for its simplicity and real-time robustness.

12157#pg14>. This metric served as a front-end proxy for user effort, chosen for its simplicity and real-time responsiveness <ref:2508. p. 6
Where the paper says it

The effort level is computed from sEMG signal features using classical root-mean-square extraction, not as a definitive physiological metric but as an interpretable proxy for front-end visualization, chosen for its simplicity and real-time robustness. We further conducted a user study (n = 12) using a standardized Likert-scale questionnaire (Fig.

12157#pg9>. The framework unifies physiological inputs and control logic into a coherent architecture that enhances autonomy, safety, and personalization across both commercial and research-grade wearable robotics <ref:2508. p. 7
Where the paper says it

For instance, torque estimation can inform low-level control laws such as impedance modulation [8], metabolic classification can guide adaptive human-in-the-loop optimization [29], and risk detection can trigger immediate safety interventions. In doing so, the framework unifies physiological inputs and control logic into a coherent architecture that enhances autonomy, safety, and personalization across both commercial and research-grade wearable robotics.

12157#pg10>.

This study presents a human-exoskeleton interaction framework that places real-time physiological perception—rather than purely kinematic sensing—at the core of intelligent assistance <ref:2508.12157#pg4>. By embedding sEMG, Strain and inertial motion sensors seamlessly in smart e-textile leg sleeves the proposed system enables mapping internal human states—skeletal motion, muscular activation, and skin deformation—onto dedicated sensing modalities <ref:2508. p. 7

Where the paper says it

Discussion This study presents a human-exoskeleton interaction framework that places real-time physiological perception—rather than purely kinematic sensing—at the core of intelligent assistance. By embedding sEMG, Strain and inertial motion sensors seamlessly in smart e-textile leg sleeves the proposed system enables mapping internal human states—skeletal motion, muscular activation, and skin deformation— onto dedicated sensing modalities, hence transforming the body into a structured, high-fidelity source of control information.

12157#pg4>. In doing so, the framework unifies physiological inputs and control logic into a coherent architecture that enhances autonomy, safety, and personalization across both commercial and research-grade wearable robotics <ref:2508. p. 7
Where the paper says it

For instance, torque estimation can inform low-level control laws such as impedance modulation [8], metabolic classification can guide adaptive human-in-the-loop optimization [29], and risk detection can trigger immediate safety interventions. In doing so, the framework unifies physiological inputs and control logic into a coherent architecture that enhances autonomy, safety, and personalization across both commercial and research-grade wearable robotics.

12157#pg10>.

The design follows a biomimetic principle along the bone-muscle-skin hierarchy: IMUs capture skeletal kinematics, textile sEMG electrodes record muscle activation, and printed graphene-based strain sensors detect skin deformation <ref:2508.12157#pg4>. The key novelty that results from the proposed system architecture lies not in the aggregation of sensors but in the structured reasoning between them, where each layer is coupled to a task-specific neural decoder <ref:2508. p. 2

Where the paper says it

aggregation of sensors but in the structured reasoning between them. Each layer is coupled to a task-specific neural decoder that closes a distinct physiological loop, i.

12157#pg4>. This approach positions sensing as the central principle, rather than a peripheral component, of human-adaptive assistance <ref:2508. p. 2
Where the paper says it

These results demonstrate that concurrent closure of multiple physiological loops is feasible and beneficial. Compared with the single-layer, proxy-based approaches summarized in Table 1, this work reframes exoskeleton perception as a unified, physiology-first problem, positioning sensing as the central principle, rather than a peripheral component, of human-adaptive assistance.

12157#pg4>.

The system is skin-conformal, mechanically compliant, and seamlessly integrated with a custom exoskeleton (. The textile strain sensors are screen-printed over the posterolateral region of the sleeve along the calcaneofibular ligament path (Fig. p. 2

Where the paper says it

This compact platform achieves perceptive sensing and modeling functions by i) decoding motion intent, ii) monitoring metabolic effort, and iii) detecting injury-related risk. The textile strain sensors are screen-printed over the posterolateral heel region along the calcaneofibular ligament path (Fig.

1b) <ref:2508.12157#pg4>. To enhance print quality and functional repeatability, a sodium carboxymethyl cellulose (CMC-Na) starching treatment is applied to ensure surface uniformity and prevent ink over-penetration into the fabric <ref:2508. p. 2
Where the paper says it

A structured graphene layer with ordered microcracks, formed via a single-step printing process, enables high sensitivity to localized skin deformation under small strains (<5%), with gauge factor exceeding 100 [23]. To enhance print quality and functional repeatability, we apply a sodium carboxymethyl cellulose (CMC-Na) starching treatment that ensures surface uniformity and prevents ink over-penetration into the fabric.

