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

Accurate detection of locomotion transitions, such as walk to sit, walk to stair ascent, and descent, is crucial to effectively control robotic assistive devices, such as lower-limb exoskeletons, as each locomotion mode requires specific assistance. Variability in collected sensor data introduced by user- or system-specific characteristics makes it challenging to maintain high transition detection accuracy while avoiding latency using non-adaptive classification models. In this study, we identified key factors influencing transition detection performance, including variations in user behavior, and different mechanical designs of the exoskeletons. To boost the transition detection accuracy, we introduced two methods for adapting a finite-state machine classifier to system- and user-specific variability: a Statistics-Based approach and Bayesian Optimization. Our experimental results demonstrate that both methods remarkably improve transition detection accuracy across diverse users, achieving up to an 80% increase in certain scenarios compared to the non-personalized threshold method. These findings emphasize the importance of personalization in adaptive control systems, underscoring the potential for enhanced user experience and effectiveness in assistive devices. By incorporating subject- and system-specific data into the model training process, our approach offers a precise and reliable solution for detecting locomotion transitions, catering to individual user needs, and ultimately improving the performance of assistive devices.

💡 Research Summary

This paper addresses the critical problem of detecting locomotion mode transitions—specifically walk‑to‑sit, walk‑to‑stair ascent, and descent—in lower‑limb exoskeletons. Accurate, low‑latency transition detection is essential for providing appropriate assistance, yet variability introduced by individual users and by the mechanical design of different exoskeletons hampers the performance of static, non‑adaptive classifiers.

The authors first evaluate a previously developed machine‑learning‑trained threshold (ML‑TH) method on two distinct exoskeleton platforms: the lightweight eWalk (2 active degrees of freedom) and the more complex autonomyo (3 active degrees of freedom with additional passive joints). The ML‑TH approach converts offline‑trained linear SVM or logistic regression models into simple one‑dimensional thresholds applied within a finite‑state machine (FSM) that governs four states (Sit, Walk, Stair Ascent, Stair Descent) and six transition events. While the baseline method achieves respectable overall accuracy (≈84 %), its performance degrades markedly for subjects whose gait patterns differ from the training set, especially when using autonomyo, where atypical sitting dynamics and “running‑like” stair climbing cause missed detections or false alarms.

To overcome these limitations, the study introduces two personalization strategies that adapt the FSM thresholds to subject‑ and system‑specific characteristics:

1. Statistics‑Based Adaptation (SBA).
   SBA computes the mean and standard deviation of the instantaneous characteristic features (ICFs) for a new user and rescales the original training thresholds according to the ratio of new‑to‑training statistics (Equation 6). This method is computationally lightweight, requires only a short calibration recording, and can be implemented directly on embedded hardware. It effectively lowers (or raises) thresholds when the new user’s feature distribution is shifted relative to the training population, thereby reducing missed detections caused by overly conservative thresholds.

2. Bayesian Optimization (BO)‑Based Tuning.
   BO treats each threshold as a continuous optimization variable and searches for the set that minimizes a performance‑based objective function J (e.g., a weighted combination of detection accuracy, latency, and false‑positive rate). Because J is not analytically defined, a Gaussian Process surrogate model with a radial‑basis‑function kernel approximates the relationship between thresholds and performance. An acquisition function (Equation 8) balances exploration and exploitation, and the algorithm iteratively refines the surrogate using a limited number of experimental evaluations. The search space for each transition pair is bounded (e.g.,