Where soft tissue artifact enters the joint torque estimation pipeline

Soft tissue artifact (STA) corrupts joint torque estimation when rigid sensors or exoskeleton attachment points move relative to the underlying skeletal segment because skin, fat, and muscle deform under load. The resulting kinematic errors propagate through inverse dynamics pipelines into spurious torque estimates that cannot be attributed to actual muscular effort. The problem is most severe at the knee and hip during the stance-to-swing transition, where ground contact forces combine with rapid angular accelerations to amplify any coupling mismatch between the rigid exoskeleton frame and the compliant biological limb.

Three interrelated problems define the challenge. First, mechanical misalignment and relative motion between the exoskeleton rigid frame and soft biological tissue introduce kinematic noise into every downstream computation. Second, passive limb dynamics — inertial, Coriolis, and gravitational torques — are difficult to separate from active muscular effort in sensor readings, so any STA-induced error in segment kinematics propagates into all three dynamic terms simultaneously. Third, the non-stationary nature of gait means pre-calibrated models degrade rapidly as walking speed, terrain, and assistive level change.

The most direct upstream intervention in the dataset is the bilevel kinematic optimization approach from Wuhan University of Technology (2022). Rather than accepting raw optical motion capture data — which is corrupted by STA — the authors iteratively optimize marker positions to recover true skeletal kinematics before any force estimation occurs. The human-exoskeleton binding interface is then modelled as a spring to estimate interaction forces. This work explicitly acknowledges that conventional kinematic data are “problematic” without preprocessing correction, making it the most upstream engineering response to STA in the literature surveyed.

Sogang University’s 2015 study on the EXOwheel platform with hip and knee torque sensors demonstrated that without accurate subtraction of dynamic effects — specifically user-specific inertial parameter identification — joint torque sensors read a composite signal that cannot be attributed to muscle effort alone. This finding established user-specific dynamic parameter identification as a prerequisite for any STA correction strategy that relies on inverse dynamics.

EMG fusion and neuromechanical models: bypassing the artifact at source

The most physiologically grounded route to eliminating STA error is to abandon rigid-body kinematic estimation entirely and instead estimate joint torques directly from electromyographic (EMG) signals fused with joint angle measurements — because EMG signals originate from the muscle itself rather than from external markers susceptible to artifact. The University of Twente (2022) demonstrated this approach with person-specific neuromechanical models that estimate biological ankle joint torques in real-time from measured EMGs and joint angles across six “unseen” walking conditions, with a low-level disturbance observer translating biological torque estimates into exoskeleton commands.

However, EMG-based approaches carry their own non-stationarity problem: EMG signals change as a function of the exoskeleton’s assistive level, causing torque prediction models trained at one assistance level to fail at others. The University of Utah (2021) addressed this directly by training a convolutional neural network (CNN) on diverse multi-level assistance data from three healthy participants. The study showed that diverse training data spanning different levels of exoskeleton assistance enables robust torque predictions across conditions, and that the CNN architecture outperforms simpler regressors. This coupling between the STA problem and the exoskeleton’s own actuation effects on the EMG signal is a critical insight: the act of providing assistance alters the very signal used to estimate how much assistance is needed.

“EMG signals change as a function of the exoskeleton’s assistive level, causing torque prediction models trained at one assistance level to fail at others — the act of providing assistance alters the very signal used to estimate how much assistance is needed.”

According to biomechanical research standards published by IEEE, the integration of multi-modal physiological sensing with model-based control represents a recognized best practice for human-robot interaction systems. The neuromechanical approach from the University of Twente aligns with this framework by combining EMG-derived muscle activation estimates with joint kinematics to produce a biologically consistent torque signal that is structurally immune to STA. The remaining challenge — EMG non-stationarity under varying assistance — is precisely what the CNN-based approach from the University of Utah addresses through data diversity during training.

Observer-based and sensorless torque estimation: eliminating the tissue-contact requirement

Where sEMG sensors are impractical — due to sweat, electrode drift, or patient compliance — Extended State Observer (ESO)-based methods provide a structurally different route to STA immunity: they eliminate the physical coupling between soft tissue and rigid sensor altogether. The modified ESO-based integral sliding mode controller (MESOISMC) from Chulalongkorn University (2023) treats unknown human muscular torque as a generalized disturbance state and reconstructs it from measurable system states alone, with no mechanical contact sensing at the limb-exoskeleton interface required. Because no tissue-contact measurement is taken, STA cannot enter the estimation pipeline.

