Shank Angle-Based Exoskeleton Control

• Shank angle-based control systems use the continuous shank-to-vertical angle to generate adaptive exoskeleton assistance, ensuring synchronization with biological ankle torque (r up to 0.92).
• They employ minimal sensing setups, combining dual IMUs and load cells to extract key gait events (foot contact, foot-off, maximum dorsiflexion) with convergence in 9–11 gait cycles.
• This method outperforms traditional time-based controllers by maintaining high phase alignment even under perturbations and reducing metabolic cost by up to 11.86%.

Searching arXiv for the cited papers to ground the article with current metadata. A shank angle-based control system is an exoskeleton control architecture that uses human shank angle—defined as the shank-to-vertical angle—as the independent variable for assistance generation, rather than elapsed time or percent gait cycle. In the formulation reported for soft ankle exoskeleton assistance during non-steady locomotion, the controller is designed to maintain real-time coordination with gait even under nonlinear phase progression and phase perturbations, while shaping assistance profiles to match biological ankle moment patterns across level walking, level running, ramp ascent, and ramp descent (Tan et al., 13 Aug 2025). Related work on a shank-mounted inertial front-end demonstrates that a single IMU on the lateral shank can also supply low-latency, low-power, event-driven locomotion context recognition—stance, level walking, or stair ascent—to an exoskeleton controller, although that system does not itself compute or use a continuous shank angle trajectory for control (Razmi et al., 24 Feb 2026).

1. Conceptual basis and problem setting

The reported motivation for shank angle-based control is that many exoskeleton controllers are built around steady, rhythmic gait assumptions. Such controllers commonly rely on time within a stride, percent gait cycle, finite-state machine timing, or feedforward and iterative strategies that depend on previous periodic cycles. The stated limitation is that these assumptions break down in non-steady locomotion, including acceleration, deceleration, recovery from disturbances, and terrain negotiation, because gait progression is not linear in time within a stride (Tan et al., 13 Aug 2025).

Within that framework, the key problem is human-exoskeleton mismatch. A time-based controller advances assistance at a constant rate even if gait progression slows, speeds up, stalls, or reverses locally. The reported consequence is degraded synchronization, with assistance arriving too early or too late and force being applied in the wrong gait sub-phase, especially during abrupt perturbations. The paper characterizes these within-cycle distortions as phase perturbations and treats them as a central obstacle for conventional timing-based assistance (Tan et al., 13 Aug 2025).

The shank angle-based alternative is intended to avoid explicit activity classification and mode switching among separate activity-specific controllers. Instead, it uses a single control architecture whose assistance timing is state-coupled within the current stride and whose profile shape is updated stride by stride. This suggests a unified controller for diverse locomotion, rather than a cascaded recognition-then-switching pipeline.

2. Definition of shank angle and sensing configuration

In the reported convention, shank angle is zero in upright standing and positive when the knee is flexed. During stance, shank angle generally transitions from negative to positive as gait progresses from foot contact to foot-off, so it functions as a continuous descriptor of stance progression. The controller uses the current shank angle $\theta$ to determine the current point on the desired assistance profile (Tan et al., 13 Aug 2025).

The sensing configuration deliberately uses a minimal sensing set. The system uses 2 IMUs per leg—one on the lateral shank and one on the shoe instep—and 1 load cell per leg mounted near the heel counter or cable endpoint. The IMUs are sampled at 100 Hz, the load cell at 1 kHz, and the motor command loop runs at 1 kHz (Tan et al., 13 Aug 2025).

The measured signals are shank angle from the shank IMU, foot pitch angle from the foot IMU, shank angular velocity, foot angular velocity, cable force from the load cell, and cable length and velocity from motor or pulley measurement. The computed signals include ankle dorsiflexion angle,

$\theta_{DF} = \theta_{SK} - \theta_{FT},$

ankle dorsiflexion angular velocity, foot contact, foot-off, Gaussian profile parameters $(\sigma_1,\sigma_2,\mu)$ for each stride, and artificial tendon length and feedforward terms through the coupled human-exoskeleton model (Tan et al., 13 Aug 2025).

