Summary

This master's thesis investigates the personalized tuning of the ELSA powered ankle-foot prosthesis using offline human-in-the-loop reinforcement learning. It addresses the time-consuming and subjective manual adjustment of prosthesis control parameters by introducing a learning layer above the existing controller.

The proposed framework combines biomechanical gait deviations with user feedback on comfort and perceived assistance. It learns bounded updates to three interpretable control parameters while preserving the safety guarantees and clinical interpretability of the existing prosthesis controller.

The work contributes a dedicated data-collection protocol, a modular computational pipeline, and an offline evaluation through structured ablation studies. The results provide a proof of concept for integrating subjective human feedback into small-data reinforcement learning for powered prosthesis personalization.

Suggested citation

Vassart, R. (2026). Offline Human-in-the-Loop Reinforcement Learning for Personalized Tuning of a Powered Ankle-Foot Prosthesis. Master's thesis, Université catholique de Louvain (UCLouvain).