National Taiwan Normal University has published consistently on model-free Linear Extended State Observer (LESO)-based control across both hip exoskeleton (2021) and hip-knee exoskeleton (2022) platforms. In both cases, the LESO treats user limb torques as lumped disturbances estimated online, decoupling the controller from STA-dependent model assumptions. Unlike model-based observers that require accurate dynamic parameters of the human limb — parameters that are themselves contaminated by STA when identified from kinematic data — the LESO avoids this circularity by treating the entire human contribution as an unknown disturbance to be estimated rather than a known quantity to be modelled.

Research standards from ISO on human-robot collaboration systems highlight disturbance observer architectures as a validated approach for handling unknown external forces in compliant human-machine interaction. The LESO implementations from National Taiwan Normal University operationalize this principle for lower-limb exoskeleton gait assistance, providing a practical sensorless torque estimation pathway that is deployable even when surface EMG is unavailable.

Deep neural networks and admittance frameworks: learning and sidestepping the nonlinear coupling

Where physics-based models cannot capture the nonlinear coupling between soft tissue, exoskeleton structure, and sensor readings, deep neural networks have demonstrated strong performance by learning the full passive dynamics of the exoskeleton as a multibody system without constraining assumptions. The University of Waterloo (2023) trained multiple neural network architectures on chirp-excited Indego exoskeleton dynamics, with the optimal architecture — a deep neural network with 250 neurons and 10 time-delays — achieving an RMSE of 1.23 on Z-normalized torques and an adjusted R² of 0.89. By learning across gait phases and interaction states without phase-dependent switching, the model implicitly captures soft tissue-induced coupling that deterministic models miss.

An architecturally different philosophy sidesteps the STA problem by reformulating the control objective to eliminate the need for accurate absolute torque estimation altogether. Honda Research Institute’s admittance shaping controller (2015) defines assistance as a desired dynamic response (admittance) for the human leg — without any estimation of muscle torques or motion intent. By replacing the need for absolute torque values with a closed-loop sensitivity transfer function scaled by positive feedback, the approach avoids the STA error pathway entirely: there is no inverse dynamics stage where artifact can corrupt the signal.

Shenyang University of Technology’s admittance-controlled four-link bionic knee exoskeleton (2020) converts human-exoskeleton interaction forces into desired trajectories through admittance principles and compensates for inertial effects, achieving a joint angle tracking error below 5% and interaction forces within ±0.5 N. According to rehabilitation robotics guidelines published by WHO, achieving sub-5% tracking error is a clinically meaningful threshold for assistive gait devices. The admittance framework reaches this benchmark precisely because it does not attempt to reconstruct the absolute torque value that STA would corrupt.

National Chiao Tung University’s admittance control approach (2020) detects wearer intention by estimating total torques applied from the wearer to the human-exoskeleton system — specifically avoiding biological signal sensors like EMG and EEG — and uses the interaction-based torque signal estimated from exoskeleton joint sensors and a dynamic model as the intention proxy. The paper explicitly acknowledges that accurately separating the wearer’s dynamic contribution from the exoskeleton’s own dynamics is the core challenge that STA introduces, and the admittance framework addresses it by operating on the combined system signal rather than attempting to isolate the biological component.

Iterative learning control and kinematic design robustness: exploiting gait periodicity and fixing the root cause

Iterative learning control (ILC) exploits the quasi-periodic nature of gait to learn and cancel systematic errors — including those caused by consistent STA patterns — stride-by-stride. Stanford University’s ILC framework (2021) demonstrated that feed-forward iterative learning, analogous to stride-wise integral control, effectively compensates for the complicated dynamics of ankle exoskeletons during human walking, including the nonlinear coupling introduced by soft tissue at the cuff interface. The optimal gain was derived theoretically as the inverse of passive actuator stiffness, and walking experiments validated this prediction. ILC is particularly powerful for STA because STA-induced errors tend to be repeatable within a gait pattern, making them learnable cycle-to-cycle.