12157#pg10>. Textile sEMG dry electrodes are fabricated by coating cotton substrates with a graphene/PEDOT:PSS composite, providing low skin–electrode impedance comparable to wet electrodes within the sEMG bandwidth <ref:2508. p. 2
Where the paper says it

Textile sEMG dry electrodes are fabricated by coating cotton substrates with a graphene/PEDOT:PSS composite. The porous coating provides a large effective contact area, resulting in low skin–electrode impedance comparable to wet electrodes within the sEMG bandwidth (10–500 Hz) [20, 21].

12157#pg10>.

The joint moment decoder processes unilateral sEMG and IMU signals using two distinct encoder branches, where the IMU input is handled by a temporal convolutional network (TCN) with causal and dilated layers <ref:2508.12157#pg5>. The metabolic trend decoder uses bilateral sEMG and IMU signals with adjusted kernel sizes to reflect the slower dynamics of metabolic changes <ref:2508. p. 6

Where the paper says it

Metabolic trend classifier The metabolic trend decoder uses bilateral sEMG and IMU signals with the same encoder structure as the moment model (Fig. 4b, left), but with adjusted kernel sizes and receptive fields to reflect the slower dynamics of metabolic changes.

12157#pg5>. The risk detection model uses a 1D ResNet encoder trained with focal loss to emphasize rare emergency events, enabling rapid and accurate detection of risk states with sub-100 ms latency <ref:2508. p. 6
Where the paper says it

To support ultra-fast detection of biomechanical risk, we designed a binary classifier based on unilateral strain signals (Fig. The encoder uses a 1D ResNet trained with focal loss to emphasize rare emergency events.

12157#pg5>.

All sensor signals, including strain, sEMG, and IMU data, are preprocessed using standard signal conditioning techniques before being used for neural network training <ref:2508.12157#pg14>. Strain signals are processed using a fourth-order low-pass Butterworth filter with a cutoff frequency of 20 Hz to remove high-frequency noise while retaining the slow deformation patterns of the skin that are informative for gait and perturbation detection <ref:2508. p. 9

Where the paper says it

Both modules incorporated a low-power ESP32 microcontroller and an integrated 9-axis inertial measurement unit (MPU-9250) for concurrent kinematic sensing. Strain signals were read through voltage divider circuits following a design established in our previous work [42], enabling high-sensitivity detection of local skin deformation.

12157#pg14>. SEMG signals are filtered using a fourth-order Butterworth bandpass filter between 20 and 450 Hz to isolate physiologically relevant frequencies <ref:2508. p. 5
Where the paper says it

Neural decoder architecture and system-level performance evaluation To translate raw multimodal signals into actionable control cues, we developed task-specific neural decoders for each of the three physiological objectives: joint moment estimation, metabolic trend classification, and injury risk detection. Each task selectively leverages the most relevant sensing modalities based on its physiological characteristics, rather than using all three modalities in every case.

12157#pg14>. IMU signals from the accelerometer, gyroscope, and magnetometer axes are denoised using a fourth-order low-pass Butterworth filter with a 20 Hz cutoff to reduce motion-induced jitter while preserving key features of segmental kinematics <ref:2508. p. 2
Where the paper says it

All sensors are stitched into a commercially available compression sleeve, with conductive threads routing signals to two compact wireless readout modules (Fig. Each module handles one sensing modality—strain or sEMG—and includes a dedicated IMU for capturing segmental kinematics from the foot and shank, respectively.

12157#pg14>.

The exoskeleton design consists of two quasi-direct-drive motors connected by a Bowden cable system that delivers controlled plantarflexion torque to the user’s foot <ref:2508.12157#pg10>. A 500 N load cell is positioned at the interface between the cable sheath and ankle brace, enabling real-time measurement of transmitted force and torque <ref:2508. p. 9

Where the paper says it

The actuation torque is transmitted to carbon fiber footplates via lightweight cables, minimizing distal mass. A 500 N load cell (DYLY08, DYSENSOR) is positioned at the interface between the cable sheath and ankle brace, enabling real-time measurement of transmitted force and torque.

12157#pg10>. All high-level control and data logging are executed on an onboard edge processor (Jetson Nano B01, Nvidia) housed within the back module <ref:2508. p. 9
Where the paper says it

A 500 N load cell (DYLY08, DYSENSOR) is positioned at the interface between the cable sheath and ankle brace, enabling real-time measurement of transmitted force and torque. All high-level control and data logging are executed on an onboard edge processor (Jetson Nano B01, Nvidia) housed within the back module.