A distinct but related architecture is reported in the shank-mounted IMU front-end paper. There, a single IMU is mounted on the lateral aspect of the participant’s shank with elastic straps, and tri-axial accelerometer and tri-axial gyroscope data are used by the embedded Machine Learning Core of the LSM6DSV16X IMU. That system outputs a discrete locomotion label rather than a continuous angle estimate (Razmi et al., 24 Feb 2026).

3. Assistance profile generation as a function of shank angle

The desired assistive force profile is formulated as a piecewise dual-Gaussian function of shank angle: $$F(\theta) = \begin{cases} A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_1^2}\right), & \theta_{FC} \le \theta \le \mu \ A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_2^2}\right), & \mu < \theta \le \theta_{FO} \end{cases} \tag{1}$$ where $F(\theta)$ is desired assistive force at shank angle $\theta$, $A$ is amplitude, $\mu$ is the mean or center of the Gaussian peak, $\sigma_1$ is the standard deviation for the rising part, $\sigma_2$ is the standard deviation for the falling part, and $\theta_{DF} = \theta_{SK} - \theta_{FT},$0 and $\theta_{DF} = \theta_{SK} - \theta_{FT},$1 are shank angle at foot contact and foot-off, respectively (Tan et al., 13 Aug 2025).

The stated motivation is that the biological ankle plantarflexion moment is single-peaked, so a dual-Gaussian can represent asymmetric rise and fall behavior with few interpretable parameters. In this parameterization, $\theta_{DF} = \theta_{SK} - \theta_{FT},$2 sets assistance magnitude, $\theta_{DF} = \theta_{SK} - \theta_{FT},$3 aligns the assistance peak with the point of maximal ankle demand, and $\theta_{DF} = \theta_{SK} - \theta_{FT},$4 and $\theta_{DF} = \theta_{SK} - \theta_{FT},$5 control how quickly force rises after foot contact and falls toward foot-off (Tan et al., 13 Aug 2025).

The central estimation step is the placement of the profile peak. The reported biomechanical finding is that across walking and running and across level and slope terrains, peak ankle plantarflexion torque consistently coincides with peak ankle dorsiflexion angle. Accordingly, the controller estimates

$\theta_{DF} = \theta_{SK} - \theta_{FT},$6

where $\theta_{DF} = \theta_{SK} - \theta_{FT},$7 is the shank angle corresponding to the maximum ankle dorsiflexion angle within the stride (Tan et al., 13 Aug 2025).

Foot contact and foot-off are detected from foot IMU features that were reported to remain valid across level walking, level running, ramp ascent, and ramp descent: foot contact is detected near maximum foot pitch angle, and foot-off near minimum foot pitch angular velocity. The corresponding shank angles are recorded as $\theta_{DF} = \theta_{SK} - \theta_{FT},$8 and $\theta_{DF} = \theta_{SK} - \theta_{FT},$9 (Tan et al., 13 Aug 2025).

The width parameters are then updated as

$(\sigma_1,\sigma_2,\mu)$0

and

$(\sigma_1,\sigma_2,\mu)$1

The denominator 4 is explicitly chosen to smooth the profile and avoid abrupt force changes at the ends. The amplitude $(\sigma_1,\sigma_2,\mu)$2 is fixed experimentally to a fraction of body weight, with 15% BW in some experiments and 20% BW in others. The reported study does not adapt magnitude online; only shape and phase alignment are adapted (Tan et al., 13 Aug 2025).

The online update is stride-by-stride. During a stride, the controller collects windows of shank angle and ankle dorsiflexion angle. At foot-off, it identifies the peak dorsiflexion angle index, computes $(\sigma_1,\sigma_2,\mu)$3, extracts $(\sigma_1,\sigma_2,\mu)$4 and $(\sigma_1,\sigma_2,\mu)$5, and updates $(\sigma_1,\sigma_2,\mu)$6, $(\sigma_1,\sigma_2,\mu)$7, and $(\sigma_1,\sigma_2,\mu)$8. The updated profile is then used in the next gait cycle. The supplementary pseudocode is reported to filter updates with a gain of 0.3 if increments are normal, and convergence to stable values is reported after 9–11 gait cycles. Initial parameter values are $(\sigma_1,\sigma_2,\mu)$9, $$F(\theta) = \begin{cases} A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_1^2}\right), & \theta_{FC} \le \theta \le \mu \ A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_2^2}\right), & \mu < \theta \le \theta_{FO} \end{cases} \tag{1}$$0, and $$F(\theta) = \begin{cases} A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_1^2}\right), & \theta_{FC} \le \theta \le \mu \ A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_2^2}\right), & \mu < \theta \le \theta_{FO} \end{cases} \tag{1}$$1 (Tan et al., 13 Aug 2025).

