PaperlayerView exact revision on arXiv

Source-linked AI summary

A Flow-Theoretic Multimodal Control Framework Fusing Mental and Physical Engagement for Upper-Limb Rehabilitation

Jun Leng, Pengcheng Li, Hanze Wang, Shuxiang Guo, Junlin Jiang

arXiv:2610.21590v1cs.RO

TL;DR

Physical-only AAN measures do not distinguish muscular inability from mental disengagement. The paper combines EEG and sEMG in a Flow-grounded adaptive control framework; compared with an sEMG-driven baseline, it reports lower robotic energy expenditure and increased Flow State Ratio.

  • Problem

    Physical-performance metrics used by traditional AAN cannot capture mental focus or distinguish genuine engagement from abnormal co-contraction, leaving users’ needs ambiguous.

    Source: §I · p. 1

  • Method

    The framework fuses EEG-based mental and sEMG-based physical engagement in a four-quadrant fuzzy decision engine that drives adaptive impedance control.

    Source: §I · p. 2

  • Results

    71.9% lower robotic energy expenditure than unimodal sEMG-driven baselines accompanied tighter trajectory tracking, while Flow State Ratio rose from 4.6% to 42.8%.

    Source: §VII · p. 9

  • Takeaways & Limitations

    The reported results support jointly regulating mental and physical engagement to mitigate intent-execution mismatch and passive disengagement in upper-limb rehabilitation.

    Source: §VII · p. 9

    Validation with healthy young adults limits generalizability to elderly stroke populations, whose intent-execution mismatches may be more severe.

    Source: C. Limitations and Future Work · p. 9

Abstract

from arXiv · show

While Assist-as-Needed (AAN) control aims to maximize user participation in robot-assisted training, existing methods relying solely on physical metrics cannot distinguish muscular inability from mental disengagement, leading to suboptimal interventions. To address this ambiguity, a Flow-theoretic multimodal AAN framework is proposed, fusing EEG-based mental engagement and sEMG-based physical engagement into an adaptive impedance control strategy. A four-quadrant fuzzy decision engine translates these dual physiological metrics into continuous control policies. This enables the upper-limb exoskeleton to inject feedforward assistance during intent-execution mismatch (e.g., motor output attenuation despite mental focus) and escalate guiding stiffness to penalize mental distraction. The framework was validated with healthy participants against a state-of-practice sEMG-only AAN baseline. Kinematically and mechanically, the proposed framework reduced robotic energy expenditure by 71.9\% while maintaining precise trajectory tracking, confirming it elicited greater voluntary human effort. Physiologically, the system increased the time spent in the neuroplasticity zone (Flow State Ratio) from 4.6% to 42.8%. These results validate that decoupling and jointly regulating mental and physical states enhances the selectivity of robotic assistance, mitigating user passivity and sustaining the engagement crucial for motor learning.

I. INTRODUCTION

Robot-assisted training can deliver precise, high-dosage guidance, but physical-only AAN measures cannot distinguish motor difficulty from mental disengagement. This work proposes a Flow-grounded system that uses EEG and sEMG to guide adaptive interventions.

I. INTRODUCTION

  • Physical-only AAN metrics cannot reveal mental focus or distinguish genuine engagement from abnormal muscle activity, limiting assessment of users’ needs.The introduction identifies this ambiguity as a barrier to maintaining an optimal training challenge.Source: §I · p. 1
  • The framework uses EEG-based mental engagement and sEMG-based physical engagement as continuous proxies to infer psychological state without directly measuring Flow’s Challenge and Skill coordinates.This approach addresses the impracticality of real-time coordinate measurement and the limits of predefined difficulty or intermittent self-reporting.Source: §I · p. 2
  • A Flow-grounded fuzzy engine and sensorless variable-impedance controller adjust guiding stiffness and feedforward assistance to address intent-execution mismatch and passive disengagement.The multimodal exoskeleton monitors mental and physical engagement during training.Source: §I · p. 2

II. SYSTEM OVERVIEW AND HARDWARE PLATFORM

  • The CURE exoskeleton provides robot-assisted therapy with precise physical guidance for repetitive, high-dosage training.The platform complements therapist-led methods and reduces clinicians’ labor burden.Source: §I · p. 1, §II · p. 2

A. CURE Exoskeleton

The CURE platform combines a wearable exoskeleton with EEG and sEMG sensing. Its control architecture maps filtered physiological signals to adaptive impedance policies.

