Personalized Gait Control Motor

• Personalized gait control motor is a systems-level approach that adapts sensing, estimation, and control to individual gait characteristics and impairment profiles.
• It leverages techniques such as phase-dependent impedance tuning, real-time motion capture, and advanced control methods like MPC and reinforcement learning for dynamic assistance.
• This tailored approach improves assistive performance by reducing tracking errors and enabling subject-specific adjustments in prostheses, exoskeletons, and hybrid gait systems.

Personalized gait control motor, in current research usage, can be understood as a class of gait-assistive systems in which sensing, actuation, and control are individualized to the gait characteristics, impairment profile, and task context of a specific user. Recent work places this concept across multichannel functional electrical stimulation (FES), powered prostheses, exoskeletons, and pelvis-coupled assistive robots, with personalization expressed through event timing, impedance trajectories, torque timing, reference gait patterns, admittance parameters, or subject-specific simulation models rather than through a single standardized controller architecture (Graffagnino et al., 30 Sep 2025, Reznick et al., 2024, Fortuna et al., 2024, Song et al., 16 Jun 2026). The field is therefore defined less by one motor topology than by a recurrent systems principle: user-specific gait information is converted into individualized control actions or individualized controller parameters.

1. Conceptual scope

Personalized gait control differs from generic assistive locomotion control by treating the target of adaptation as user-specific. In powered prostheses, this may mean individualized phase-dependent impedance parameters over stance flexion, stance extension, swing flexion, and swing extension; in exoskeletons, it may mean individualized gait phase estimation, personalized reference trajectories, or subject-specific torque timing; in hybrid robot-FES systems, it may mean assistance sharing that depends on tracking performance and estimated muscle fitness; and in pelvis-coupled walkers, it may mean tuning virtual mass and damping to user preference rather than imposing fixed interaction dynamics (Li et al., 2020, Song et al., 16 Jun 2026, Christou et al., 2024, Fortuna et al., 2024).

The literature also shows that personalization is not confined to direct hardware tuning. Some studies treat it as a laboratory prototyping problem, such as real-time motion-capture-triggered multichannel stimulation for future cerebral palsy rehabilitation, while others treat it as an optimization problem over controller parameters, a continual online adaptation problem, or a simulation-first policy-learning problem using musculoskeletal digital humans (Graffagnino et al., 30 Sep 2025, Christou et al., 1 Mar 2025, Choi et al., 10 Apr 2026). This suggests that “personalized gait control motor” is best regarded as a systems-level category spanning sensing, state estimation, controller synthesis, and actuation allocation.

2. Sensing, state estimation, and system architecture

A defining feature of personalized gait control is the reliance on user-specific measurements. One laboratory architecture uses 22 reflective markers tracked at 100 Hz in the GRAIL system with a 10-camera Vicon setup, real-time D-Flow processing with the Human Body Model, trigger transmission to an 8-channel Motimove stimulator, and EMG artifact verification at 1000 Hz with a Cometa system; the closed loop detects heel strike, knee flexion, and ankle dorsiflexion online and measures end-to-end latency from event detection to delivered stimulation (Graffagnino et al., 30 Sep 2025). A wearable ankle-assistance pipeline instead uses three IMUs on the paretic limb plus an ankle encoder, reduces the sensing space to 16 channels, and performs 100 Hz embedded inference of paretic ankle torque using a temporal convolutional network on a Raspberry Pi 5 (Weigend et al., 1 Aug 2025).

High-level state estimation has become a major personalization locus. In one exoskeleton architecture, bilateral hip encoders streamed at 100 Hz feed a temporal convolutional network that predicts unilateral gait phase in Cartesian form, $(\cos\phi,\sin\phi)$, and only the final linear layer is updated online; the phase estimate then drives predefined task-dependent torque splines (Song et al., 16 Jun 2026). In another line, personalized gait recognition is treated as a perception problem rather than a control law: ExoGait-MS combines a gait nonlinear periodic dynamics learning module with a multi-scale global dense graph convolutional network and reports 94.34% accuracy, 0.9428 F1-score, and 0.9990 AUC on its constructed gait dataset, explicitly positioning individualized recognition as a prerequisite for tailored exoskeleton gait control (Liu et al., 23 May 2025).