12157#pg10>. The mobile interface is developed using Flutter to visualize real-time sensor decoding outputs via a UDP socket at a 5 Hz refresh rate <ref:2508. p. 9
Where the paper says it

sEMG signals were amplified and digitized using a dedicated analog front-end (ADS1293), communicating with the ESP32 via SPI. All sensor data were transmitted over Wi-Fi to an on-body edge processor (Jetson Nano), where task-specific neural decoders performed real-time inference.

12157#pg14>.

The system's validation across three scenarios—biomechanics-lab gait trials, outdoor assisted walking with metabolic monitoring, and controlled perturbations simulating safety-critical events—demonstrates the feasibility of concurrent closure of multiple physiological loops <ref:2508.12157#pg4>. The results show that variations in torque scaling (α) and actuation delay (Δt) lead to large inter-condition differences in metabolic rate, reinforcing the need for adaptive control schemes based on sensing feedback <ref:2508. p. 4

Where the paper says it

As illustrated in Fig. 3c, variations in α and Δt led to large inter-condition differences in metabolic rate, with certain combinations (e.

12157#pg6>. The system's ability to detect biomechanical anomalies based on cutaneous-level sensing highlights the additional safety value provided by skin-deformation monitoring <ref:2508.12157#pg7>. Future work will focus on implementing new capabilities to update the embedded perception model in real-time with minimal calibration and without interrupting usage <ref:2508. p. 7
Where the paper says it

Third, while our decoding models support real-time inference and demonstrate strong cross-subject generalization, they still rely on offline training and manual parameter tuning. To overcome this limitation, future work will focus on implementing new capabilities to update the embedded perception model in real-time, with minimal calibration and without interrupting usage.

12157#pg10>.

The study's findings are transferable, providing a foundation for future technologies such as upper-limb prosthetic devices capable of effort-aware control and personalized rehabilitation systems that adapt based on patient progress <ref:2508.12157#pg10>. The abstraction of multimodal sensor data into interpretable indicators like torque intent, effort dynamics, and risk events could be used to populate real-time digital twins of human users <ref:2508. p. 8

Where the paper says it

Furthermore, the abstraction of multimodal sensor data into interpretable indicators such as torque intent, effort dynamics, and risk events could be used to populate real-time digital twins of human users, supporting simulation-informed interventions and long-term health tracking [37, 38]. The same sensing logic can also support the development of general-purpose somatosensory skins for robotic systems that require human-like perception [39, 40].

12157#pg10>. This sensing framework provides a transferable foundation for future technologies that require embodied intelligence <ref:2508. p. 7
Where the paper says it

As intelligent systems become increasingly embedded into the daily activities of individuals, there is growing demand for machines that can interpret internal physiological cues and not just respond to external behaviors. The sensing framework presented here provides a transferable foundation for future technologies that require embodied intelligence [36].

12157#pg10>. The same sensing logic can also support the development of general-purpose somatosensory skins for robotic systems that require human-like perception <ref:2508. p. 8
Where the paper says it

Furthermore, the abstraction of multimodal sensor data into interpretable indicators such as torque intent, effort dynamics, and risk events could be used to populate real-time digital twins of human users, supporting simulation-informed interventions and long-term health tracking [37, 38]. The same sensing logic can also support the development of general-purpose somatosensory skins for robotic systems that require human-like perception [39, 40].

p. 8
Where the paper says it

Furthermore, the abstraction of multimodal sensor data into interpretable indicators such as torque intent, effort dynamics, and risk events could be used to populate real-time digital twins of human users, supporting simulation-informed interventions and long-term health tracking [37, 38]. The same sensing logic can also support the development of general-purpose somatosensory skins for robotic systems that require human-like perception [39, 40].

p. 8
Where the paper says it

Furthermore, the abstraction of multimodal sensor data into interpretable indicators such as torque intent, effort dynamics, and risk events could be used to populate real-time digital twins of human users, supporting simulation-informed interventions and long-term health tracking [37, 38]. The same sensing logic can also support the development of general-purpose somatosensory skins for robotic systems that require human-like perception [39, 40].

p. 8
Where the paper says it

Furthermore, the abstraction of multimodal sensor data into interpretable indicators such as torque intent, effort dynamics, and risk events could be used to populate real-time digital twins of human users, supporting simulation-informed interventions and long-term health tracking [37, 38]. The same sensing logic can also support the development of general-purpose somatosensory skins for robotic systems that require human-like perception [39, 40].