4. Model-based feedforward tracking and coupled mechanics

The assistance profile generator specifies desired force, but the soft exoskeleton must also track that force in the presence of compliance and low mechanical bandwidth. The reported solution is a model-based feedforward control method built on a coupled human-exoskeleton kinematics and stiffness model (Tan et al., 13 Aug 2025).

The exoskeleton cable is modeled as an artificial tendon running approximately parallel to the Achilles tendon. Its effective length depends on ankle kinematics, suit deformation or compliance, and migration of the soft interface on the leg. The tendon is split into two segments, $$F(\theta) = \begin{cases} A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_1^2}\right), & \theta_{FC} \le \theta \le \mu \ A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_2^2}\right), & \mu < \theta \le \theta_{FO} \end{cases} \tag{1}$$2 and $$F(\theta) = \begin{cases} A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_1^2}\right), & \theta_{FC} \le \theta \le \mu \ A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_2^2}\right), & \mu < \theta \le \theta_{FO} \end{cases} \tag{1}$$3, with stiffnesses $$F(\theta) = \begin{cases} A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_1^2}\right), & \theta_{FC} \le \theta \le \mu \ A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_2^2}\right), & \mu < \theta \le \theta_{FO} \end{cases} \tag{1}$$4 and $$F(\theta) = \begin{cases} A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_1^2}\right), & \theta_{FC} \le \theta \le \mu \ A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_2^2}\right), & \mu < \theta \le \theta_{FO} \end{cases} \tag{1}$$5, and assistance lever arm $$F(\theta) = \begin{cases} A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_1^2}\right), & \theta_{FC} \le \theta \le \mu \ A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_2^2}\right), & \mu < \theta \le \theta_{FO} \end{cases} \tag{1}$$6. The segment length equations are given as

$$F(\theta) = \begin{cases} A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_1^2}\right), & \theta_{FC} \le \theta \le \mu \ A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_2^2}\right), & \mu < \theta \le \theta_{FO} \end{cases} \tag{1}$$7

and

$$F(\theta) = \begin{cases} A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_1^2}\right), & \theta_{FC} \le \theta \le \mu \ A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_2^2}\right), & \mu < \theta \le \theta_{FO} \end{cases} \tag{1}$$8

The combined tendon length is then expressed as

$$F(\theta) = \begin{cases} A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_1^2}\right), & \theta_{FC} \le \theta \le \mu \ A \cdot \exp\left(-\dfrac{(\theta-\mu)^2}{2\sigma_2^2}\right), & \mu < \theta \le \theta_{FO} \end{cases} \tag{1}$$9

with

$F(\theta)$0

and

$F(\theta)$1

In the paper’s presentation, Eq. (6) is written as

$F(\theta)$2

The lever arm $F(\theta)$3 is manually measured for each subject (Tan et al., 13 Aug 2025).

The coupled stiffness is identified experimentally by loading and unloading cable force from 5 N to 180 N over 10 cycles while the user holds a push-off-like pose. A linear fit gives

$F(\theta)$4

with

$F(\theta)$5

Migration compensation is estimated as

$F(\theta)$6

This is estimated once during initial cable tightening in each gait cycle and is assumed approximately constant within that cycle. Migration is reported to stabilize after about 7–10 assisted gait cycles (Tan et al., 13 Aug 2025).

During swing, the desired slack-state tendon length is updated stride by stride as

$F(\theta)$7

where the target swing force is 3 N. The corresponding velocity command during swing is

$F(\theta)$8

with $F(\theta)$9, $\theta$0, and $\theta$1 (Tan et al., 13 Aug 2025).

Because the motor driver is in velocity mode, force error is mapped to cable retraction speed through

$\theta$2

with experimental values $\theta$3 and $\theta$4. The stance command is then

$\theta$5

where the feedforward term is

$\theta$6

and

$\theta$7

Here $\theta$8 is $\theta$9 or $A$0 depending on whether $A$1 or $A$2, and $A$3 is shank angular velocity from the IMU (Tan et al., 13 Aug 2025).