At the human-robot interface, pelvis-level sensing is prominent. WANDER couples the user at the pelvis/lumbar region through a rigid interface with a LaxOne 6-axis force/torque sensor and translates $\mathbf{T}^{ext}_m=[F_x,F_y,\tau_z]$ into platform motion through an admittance controller (Fortuna et al., 2024). Human-in-the-loop simulation for passive-support robots similarly emphasizes the physical human-robot interaction channel, inserting a six-DoF virtual free joint between pelvis and robot and modeling the interface as a mass-spring-damper system rather than as a rigid attachment (Wang et al., 5 Mar 2025). A related surrogate-testing line uses a 4-DOF bipedal robot with hip and knee actuators, IMUs, and SDRE-optimized torque control to reproduce measured human gait for knee-type exoskeleton evaluation under repeatable conditions (Huang et al., 5 Jun 2025).

3. Control formulations

The control formulations used in personalized gait control motors span discrete event triggering, continuous impedance control, admittance control, model predictive control, and reinforcement learning. The prosthesis literature retains a strong impedance lineage. A robotic knee prosthesis with finite-state machine impedance control divides gait into stance flexion, stance extension, swing flexion, and swing extension, and in each phase applies

$\tau=K(\theta_e-\theta)-C\omega$

with 12 impedance parameters to be tuned across the gait cycle (Li et al., 2020). A clinical tuning framework for a powered knee-ankle prosthesis generalizes this idea into continuous phase/task models: stance uses phase-varying impedance, while swing uses continuous kinematic references parameterized by gait phase, walking speed, and slope (Reznick et al., 2024).

Hybrid rehabilitation controllers add explicit allocation logic across multiple actuators. One robot-FES path controller defines a dead band, an FES band, and a hybrid band around a reference gait path in hip-knee joint space; FES acts outside the dead band, and robot torque is recruited only outside the FES band, so voluntary movement is prioritized first, FES second, and robotic assistance last (Christou et al., 2024). The FES command is further modulated by muscle fitness and by an iterative-learning gain, whereas exoskeleton stiffness is adapted from phase-specific RMS tracking error (Christou et al., 2024).

Admittance control is central in pelvis-coupled assistive robots. WANDER uses the second-order law

$$\mathbf{M}_{adm}\mathbf{\ddot{q}^{des}_m} + \mathbf{D}_{adm}\mathbf{\dot{q}^{des}_m} = \mathbf{T}^{ext}_m$$

with personalized virtual mass and damping, and adds a direction-based variable damping law that reduces damping along the intended direction of motion (Fortuna et al., 2024). In this setting, personalization occurs at the level of interaction dynamics rather than at the level of joint reference trajectories.

Predictive control appears most clearly in ankle FES assistance. A Koopman-based MPC framework models stance plantarflexion and swing dorsiflexion as a switched nonlinear system, lifts the dynamics into a Koopman observable space, derives phase-specific linear predictors, and solves constrained real-time MPC for stimulation current amplitude while keeping frequency fixed at 33 Hz (Singh et al., 10 Jan 2025). The controller assists both Tibialis Anterior during swing and Gastrocnemius during stance, thereby targeting both toe clearance and push-off rather than swing-only drop-foot correction (Singh et al., 10 Jan 2025).

Reinforcement learning appears in both online human-in-the-loop control and simulation-first control design. An adaptive ankle-foot orthosis learns a subject-specific phase-dependent impedance landscape $g(\phi)$ online with a modified PI$^2$ algorithm, using a dual-objective cost that penalizes both tracking error and robot assistance and a high-level hysteretic switch between intervention and compliance modes (Zhang et al., 2021). A musculoskeletal simulation framework instead learns closed-loop exoskeleton torque policies jointly with muscle excitations, yielding deficit-specific asymmetric assistance without explicit prescription of a target impaired gait pattern (Choi et al., 10 Apr 2026).

