Learning-Based Locomotion Transition Detection: Offline Optimization to Tackle System- and User-Specific Variability in Lower Limb Exoskeletons

Accurate detection of locomotion transitions (e.g., walk-to-sit, stair ascent/descent) is critical for controlling lower limb exoskeletons, as each mode requires specific assistance. Variability in sensor data due to userand system-specific factors challenges nonadaptive classifiers in maintaining accuracy without latency. We evaluated a threshold-based finite-state machine trained via machine learning across 18 subjects using two exoskeletons, eWalk and autonomyo. Transition detection exceeded 90% in most cases, but some transitions reached only 70% median accuracy. The performance was influenced by user behavior and exoskeleton design. To address this, we introduced two offline adaptation methods: a statistics-based approach (SBA) and Bayesian optimization (BO). For eWalk, SBA improved mean accuracies by 2% 8% (e.g., walk-to-si 92% → 94%, sit-to-walk 86% → 90%, stair ascent-to-walk 80% → 88%). BO achieved similar gains and further enhanced stair descent (W SD 88% → 96%, SD W 92% → 98%). For autonomyo, SBA showed no effect, while BO improved accuracies by 7.5% 15% (e.g., walk-to-sit 77.5% → 90%). At the individual level, BO s effect ranged from no benefit to improvements up to 80% for specific transitions. By integrating subject- and system-specific data, this approach provides a reliable, interpretable solution for locomotion transition detection, enhancing personalization and exoskeleton performance..