The reported assumptions are that the cable-suit-human system can be approximated by linear equivalent stiffness $A$4, the lever arm $A$5 is approximately constant for each subject, suit migration is slowly varying and constant within a stride after estimation, assistance is primarily sagittal ankle plantarflexion assistance, and IMU-derived angles are sufficiently accurate (Tan et al., 13 Aug 2025).

5. Real-time coordination, perturbation robustness, and comparison with time-based control

The central claimed advantage of shank angle as the control variable is that it is state-coupled. If the user delays push-off, accelerates, decelerates, or undergoes a perturbation, the progression of the force point along the profile changes accordingly. The paper explicitly notes that the force point can move nonlinearly or even in reverse to stay coordinated with gait progression. This is contrasted with time, which is described as irreversible and independent of human state (Tan et al., 13 Aug 2025).

The architecture handles irregularity through two mechanisms: instantaneous shank angle drives profile progression within the current stride, and the next-cycle profile is regenerated from the most recent measured foot contact, foot-off, and maximum dorsiflexion features. The same profile structure and event logic are reported across level walking, level running, ramp ascent, and ramp descent without explicit activity classification (Tan et al., 13 Aug 2025).

A direct comparison is reported under treadmill perturbations. Pearson correlation with biological ankle torque over stance is given as follows.

Condition	Reported correlation
Shank-angle-based profiles	average $A$6, almost all $A$7
Time-based profiles, forward perturbations	average $A$8
Time-based profiles, backward perturbations	average $A$9

Reported examples are level walking forward $\mu$0, $\mu$1; level walking backward $\mu$2, $\mu$3; level running forward $\mu$4, $\mu$5; level running backward $\mu$6, $\mu$7; ramp ascent forward $\mu$8, $\mu$9; ramp ascent backward $\sigma_1$0, $\sigma_1$1; and ramp descent with only shank-angle values reported, $\sigma_1$2 forward and $\sigma_1$3 backward (Tan et al., 13 Aug 2025).

This supports the reported interpretation that time-based assistance degrades especially under backward perturbations, whereas shank-angle-based profiles remain strongly aligned with biological ankle torque. A plausible implication is that the state-coupled parameterization is particularly beneficial when within-stride timing is distorted.

6. Experimental platform and reported outcomes

The reported soft ankle exoskeleton comprises a lightweight leather wearable suit, Bowden cable actuation, off-board motors and control box behind the treadmill, and cable pulling parallel to the Achilles tendon. Hardware details include a suit mass of 0.067 kg per leg, total wearable exoskeleton plus sensors of less than 0.5 kg on body, Maxon EC-4pole motors with 66:1 reduction ratio, Elmo Gold Solo Twitter motor drivers, a 24 V, 11.6 Ah battery, RT Ubuntu 14.04, EtherCAT, motor velocity command rate of 1 kHz, Xsens Awinda IMUs at 100 Hz, and a Futek LSB205 load cell at 1 kHz (Tan et al., 13 Aug 2025).

The participant cohort includes Session 1 with 1 subject, Session 2 with 1 subject, and Session 3 with 8 male subjects, age $\sigma_1$4 years, mass $\sigma_1$5 kg, and height $\sigma_1$6 m; 5 of the 8 were novices to exoskeleton use (Tan et al., 13 Aug 2025).

Three experiments are reported. Experiment 1 validated sensing, online profile generation, and profile tracking over 2-minute bouts of level walking at 1.33 m/s, level running at 2.20 m/s, ramp ascent at 1.33 m/s and $\sigma_1$7, and ramp descent at 1.33 m/s and $\sigma_1$8. Experiment 2 assessed perturbation robustness with treadmill belt perturbations during stance, applied at 15% gait cycle onset, belt speed changed by 80% within 0.1 s and returned in another 0.1 s, acceleration set to $\sigma_1$9, and one perturbation maximum per cycle. Experiment 3 evaluated physiological and biomechanical effects under no exoskeleton, passive, and active assistance conditions over 2-minute bouts of level walking at 1.0 m/s, level running at 2.0 m/s, ramp ascent at 1.0 m/s, and ramp descent at 1.0 m/s, with assistance magnitude 20% BW (Tan et al., 13 Aug 2025).

