A Personalized Lightweight Framework for Kinematic and Dynamic Estimation in Exoskeleton-Assisted Locomotion Under Varying Walking Conditions

Existing methods for exoskeleton kinematic and dynamic estimation often rely on multi-channel electromyographic (EMG) signals and extensive training data, yet they frequently exhibit accuracy degradation under varying gait conditions. To address these issues, this paper proposes a lightweight, personalized, and adaptive framework that uses EMG signals to estimate ankle joint angle and torque in real time during variable-gait walking, both with and without exoskeleton assistance. The framework integrates an autoregressive module, a multilayer perceptron, and Bayesian-optimized time-delay inputs, with personalized parameter selection to account for individual physiological differences. It is designed to be lightweight, utilizing signals from only two muscles (soleus and tibialis anterior) and only 160 seconds of training data across eight walking conditions. An adaptive unscented Kalman filter is employed to enhance robustness via sensor fusion and noise suppression, enabling dynamic adaptation to gait changes. Experiments involving ten participants under high-speed and varying-gait conditions demonstrated that the proposed method consistently outperforms four baseline models in terms of root mean square error (RMSE). Without exoskeleton assistance, it reduced angle RMSE by up to 32.8% ± 9.1% and torque RMSE by 16.0% ± 13.3% during high-speed walking. With assistance, RMSE decreased by 22.2% ± 12.8% for angle and 7.9% ± 8.8% for torque. In varyinggait scenarios, it maintained stable accuracy, achieving RMSE reductions of up to 22.3% (angle) and 15.6% (torque) without assistance, and 14.6% and 7.5% with assistance. Robustness analysis under steady walking conditions revealed limited error variability, with maximum standard deviation increases of 18.0% (angle) and 10.0% (torque) in zero-torque mode, and 15.9% and 19.7% under adaptive torque, indicating strong resilience to gait dynamics. These results confirm the method’s ability to provide precise, continuous, and personalized motion intent estimation, thereby enhancing exoskeleton control for rehabilitation and daily mobility applications.