12157#pg10>. The boundary between biological and artificial systems becomes increasingly porous, requiring compact, responsive, and physiologically grounded sensing platforms <ref:2508. p. 8
Where the paper says it

The same sensing logic can also support the development of general-purpose somatosensory skins for robotic systems that require human-like perception [39, 40]. As the boundary between biological and artificial systems becomes increasingly porous, sensing platforms that are compact, responsive, and physiologically grounded will play a central role in shaping the next generation of intimate and symbiotic human-machine co-adaptation, potentially contributing to future human–machine co-adaptation strategies that could enhance mobility and safety [41].

p. 8
Where the paper says it

The same sensing logic can also support the development of general-purpose somatosensory skins for robotic systems that require human-like perception [39, 40]. As the boundary between biological and artificial systems becomes increasingly porous, sensing platforms that are compact, responsive, and physiologically grounded will play a central role in shaping the next generation of intimate and symbiotic human-machine co-adaptation, potentially contributing to future human–machine co-adaptation strategies that could enhance mobility and safety [41].

p. 8
Where the paper says it

The same sensing logic can also support the development of general-purpose somatosensory skins for robotic systems that require human-like perception [39, 40]. As the boundary between biological and artificial systems becomes increasingly porous, sensing platforms that are compact, responsive, and physiologically grounded will play a central role in shaping the next generation of intimate and symbiotic human-machine co-adaptation, potentially contributing to future human–machine co-adaptation strategies that could enhance mobility and safety [41].

p. 8
Where the paper says it

The same sensing logic can also support the development of general-purpose somatosensory skins for robotic systems that require human-like perception [39, 40]. As the boundary between biological and artificial systems becomes increasingly porous, sensing platforms that are compact, responsive, and physiologically grounded will play a central role in shaping the next generation of intimate and symbiotic human-machine co-adaptation, potentially contributing to future human–machine co-adaptation strategies that could enhance mobility and safety [41].

12157#pg10>.

The paper addresses the need for a paradigm shift where exoskeletons must evolve from perceiving motion to decoding physiology <ref:2508.12157#pg3>. This shift is achieved by defining three key physiological dimensions that exoskeletons must sense: i) joint moment as the mechanical target for assistance, ii) metabolic cost as the energetic objective for optimization, and iii) injury risk as the safety constraint requiring immediate feedback <ref:2508. p. 1

Where the paper says it

As summarized in Table 1, representative systems remain single-modality and shallow in perceptual depth, excelling at phase detection or kinematic estimation but rarely accessing the layered physiological variables that govern assistance, effort, and risk. We define three key physiological dimensions that exoskeletons must sense to operate effectively in real-world environments: i) joint moment as the mechanical target for assistance, ii) metabolic cost as the energetic objective for optimization, and iii) injury risk as the safety constraint requiring immediate feedback.

12157#pg3>. The unified, physiology-layered perception architecture functions as a form of physiological tomography that decodes the body's state layer by layer <ref:2508.12157#pg4>. This architecture allows for concurrent closure of multiple physiological loops, which is feasible and beneficial across different experimental conditions <ref:2508. p. 2
Where the paper says it

1%, and injury risk detection within 100 ms with a recall of 0. These results demonstrate that concurrent closure of multiple physiological loops is feasible and beneficial.

12157#pg4>.

The results show that the IMU+EMG combination achieved the lowest RMSE (0.089 Nm/kg within-subjects, 0. p. 5

Where the paper says it

Across eight subjects, the LOSO RMSE averaged 0. 015 Nm/kg, while within-subject training achieved 0.

13 Nm/kg LOSO), consistently below 0. p. 5
Where the paper says it

Each task selectively leverages the most relevant sensing modalities based on its physiological characteristics, rather than using all three modalities in every case. All models were trained using a leave-one-subject-out (LOSO) protocol and optimized for real-time deployment on a wearable edge processor [31].

2 across all test subjects <ref:2508. p. 5
Where the paper says it

4a (right), the model accurately tracks rapid fluctuations in joint torques. Across eight subjects, the LOSO RMSE averaged 0.

12157#pg6>. The metabolic trend classifier achieved 98. p. 6
Where the paper says it

2 Nm/kg, highlighting the model’s generalization capability (Fig. Metabolic trend classifier The metabolic trend decoder uses bilateral sEMG and IMU signals with the same encoder structure as the moment model (Fig.

4% accuracy within-subject and 97. p. 6
Where the paper says it

Classification accuracy reached 98. 4% within-subject and 97.