For online profile generation, Session 1 is reported to show that the IMU-based variables were sufficient to estimate $\sigma_2$0, $\sigma_2$1, and $\sigma_2$2, that profile parameters converged after 9–11 gait cycles, and that generated profiles adapted across walking, running, ascent, and descent. Supplementary speed-change results reportedly showed that $\sigma_2$3, $\sigma_2$4, and $\sigma_2$5 changed with locomotion speed and maintained profile alignment with biological ankle torque patterns (Tan et al., 13 Aug 2025).

For model-based feedforward tracking, tracking RMSE relative to max desired force at 15% BW is reported as $\sigma_2$6 for level walking, $\sigma_2$7 for level running, $\sigma_2$8 for ramp ascent, and $\sigma_2$9 for ramp descent. Tracking error is reported to be worst in running because motor maximum speed limited engagement in early stance (Tan et al., 13 Aug 2025).

Whole-system biomechanical results include delivered peak exoskeleton mechanical plantarflexion moment of about 0.18–0.19 Nm/kg across tasks, and delivered peak exoskeleton plantarflexion work rate of $\theta_{DF} = \theta_{SK} - \theta_{FT},$00 W/kg for level walking, $\theta_{DF} = \theta_{SK} - \theta_{FT},$01 W/kg for level running, $\theta_{DF} = \theta_{SK} - \theta_{FT},$02 W/kg for ramp ascent, and $\theta_{DF} = \theta_{SK} - \theta_{FT},$03 W/kg for ramp descent. With active assistance, gait cycle duration is reported to reduce by about 4.12% versus no exoskeleton and 2.45% versus passive on average; peak total ankle plantarflexion moment decreased across activities versus passive, and in level running versus no exoskeleton by $\theta_{DF} = \theta_{SK} - \theta_{FT},$04 or $\theta_{DF} = \theta_{SK} - \theta_{FT},$05, $\theta_{DF} = \theta_{SK} - \theta_{FT},$06; peak total plantarflexion work rate decreased relative to no exoskeleton and passive across all activities, by an average of $\theta_{DF} = \theta_{SK} - \theta_{FT},$07 or $\theta_{DF} = \theta_{SK} - \theta_{FT},$08 versus no exoskeleton and $\theta_{DF} = \theta_{SK} - \theta_{FT},$09 or $\theta_{DF} = \theta_{SK} - \theta_{FT},$10 versus passive (Tan et al., 13 Aug 2025).

Average change in normalized EMG RMS relative to no exoskeleton is reported as ACT $\theta_{DF} = \theta_{SK} - \theta_{FT},$11, PAS $\theta_{DF} = \theta_{SK} - \theta_{FT},$12 for level walking; ACT $\theta_{DF} = \theta_{SK} - \theta_{FT},$13, PAS $\theta_{DF} = \theta_{SK} - \theta_{FT},$14 for level running; ACT $\theta_{DF} = \theta_{SK} - \theta_{FT},$15, PAS $\theta_{DF} = \theta_{SK} - \theta_{FT},$16 for ramp ascent; and ACT $\theta_{DF} = \theta_{SK} - \theta_{FT},$17, PAS $\theta_{DF} = \theta_{SK} - \theta_{FT},$18 for ramp descent. The largest and most consistent benefit is reported for lateral gastrocnemius reductions: $\theta_{DF} = \theta_{SK} - \theta_{FT},$19 in level walking, $\theta_{DF} = \theta_{SK} - \theta_{FT},$20 in level running, $\theta_{DF} = \theta_{SK} - \theta_{FT},$21 in ramp ascent, and $\theta_{DF} = \theta_{SK} - \theta_{FT},$22 in ramp descent (Tan et al., 13 Aug 2025).

Active assistance reduced net metabolic rate versus no exoskeleton by $\theta_{DF} = \theta_{SK} - \theta_{FT},$23, $\theta_{DF} = \theta_{SK} - \theta_{FT},$24 in level walking; $\theta_{DF} = \theta_{SK} - \theta_{FT},$25, $\theta_{DF} = \theta_{SK} - \theta_{FT},$26 in level running; $\theta_{DF} = \theta_{SK} - \theta_{FT},$27, $\theta_{DF} = \theta_{SK} - \theta_{FT},$28 in ramp ascent; and $\theta_{DF} = \theta_{SK} - \theta_{FT},$29, $\theta_{DF} = \theta_{SK} - \theta_{FT},$30 in ramp descent. Passive exoskeleton use increased metabolic rate, and veteran users are reported to perform better than novices, with veterans showing PAS $\theta_{DF} = \theta_{SK} - \theta_{FT},$31, ACT $\theta_{DF} = \theta_{SK} - \theta_{FT},$32, and novices PAS $\theta_{DF} = \theta_{SK} - \theta_{FT},$33, ACT $\theta_{DF} = \theta_{SK} - \theta_{FT},$34 (Tan et al., 13 Aug 2025).