Shenzhen Institutes of Advanced Technology (2020) extended ILC to multi-joint assistance with a parameter-optimal ILC (POILC) strategy specifically designed to reduce errors caused by differences between the wearing position and the biological features of different wearers — a direct description of STA-induced inter-individual variability in the sensor-to-joint relationship. The system adapts the assistance profile based on biological torque data from different terrains, effectively personalising the STA correction across subjects and walking environments.

Kinematic design robustness: addressing the root cause mechanically

A root-cause approach to STA reduction is to minimise the relative motion between exoskeleton attachment points and the underlying skeletal anatomy before any signal processing occurs. Istituto Italiano di Tecnologia (2020) introduced a model-based method that represents kinematic misalignments as disturbance forces at anchor points, using robotic kinematic tools to systematically evaluate and minimise these disturbances across the inter- and intra-subject variability space. By designing exoskeleton kinematics to accommodate biological joint axis migration — a key driver of relative motion and STA — this method reduces the magnitude of the artifact before it enters the signal chain.

Ajou University (2019) addressed the migration of the knee center of rotation during motion — a fundamental source of STA and axis misalignment error — by tracking the changing distance between the knee center and ankle (LBKA) in real-time and using this biomechanically varying quantity as an intention detection signal. The self-aligning lower frame design reduces mechanical constraint forces that otherwise cause STA by forcing the limb into geometrically incorrect trajectories. KTH Royal Institute of Technology (2021) reinforced this direction using musculoskeletal modelling to quantify how different mechanical interface designs affect the distribution of forces at attachment points during simulated gait, providing a simulation-based tool for evaluating STA risk before physical prototyping.

Université Paris-Saclay (2021), working on upper-limb exoskeletons, demonstrated directly transferable methodology: modelling joint misalignments between the user and the exoskeleton and incorporating them into a feedforward weight compensation model across 17–29 participant cohorts. The study showed that ignoring such misalignments corrupts effective torque delivery, establishing misalignment monitoring as a generalizable framework applicable to lower-limb STA contexts. Biomechanical research published by Nature has similarly highlighted joint axis migration as a critical variable in wearable robot design, reinforcing the importance of kinematic alignment as a first-line STA defence.

Innovation landscape: key institutions and the convergence of physics and data

Analysis of more than 50 sources reveals distinct institutional clusters driving innovation in torque estimation accuracy and STA mitigation, with a clear trend toward hybrid architectures that combine model-based and data-driven approaches. No single method dominates — the highest-performing systems in the dataset combine at least two strategies.

The dominant trend identified across the dataset is the convergence of physical model-based and data-driven approaches. Rather than relying solely on inverse dynamics or solely on learned mappings, the highest-performing systems combine physics-informed state observers or dynamic compensation with neural network corrections. This hybrid architecture provides the structural guarantees of model-based control — stability, interpretability, known failure modes — while adapting to the irreducible uncertainty introduced by soft tissue mechanics that no deterministic model can fully capture.

Innovation intelligence standards from WIPO on emerging technology tracking highlight hybrid physics-data architectures as a fast-growing patent category in medical robotics, consistent with the convergence pattern observed in this dataset. The PatSnap Life Sciences intelligence platform provides real-time tracking of exoskeleton patent filings across all major jurisdictions, enabling R&D teams to monitor this convergence as it unfolds.

“The highest-performing exoskeleton torque estimation systems in the dataset combine physics-informed state observers with neural network corrections — providing structural guarantees of model-based control while adapting to the irreducible uncertainty of soft tissue mechanics.”

A key practical implication is that no single strategy is sufficient for all deployment contexts. EMG-based neuromechanical models are optimal when electrode placement is feasible and stable, but ESO-based sensorless methods are preferable for clinical rehabilitation settings where patient compliance with surface electrodes is limited. Admittance frameworks are architecturally robust to STA but require careful tuning of the admittance model to the individual user. ILC is powerful for community ambulators with consistent gait patterns but degrades for highly variable or pathological gait. The PatSnap R&D Intelligence platform enables systematic comparison of these approaches across the global patent and literature landscape to inform technology selection decisions.