1% under LOSO testing, with class-wise F1-scores exceeding 0. p. 6
Where the paper says it

1% under LOSO testing (Fig. 5b, top), with class-wise F1-scores exceeding 0.

97 <ref:2508.12157#pg5>. These results underscore the system’s usability, physiological transparency, and translational potential for real-world exoskeleton applications <ref:2508. p. 7
Where the paper says it

Expanding training to include slopes, stairs, and transitions will be a focus of future work. Nevertheless, these results underscore the system’s usability, physiological transparency, and translational potential for real-world exoskeleton applications.

12157#pg9>. The study's findings are transferable, providing a foundation for future technologies that require embodied intelligence <ref:2508.12157#pg10>. The same sensing logic can also support the development of general-purpose somatosensory skins for robotic systems that require human-like perception <ref:2508.12157#pg10>. The boundary between biological and artificial systems becomes increasingly porous, requiring compact, responsive, and physiologically grounded sensing platforms <ref:2508.12157#pg10>.

The paper's findings are transferable, providing a foundation for future technologies that require embodied intelligence <ref:2508.12157#pg10>. The same sensing logic can also support the development of general-purpose somatosensory skins for robotic systems that require human-like perception <ref:2508.12157#pg10>. The boundary between biological and artificial systems becomes increasingly porous, requiring compact, responsive, and physiologically grounded sensing platforms <ref:2508.12157#pg10>. The paper's findings are transferable, providing a foundation for future technologies that require embodied intelligence <ref:2508.12157#pg10>. The same sensing logic can also support the development of general-purpose somatosensory skins for robotic systems that require human-like perception <ref:2508.12157#pg10>. The boundary between biological and artificial systems becomes increasingly porous, requiring compact, responsive, and physiologically grounded sensing platforms <ref:2508.12157#pg10>.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Development of an adaptive control scheme based on real-time effort-aware classification can optimize assistive benefits on a per-user, per-condition basis, as demonstrated by results showing that real-time effort-aware classification provides a pathway to optimize assistive benefits on a per-user, per condition basis. This allows the system to dynamically adjust assistance parameters like scaling factor and actuation delay based on the user's current metabolic state.

  2. Implementation of a physiology-aligned sensing architecture can transform exoskeleton perception from motion tracking to real-time physiological decoding. This enables three core perceptual tasks: controlling personalized assistance, optimizing user effort, and safeguarding against injury risks, by mapping internal human states onto dedicated sensing modalities.

  3. Creation of a unified, physiology-layered perception architecture can function as a form of physiological tomography that decodes the body's state layer by layer. This structure provides simultaneous access to kinematic, muscular, and cutaneous layers through integrated IMUs, sEMG electrodes, and strain sensors.

  4. Integration of task-specific neural decoders allows for selective use of modalities based on physiological characteristics rather than using all three in every case. For example, the joint moment decoder uses unilateral sEMG and IMU signals, while the risk detection model focuses on unilateral strain signals.

  5. The system can achieve real-time injury risk detection within 100 ms (recall = 0.96), complementing traditional force-threshold-based approaches by detecting anomalies based on cutaneous-level sensing, as shown by the ability to detect biomechanical anomalies based on cutaneous-level sensing.

Abstract

Wearable exoskeletons hold transformative promise for restoring mobility across diverse users with muscular weakness or other impairments. However, their translation beyond laboratory environments remains limited by sensing systems that capture movement but not underlying physiology. Here, we present a soft, lightweight smart leg sleeve that achieves anatomically aligned, layered multimodal sensing by integrating textile-based surface electromyography (sEMG) electrodes, ultrasensitive textile strain sensors, and inertial measurement units (IMUs). Each sensing modality targets a distinct physiological layer: IMUs track joint kinematics at the skeletal level, sEMG monitors muscle activation at the muscular level, and strain sensors detect skin deformation at the cutaneous level. Together, these sensors provide real-time perception to support three core objectives: controlling personalized assistance, optimizing user effort, and safeguarding against injury risks. The system is skin-conformal, mechanically compliant, and seamlessly integrated with a custom exoskeleton (<20 g total sensor and electronics weight). We demonstrate: (1) accurate ankle joint moment estimation (RMSE = 0.13 Nm/kg), (2) real-time classification of metabolic trends (accuracy = 97.1%), and (3) injury risk detection within 100 ms (recall = 0.96), all validated on unseen users using a leave-one-subject-out protocol. This work establishes a physiology-aligned sensing architecture that reframes exoskeleton perception from motion tracking to real-time physiological decoding, offering a pathway towards intelligent, adaptive, and personalized wearable robotics.

Sources

Related papers