Controller family	Primary control variable	Personalized target
Event-triggered FES	Trigger timing from gait events	Event type, threshold, monitored leg (Graffagnino et al., 30 Sep 2025)
FSM or continuous impedance	$K$, $B$, $\theta_{eq}$ or phase/task models	User-specific stance and swing behavior (Li et al., 2020, Reznick et al., 2024)
Hybrid robot-FES path control	FES pulse width and exoskeleton stiffness	Performance- and fatigue-dependent assistance sharing (Christou et al., 2024)
Admittance control	Virtual mass and damping	User-preferred interaction dynamics (Fortuna et al., 2024)
Phase-estimation-driven exoskeleton control	Gait phase estimator weights	User-specific torque timing across tasks (Song et al., 16 Jun 2026)

4. Mechanisms of personalization

The literature operationalizes personalization through several distinct mechanisms. One class of methods tunes explicit controller parameters from user-specific gait errors. In robotic knee prostheses, Policy Iteration with Constraint Embedded treats each gait phase independently, defines the state as the error in local gait features relative to able-bodied targets, and learns a policy that maps those errors to updates in stiffness, equilibrium angle, and damping (Li et al., 2020). In robot-assisted gait training, model-based optimization builds a subject-specific OpenSim human-robot model, estimates voluntary torques from transparent-mode trials, and optimizes four phase-dependent hip and knee stiffness parameters for assistance as needed (Christou et al., 1 Mar 2025).

A second class personalizes the interaction channel rather than the joint trajectory. WANDER uses Preference-Based Optimization over a two-dimensional parameter space of translational virtual mass and damping, learning a latent subjective objective from pairwise user preferences rather than from an explicit analytic comfort metric (Fortuna et al., 2024). Closely related preference-landscape learning in exoskeleton gait tuning uses a GP-based Bayesian model over step length, step duration, pelvis roll, and pelvis pitch, and explicitly separates a Region of Avoidance from a Region of Interest so that highly undesirable gaits are no longer queried once confidence is sufficient (Li et al., 2020).

A third class personalizes high-level state estimation online. Continual online personalization of exoskeleton control via manifold-aware experience replay adapts only the final linear layer of a pretrained gait phase network, preserving performance across speed and incline transitions by replaying previously encountered locomotor contexts organized in a compact gait manifold rather than by explicit task labels (Song et al., 16 Jun 2026). A related data-driven ankle controller fine-tunes a multi-task TCN on post-stroke data so that real-time torque assistance depends continuously on the user’s current wearable-sensor kinematics rather than on a fixed event-triggered torque profile (Weigend et al., 1 Aug 2025).

A fourth class personalizes through simulation and digital twins. Human-in-the-loop simulation for passive-support robots creates a personalized full-body human digital twin with 27 DoF and a learned locomotion policy derived from subject-specific markerless motion capture (Wang et al., 5 Mar 2025). Musculoskeletal motion imitation trains a physiologically plausible 21-DoF, 90-muscle human model, then fine-tunes exoskeleton assistance under unilateral plantarflexor or hip-flexor weakness so that asymmetric assistance emerges from asymmetric impairment (Choi et al., 10 Apr 2026). A related data-physics hybrid generative framework reconstructs a stroke survivor’s locomotor controller from a single 20 m level-ground walking trial with five IMUs and predicts slope ascent and stair climbing for rehabilitation planning (Dai et al., 16 Dec 2025). This suggests that personalization increasingly includes not only controller tuning but also patient-specific predictive modeling.

5. Empirical performance and translational evidence

Published results show substantial technical progress, although the evidence is highly heterogeneous. In passive-support gait robotics, a speed-adaptive controller for the Mobile Robotic Balance Assistant reduced average forward tracking error from 4.25 cm to 0.37 cm and average lateral tracking error from 3.32 cm to 0.72 cm in real walking relative to a conventional PID controller, while preserving stride length and gait speed much better than PID (Wang et al., 5 Mar 2025). In powered prosthesis tuning, offline pretraining with PICE reduced online tuning burden: on average, starting from a pre-trained policy caused only 1 phase to require online policy updates, whereas random initialization required updates in 4 phases, with 58 fewer impedance updates on average, equivalent to about 7 minutes less walking time (Li et al., 2020).