A. CURE Exoskeleton

  • The CURE exoskeleton uses a Bowden-cable transmission to relocate actuation to a backpack, reducing distal arm mass to approximately 0.7 kg; this study validates control on the 1-DoF elbow joint.The platform supports three degrees of freedom across the elbow and wrist.Source: A. CURE Exoskeleton · p. 2

B. Multimodal Sensing Interface

  • An integrated sensing interface uses EEG to decode mental engagement and biceps and triceps sEMG to measure voluntary effort and antagonist co-activation.Signals are synchronized with the robot’s torque controller at a 100 Hz loop rate.Source: B. Multimodal Sensing Interface · p. 2

III. FLOW-GROUNDED STRATEGY AND DECISION MAKING

  • A hierarchical controller evaluates physiological states at the high level and translates strategic policies into adaptive impedance parameters for physical execution.The architecture separates multimodal decision-making from low-level execution.Source: §III · p. 2

1) Signal Preprocessing:

  • EEG and sEMG are filtered and segmented into 1.0 s windows with 0.25 s steps; EEG engagement combines normalized frontal Theta/Beta Ratio and Alpha Suppression Index measures.Physical engagement uses MVC-normalized sEMG RMS, penalizes antagonist co-contraction, and applies peak-hold smoothing with exponential decay.Source: 1) Signal Preprocessing · p. 3

B. Flow-Grounded Theoretical Mapping

The framework maps mental and physical engagement to four Flow-theoretic states, then assigns a distinct intervention to each state.

B. Flow-Grounded Theoretical Mapping

  • Mental engagement proxies perceived task Challenge, while effective physical engagement reflects execution Skill; their congruence supports reverse inference of psychological state.The mapping avoids direct real-time measurement of abstract Challenge and Skill coordinates.Source: B. Flow-Grounded Theoretical Mapping · p. 3
  • The four quadrants pair high mental and low physical engagement with Anxiety, low mental and high physical engagement with Boredom, high values in both with Flow, and low values in both with Apathy.The mappings correspond to intent-execution mismatch, rote automaticity, synchronous activation, and complete disengagement, respectively.Source: B. Flow-Grounded Theoretical Mapping · p. 4
  • The controller assigns assistance and guidance to Anxiety, correction and alerting to Boredom, maintain-and-challenge to Flow, and safe re-engagement to Apathy.The feature-engineered mapping makes Flow-theory dimensions computable for closed-loop robotic modulation.Source: B. Flow-Grounded Theoretical Mapping · p. 4

C. Decision Engine Implementation

A fuzzy inference engine turns physiological engagement scores into continuous assistance and challenge policies. Gaussian membership functions and stabilization support smooth control despite changing physiological inputs.

C. Decision Engine Implementation

1) Four-Quadrant Fuzzy Inference:

  • Mental engagement uses a 0.4 inflection point, while physical engagement is clamped below 0.3 to classify low muscular effort.The mental threshold accommodates natural attentional fluctuations; the physical threshold is intended to restrict sub-optimal effort.Source: 1) Four-Quadrant Fuzzy Inference · p. 4
  • Gaussian membership functions and weighted-average defuzzification create a continuously differentiable control manifold that bounds impedance changes and avoids the mechanical jerk of Boolean triggers.Assistance demand peaks in Anxiety and tapers when voluntary effort crosses the effective threshold; challenge rises during distraction.Source: 1) Four-Quadrant Fuzzy Inference · pp. 4–5, Figure 3 · p. 4

2) Dual-Stage Decision Stabilizer:

A dwell-time gate and low-pass filtering stabilize both quadrant decisions and continuous control outputs. The stabilized policies map into adaptive assistance, stiffness, and damping while smoothing transitions for safe physical interaction.

2) Dual-Stage Decision Stabilizer:

  • A pending quadrant becomes confirmed only after persisting for T_dwell; earlier fluctuations reset the timer, rejecting high-frequency physiological artifacts.A first-order low-pass filter also smooths continuous assistance and challenge outputs, keeping strategic mode switches deliberate while parameter adjustments remain continuous.Source: 2) Dual-Stage Decision Stabilizer · p. 5