7. Relation to shank-mounted locomotion recognition and scope limitations

The shank-mounted IMU locomotion-recognition system reported in 2026 is not a direct shank-angle control paper. It does not present a controller that estimates continuous shank angle and maps that angle to exoskeleton torque. Instead, it presents a shank-mounted, on-sensor locomotion mode recognition front-end: a single IMU on the lateral shank runs a tiny decision-tree classifier inside the IMU itself and outputs a discrete locomotion label—stance, level walking, or stair ascent—to a host microcontroller, which can then pass that mode to an exoskeleton controller (Razmi et al., 24 Feb 2026).

Its architecture is event-driven. The IMU performs feature extraction and decision-tree inference internally, raises an interrupt when a WAKEUP or MLC event occurs, and the host microcontroller wakes from low power, reads the MLC output or status registers, retrieves the recognized locomotion mode, and forwards that mode to the exoskeleton controller. The decision-tree output register is dec_tree_out_1, with reported real-time encoding 8 = stance, 0 = walk, and 4 = stairsUp. The host does not require custom machine learning code, because the classifier is configured in ST MEMS Studio and exported as an MLC-compatible configuration (Razmi et al., 24 Feb 2026).

The classifier uses raw tri-axial accelerometer $\theta_{DF} = \theta_{SK} - \theta_{FT},$35 and tri-axial gyroscope $\theta_{DF} = \theta_{SK} - \theta_{FT},$36 measurements from a shank-mounted IMU and a compact set of 15 features selected with Activity Feature Selection in expert mode using ANOVA-based ranking, AdaBoost evaluations, Random Forest evaluations, and Recursive Feature Elimination. The MLC operating data rate is set to 240 Hz with a window length of 240 samples, corresponding to a 1-second window, while raw IMU acquisition is sampled at the highest possible sampling rate of 7.68 kHz (Razmi et al., 24 Feb 2026).

The paper explicitly discusses shank angular motion patterns, peak-to-peak acceleration, gyroscope mean feature, tri-axial acceleration, and tri-axial angular velocity, and states that “the peak-to-peak acceleration feature captures the transition from quiet stance to dynamic locomotion, while the gyroscope mean feature reflects differences in shank angular motion patterns between level walking and stair ascent” (Razmi et al., 24 Feb 2026). However, it explicitly does not compute continuous shank angle, shank tilt angle, inclination, orientation from sensor fusion, Euler angles, quaternions, gait phase percentage, explicit heel-strike or toe-off timing, knee angle, or joint biomechanics.

In relation to shank angle-based control, the significance of that paper is therefore supervisory rather than substitutive. It shows how a shank-mounted IMU can provide low-latency, low-power, event-driven locomotion context recognition that could complement a shank-angle-based controller or gait-state machine. A plausible implication is a hybrid architecture in which continuous shank-angle control provides phase-dependent torque while the on-sensor classifier selects among stance, level-walking, and stair-ascent assistance contexts. The reported evidence does not support interpreting the on-sensor recognition front-end itself as a shank angle-based control system (Razmi et al., 24 Feb 2026).

The shank angle-based controller also has reported limitations. Shank angle alone does not adapt assistance magnitude, since $\theta_{DF} = \theta_{SK} - \theta_{FT},$37 remains fixed as a percentage of body weight; running tracking suffered from actuator speed saturation; the system was tested with off-board actuation; performance depends on robust gait event detection and IMU quality; user adaptation and familiarity matter; and very unusual motions such as backward walking or highly non-sagittal maneuvers were discussed conceptually but not experimentally validated (Tan et al., 13 Aug 2025). These constraints delimit the present scope of the approach while clarifying why shank angle is treated as a control variable for phase and profile alignment rather than a complete solution to all aspects of exoskeleton adaptation.