Real-time gait phase estimation is needed to coordi nate assistance from wearable robots, but relying on footswitches and training a separate model for each user complicate deploy ment. We evaluated an embedded hybrid estimator combining machine-learning inference through a data-trained hidden Markov model (HMM) with model-based tracking through a digital phase locked loop (DPLL) on an adult knee exoskeleton with five healthy adults. During treadmill trials, the complete estimator ran in the 100-Hz control task of an ESP32-S3 microcontroller and used only hip and knee kinematics. Footswitch signals were recorded separately for HMM training and as the operational reference, but were not supplied to the estimator. We compared subject-specific and leave-one-subject-out (LOSO) training through host-side causal replay of the firmware estimator from reset on identical test signals with fold-specific parameters; recorded online estimates were excluded. LOSO yielded a mean absolute phase error of 4.7% of the gait cycle and a balanced accuracy of 75.7% without individual retraining; among the evaluated gait-phase estimation branches, the nominal-map DPLL had the lowest mean absolute phase error, 3.7% of the gait cycle. Variability among participants was greater under LOSO. These results support the feasibility of real-time gait phase estimation on the embedded controller under the tested conditions. Within this preliminary cohort, LOSO suggests that an HMM trained on other healthy adults may generalize to a new user during treadmill walking without individual retraining. Confirmation requires larger cohorts and repeated-session, active-assistance, and impaired-user validation.

cross-validation,embedded system,exoskeletons,gait phase estimation,hidden Markov models,phase-locked loops