A. Adaptive Parameter Synthesis

  • Strength-reduction and co-activation penalty factors modulate user control authority; Flow retains full voluntary authority, while Anxiety and Apathy assign zero factors.Boredom uses a factor of 0.5; the authors state that zero factors in Anxiety or Apathy transfer full authority to the robot for stable guidance.Source: A. Adaptive Parameter Synthesis · p. 5
  • Assistance force is synthesized from the state-dependent policy, and any spontaneous valid physical effort immediately attenuates robotic output, preserving AAN behavior.The passage describes this as zero-latency assistance withdrawal.Source: A. Adaptive Parameter Synthesis · p. 5
  • Stiffness follows the challenge policy, while damping increases with co-activation, making excessive co-contraction feel heavier and encouraging relaxed motor synergies.The control maps challenge and the macroscopic state into stiffness, damping, and assistance force.Source: A. Adaptive Parameter Synthesis · p. 5
  • An exponential moving average bridges synthesized targets for stiffness, damping, and assistance to commanded parameters, smoothing discrete modulation changes into haptic interventions without mechanical jerk.The filter uses α=0.8 at 100 Hz; a dual-stage stabilizer also bounds transition frequency.Source: A. Adaptive Parameter Synthesis · p. 5

B. Variable Impedance Control

A position-based variable-impedance controller executes the high-level policies through feedback impedance and feedforward assistance, without relying on potentially unstable force-feedback loops.

B. Variable Impedance Control

  • The position-based Variable Impedance Control scheme modulates interaction compliance directly to support intrinsic safety and robustness against sensor noise in dynamic settings.Source: B. Variable Impedance Control · p. 5
  • Command torque combines feedback impedance, feedforward assistance, and static friction compensation.Source: B. Variable Impedance Control · p. 5
  • The controller decouples time-varying stiffness and damping from feedforward assistance, enabling independent execution of multidimensional AAN interventions directed by high-level strategies.The desired and actual joint positions are q_d and q, respectively.Source: B. Variable Impedance Control · p. 5

V. EXPERIMENT AND CHARACTERISTIC EVALUATION

The validation procedure combined individualized calibration with two studies, and the experimental setup paired active tracking with visual feedback.

V. EXPERIMENT AND CHARACTERISTIC EVALUATION

  • The procedure comprised individualized calibration followed by two primary validation studies.Fig. 5 depicts calibration, a 240-second continuous adaptive response test, and a five-minute comparative efficacy analysis against baseline AAN.Source: §V · p. 6
  • During active tracking, the interface displayed the target trajectory, the user’s real-time cursor, and synchronized phase prompts.Source: Figure 4 · p. 6
  • The experimental overview and interface screenshot document the tracking scenario and its visual feedback.Source: Figure 4 · p. 6

A. Experimental Participants and Setup

Ten healthy volunteers completed individualized calibration and protocol phases designed to characterize adaptive responses and compare efficacy against a baseline EMG-driven AAN mode.

A. Experimental Participants and Setup

  • Ten healthy, right-handed volunteers performed one-degree-of-freedom elbow tracking with a 1.2 kg wrist load to sustain muscular effort and induce natural fatigue.Participants tracked a 0.2 Hz sinusoidal target spanning 0°–90° using a cursor and synchronized audio-visual cues.Source: A. Experimental Participants and Setup · p. 6
  • Individual EEG normalization used resting, passive-following, and focused active-tracking conditions; sEMG normalization used a five-second MVC test.Source: A. Experimental Participants and Setup · p. 6

B. Experiment Protocol

  • The protocol was organized into three sequential phases to establish physiological baselines, characterize dynamic stability, and evaluate functional efficacy.Source: B. Experiment Protocol · p. 6, Figure 5 · p. 6
  • Preliminary one-minute isolated trials established decision-engine inflection points using focused versus distracted tracking and zero-assist versus passive-following conditions.The stated inflection points were 0.4 and 0.3.Source: B. Experiment Protocol · p. 6
  • A 240-second adaptive-response session induced Flow, Boredom, Anxiety, and Apathy at 60-second intervals, with transition latency and maximum assistive-force derivative used to assess stability.Source: B. Experiment Protocol · p. 6
  • Three five-minute sessions compared Fusion Mode with baseline EMG-driven AAN using tracking RMSE and raw assistive force; offline iEMG and EEG PSD supported Flow State Ratio assessment.The baseline decays feedforward assistance with physical engagement while keeping guiding stiffness constant.Source: B. Experiment Protocol · p. 6

C. Experimental Results

Preliminary tuning established separable physiological boundaries and data-driven thresholds; the continuous test then showed state-specific interventions and smooth transitions across the four induced states.