In FES-based ankle assistance, Koopman-based MPC reported overall ankle trajectory tracking RMSE of $1.625^\circ$ when muscles were rested and maintained closed-loop assistance for both plantarflexion and dorsiflexion, while a participant with Multiple Sclerosis progressed from toe drag and inability to sustain treadmill walking at 0.1 m/s before control to sustained walking at 0.1, 0.2, and 0.3 m/s after control (Singh et al., 10 Jan 2025). In post-stroke ankle torque estimation for exoskeleton control, the best fine-tuned multi-task TCN achieved $\mathbf{T}^{ext}_m=[F_x,F_y,\tau_z]$0 Nm/kg, $\mathbf{T}^{ext}_m=[F_x,F_y,\tau_z]$1 Nm/kg, and $\mathbf{T}^{ext}_m=[F_x,F_y,\tau_z]$2 across leave-one-subject-out evaluation, and the real-time prototype achieved $\mathbf{T}^{ext}_m=[F_x,F_y,\tau_z]$3 Nm/kg, $\mathbf{T}^{ext}_m=[F_x,F_y,\tau_z]$4 Nm/kg, and $\mathbf{T}^{ext}_m=[F_x,F_y,\tau_z]$5 during wearable embedded control (Weigend et al., 1 Aug 2025).

Clinical usability has also improved. A powered knee-ankle prosthesis with a Clinical Tuning Interface was fully tuned by a prosthetist in under 20 min, with each walking iteration taking 2 min on average and each sit-stand iteration taking 1 min on average; the tuned behavior changes were manifested not only in manually tuned tasks but also in automatically tuned incline tasks (Reznick et al., 2024). Laboratory event-triggered multichannel stimulation within GRAIL confirmed end-to-end latency of less than 100 ms between gait event detection and the first stimulation artifact in EMG, establishing feasibility for future closed-loop gait correction applications (Graffagnino et al., 30 Sep 2025). These results support technical viability, but not yet broad clinical efficacy.

6. Limitations, misconceptions, and research directions

A recurrent misconception is that trajectory personalization alone is sufficient. A pilot study with ten unimpaired participants found no relevant differences in comfort, naturalness, or overall experience between personalized, average, and random gait patterns when all trajectories were executed with high accuracy by a stiff position-derivative controller; later trials were rated as more comfortable and natural than the first trial, suggesting that user adaptation to the exoskeleton can dominate the effect of kinematic personalization (Luciani et al., 19 Dec 2025). This implies that personalization of interaction dynamics may matter at least as much as personalization of reference motion.

A second misconception is that architectural personalization automatically implies clinically individualized control. The proof-of-concept motion-capture-triggered multichannel stimulation system explicitly supports personalization at the architectural level but does not yet provide subject-specific channel maps, numerical thresholds, participant-specific tuning rules, or muscle-specific stimulation amplitudes (Graffagnino et al., 30 Sep 2025). Similarly, model-based offline optimization predicted an average objective improvement of about 30.4% in simulation, but experimental responses varied: six subjects improved significantly, eight showed no obvious change, and four performed worse, highlighting the limits of static human models and simplified contact models (Christou et al., 1 Mar 2025).

A third misconception is that online adaptation is intrinsically stable across tasks. Continual online personalization of exoskeleton control shows that gait-phase estimators can catastrophically forget previously learned locomotor contexts during speed and incline transitions unless replay mechanisms are introduced; manifold-aware replay improved torque tracking accuracy by 40% and gait phase tracking accuracy by 60% relative to a baseline without replay in forgetting scenarios (Song et al., 16 Jun 2026). This indicates that long-horizon deployment requires continual-learning safeguards, not just local adaptation.

The current research frontier therefore lies in combining individualized sensing, structured control, and robust adaptation. Likely near-term directions include explicit handling of fatigue and spasticity in FES, impaired-user datasets for digital-twin personalization, wearable substitutes for laboratory phase labeling, richer multi-joint coordination beyond knee or ankle only, and integration of subjective comfort and naturalness into the objective function rather than treating them as secondary outcomes (Christou et al., 2024, Singh et al., 10 Jan 2025, Luciani et al., 19 Dec 2025, Dai et al., 16 Dec 2025). A plausible implication is that mature personalized gait control motors will increasingly be hybrid systems: personalized plus compliant plus adaptive, with the low-level motor loop kept reliable and deterministic while the high-level gait-state, interaction, and assistance-allocation layers remain user-specific and updateable.