A preliminary study on deep neural network-based adaptive control for robotic ankle-foot prosthesis with speed and terrain flexibility

Document Type

Article

Publication Date

3-4-2026

Abstract

Powered prostheses assist lower-limb amputees by providing positive mechanical energy during ambulation. Optimal control parameters are crucial in providing proper support and smooth gait transition to the user. However, tuning remains a complex process. This research proposes a long short-term memory (LSTM)-based deep neural network (DNN) system to adaptively tune the virtual stiffness parameters of the robotic prosthesis across multiple walking speeds and terrains. It involves data collection from 19 healthy subjects performing walking, stairs ascent/descent and running. The DNN is then trained and its generalisability is tested with the presence of load. Overall, the trained model achieves a normalised root mean squared error (NRMSE) below 0.09, a correlation coefficient above 0.98 and a lag time below 100  µs, demonstrating high generalisation and strong adaptibility for prosthesis impedance control.

Keywords

Powered prosthesis, adaptive mechanism, deep neural network, environment uncertainty, ambulation speed

Publication Title

International Journal of Control

ISSN

0020-7179

DOI

10.1080/00207179.2025.2537253

Volume

99

Issue

3

First Page

783

Last Page

794

Publisher

Taylor & Francis

Share

COinS