1) Results of Preliminary Parameter Tuning:

  • The 0.3 physical-engagement and 0.4 mental-engagement thresholds separate the high-state 95% confidence intervals, whereas a symmetric 0.5 threshold would overlap and risk punitive control.The tuning data came from isolated one-minute blocks across all 10 participants.Source: 1) Results of Preliminary Parameter… · p. 7

2) Results of Continuous Adaptive Response Test:

  • Across four induced states, the system minimizes Flow assistance while maintaining firm guiding stiffness, withholds power and applies a haptic constraint in Boredom, lowers Apathy stiffness to about 2 N·m/rad, and injects feedforward assistance during Anxiety-related intent-execution mismatch.Source: 2) Results of Continuous Adaptive… · p. 7
  • Across 23 inter-quadrant shifts, transition latency averaged 3.10 ± 0.87 s and assistance-force change averaged 1.86 ± 1.03 N/s, supporting smooth, chatter-free interventions.Fig. 7 presents boxplots of response latency and maximum assistance-force change.Source: 2) Results of Continuous Adaptive… · pp. 7–8, Figure 7 · p. 7

3) Results of Comparative Efficacy Analysis:

Compared with baseline AAN, Fusion Mode reduced robotic assistance while maintaining comparable tracking and increased Flow-state time. The framework pairs assistance during motor-output attenuation with increased stiffness during distraction, at the cost of greater deployment complexity.

3) Results of Comparative Efficacy Analysis:

  • 72.4% lower mean robotic assistive force with comparable trajectory tracking characterized Fusion Mode versus baseline AAN; Flow State Ratio rose from 4.6% to 42.8%.Fig. 8 reports tracking and assistive force; Fig. 9 shows state distributions and Flow State Ratio.Source: 3) Results of Comparative Efficacy… · p. 8, Figure 8 · p. 8, Table III · p. 8
  • During motor-output attenuation, assistance reached about 3.5 N; during visual distraction, stiffness rose to about 14 N·m/rad and cortical re-engagement followed within 5–10 s.These are representative microscopic windows from the first Fusion Mode trial of Subject 2.Source: 3) Results of Comparative Efficacy… · p. 8

A. Resolving Intent-Execution Ambiguity in Multimodal AAN Control

  • The framework uses a continuously differentiable fuzzy manifold to diagnose intent-execution mismatch and deliver chatter-free assistance or haptic penalties.It assists during motor-output attenuation and penalizes slacking during Boredom.Source: A. Resolving Intent-Execution… · p. 8
  • Simultaneous EEG and sEMG preparation, individualized fuzzy-parameter tuning, and computational overhead make deployment less efficient than unimodal systems.The study validates control on the 1-DoF elbow joint of a three-DoF platform.Source: A. Resolving Intent-Execution… · p. 9

B. Multimodal Integration and Functional Implications

The framework uses multimodal physiological signals to adapt robotic assistance, with reported reductions in robotic energy and increases in Flow State Ratio. Its validation and fatigue monitoring remain bounded by the tested population and the current disengagement-based fatigue limitation.

B. Multimodal Integration and Functional Implications

  • Physiological smoothing and a macroscopic Dwell-Time Stabilizer are reported to reduce BCI chattering and support safe impedance modulation despite biosignal non-stationarity.The physical engagement metric penalizes antagonist co-activation, while insufficient challenge triggers Boredom haptic penalties.Source: B. Multimodal Integration and… · p. 9
  • 72.4% lower robotic energy and a nearly tenfold increase in Flow State Ratio were reported for Fusion Mode, which penalizes distraction and assists during motor output attenuation.The passage frames this as a transition from execution-aware to intent-aware training relative to EMG-driven baselines.Source: B. Multimodal Integration and… · p. 9

C. Limitations and Future Work

  • Validation with healthy young adults limits generalizability to elderly stroke populations, where severe neurologically derived intent-execution mismatches remain to be tested.The passage calls for longitudinal trials with age-matched stroke survivors to validate parameter generalizability and therapeutic robustness.Source: C. Limitations and Future Work · p. 9
  • The Apathy quadrant may not detect muscle fatigue that arises before complete mental disengagement.The authors propose adding an sEMG spectral-shift safety layer to monitor fatigue and modulate impedance or initiate rest.Source: C. Limitations and Future Work · p. 9

VII. CONCLUSION

  • 71.9% lower robotic energy expenditure and Flow State Ratio rising from 4.6% to 42.8% were reported in the conclusion versus unimodal sEMG-driven baselines.The conclusion also reports tighter trajectory tracking and describes feedforward assistance during motor output attenuation and increased guiding stiffness during mental distraction.Source: §VII · p. 9
Loading 2610.21590v1…