Day-To-Day User Adaptation to a Robotic Ankle-Foot Prosthesis

Adaptation in locomotion of lower limb amputees is essential for effective interaction with assistive devices such as robotic ankle-foot prostheses (AFP), which can provide personalized assistance. Bayesian Optimization (BO) and Human-in-the-Loop (HIL) optimization frameworks have been utilized to personalize device parameters based on user feedback; however, these methods may not always fully consider the dynamic nature of human adaptation. This study investigates day-to-day adaptation to a robotic ankle-foot prosthesis by modeling socket pressure distributions across different AFP stiffness conditions using Gaussian processes. We performed a secondary analysis on data from three unilateral transtibial amputees who participated in a four-day experimental protocol involving discrete parameter sweeps and HIL optimization using BO. Gaussian process models were constructed from data collected on the discrete sweep day and the HIL optimization day. Comparative analyses using Kullback-Leibler (KL) divergence and a mean absolute error (MAE) loss metric revealed subject-specific shifts in optimal stiffness values and changes in the pressure landscape between the two days. The average KL divergence was 1.15, and the MAE loss was 0.59, suggesting adaptations in gait and limb dynamics. These findings suggest the importance of using adaptation in a short time period and personalization methods may need to not only optimize device parameters based on static performance snapshots but also accommodate the dynamic shifts that occur as users adapt over time.