Locomotion Mode Transitions: Tackling System- and User-Specific Variability in Lower-Limb Exoskeletons

1. Problem Statement

Effective control of lower-limb exoskeletons requires accurate, real-time detection of locomotion mode transitions (e.g., walking to sitting, walking to stair ascent/descent). While steady-state activity recognition is well-established, transition detection remains challenging due to:

• Variability: Significant differences in user gait patterns and mechanical constraints imposed by different exoskeleton designs (e.g., degrees of freedom, weight, actuation).
• Limitations of Fixed Models: Non-adaptive classification models (using fixed thresholds) often fail to maintain high accuracy across diverse users and devices, leading to latency or missed transitions.
• Trade-offs: Deep learning offers high accuracy but lacks interpretability and requires heavy computational resources, while traditional machine learning (ML) often requires extensive feature engineering.

The core problem addressed is how to adapt transition detection algorithms to specific users and exoskeleton systems without incurring prohibitive computational costs or lengthy training times.

2. Methodology

A. Experimental Setup & Hardware

• Subjects: 18 healthy adults (10 for each exoskeleton) aged 18–65.
• Exoskeletons:
  • eWalk: A lightweight device with 1 active degree of freedom (DOF) at the hip.
  • autonomyo: A more complex device with 3 active DOFs per leg (hip flex/ext, hip abd/add, knee flex/ext) and 3 passive DOFs.
• Tasks: Four locomotion modes (Walking, Sitting, Stair Ascent, Stair Descent) and six transitions between them.
• Sensors: Thigh angle/velocity, ground reaction forces (GRF), and inertial measurement units (IMU).

B. Core Algorithm: Finite State Machine (FSM) with ML-Trained Thresholds

The study utilizes a Finite State Machine (FSM) with four states. Instead of complex classifiers running in real-time, the authors use an ML-trained Threshold-Based (TH) method:

• Features: Instantaneous Characteristic Features (ICFs) are extracted from thigh angle ($\theta_{th}$) and velocity ($\dot{\theta}_{th}$) at specific gait events (Maximum Hip Flexion, Heel Strike, and specific angle ranges).
• Training: Offline ML models (Linear SVM and Logistic Regression) are trained to determine optimal decision boundaries. These boundaries are converted into simple if/else thresholds for real-time embedded execution.

C. Addressing System Variability: Joint Misalignment Correction

To handle kinematic discrepancies between the exoskeleton joints and human anatomy (common in the complex autonomyo design), a regression-based mapping algorithm was developed.

• Method: A function $f(w, \phi(x)) \to t$ maps measured encoder data ($x$, velocity, acceleration) to corrected angles ($t$) based on literature ground truth.
• Optimization: Weights ($w$) are optimized using a sum-of-squares loss function to minimize the error between measured and natural human biomechanics.

D. Addressing User Variability: Personalization Strategies

Two methods were proposed to adapt the detection thresholds to individual users:

1. Statistics-Based Approach (SBA):

   • Analyzes statistical differences (mean and standard deviation) between the training population and the specific user.
   • Scales the original thresholds using a ratio of these statistics to adjust for the new user's gait characteristics.
   • Formula: $TH_{new} = TH_{tr} \cdot \frac{\text{mean}(ICF_{new}) \pm \text{std}(ICF_{new})}{\text{mean}(ICF_{tr}) \pm \text{std}(ICF_{tr})}$

2. Bayesian Optimization (BO):

   • Treats threshold tuning as an optimization problem to minimize a cost function (penalties for missed or delayed transitions).
   • Uses Gaussian Processes to model the objective function and an acquisition function to guide the search for optimal thresholds with minimal experiments (Human-in-the-Loop).
   • Includes regularization to prevent thresholds from drifting into unrealistic values.

3. Key Contributions

1. Personalization Framework: Introduction of SBA and Bayesian Optimization to adapt transition detection thresholds to specific users and exoskeleton designs, significantly outperforming non-personalized methods.
2. Joint Misalignment Correction: Development of a data-driven regression model to correct kinematic misalignments in complex exoskeletons, improving the fidelity of joint trajectory data.
3. Open-Source Dataset: Release of a comprehensive lower-limb kinematics dataset involving 18 subjects, two distinct exoskeletons, and four locomotion modes to facilitate future research.
4. Computational Efficiency: Demonstration that ML-trained thresholds combined with BO offer a computationally efficient alternative to deep learning, suitable for embedded systems.

4. Results

• Joint Alignment: The mapping function reduced the Mean Root Mean Squared Error (RMSE) of hip angles by 3.61° and aligned the Mean Maximum Hip Flexion (MHF) closer to literature values (from 73.87° to 55.27°).
• Detection Accuracy:
  • Baseline: The non-personalized method achieved varying accuracy (e.g., 70–100% depending on the transition and device).
  • SBA: Improved accuracy for the simpler eWalk but showed limited gains for the complex autonomyo.
  • Bayesian Optimization: Achieved the highest performance, increasing detection accuracy by up to 80% in specific scenarios compared to the baseline.
    • Example: For eWalk Subject 5, W-SD accuracy jumped from 20% to 100%.
    • Example: For autonomyo Subject 6, S-W and W-S accuracy improved by 40% to reach 100%.
• Efficiency: Bayesian Optimization was 4x faster than Grid Search (GSA) in finding optimal thresholds (approx. 2 minutes on a standard PC vs. 8+ minutes for GSA), making it viable for real-time or near-real-time tuning.

5. Significance and Conclusion

This study highlights that personalization is critical for the effective deployment of assistive exoskeletons. A "one-size-fits-all" control strategy fails to account for the inter-subject variability in gait and the specific mechanical constraints of different devices.

• Adaptive Control: The proposed methods allow control systems to dynamically adjust to individual users, enhancing safety, stability, and user experience.
• Scalability: By using lightweight, threshold-based logic enhanced by efficient optimization (BO), the approach bridges the gap between high-accuracy machine learning and the computational constraints of embedded medical devices.
• Future Impact: The work paves the way for more responsive, user-centered assistive technologies that can be rapidly tuned to new users without extensive manual calibration or expensive hardware requirements.

The authors conclude that while threshold optimization is powerful, future work must address non-optimized FSM components and include users with mobility impairments to fully validate the system's robustness in clinical settings.