Adaptive Lower-Limb Prosthetic Control: Towards Personalized Intent Recognition & Context Estimation

Abstract

Historical advancements in lower-limb prostheses have reflected the challenges of diverse anthropomorphic biomechanics, limiting intelligent control systems from being implemented and reflecting true user intent. With recent advancements in machine learning (ML), however, this notion is being challenged. In transfemoral-powered prostheses, time series information has been used to infer context (slope angle and walking speed) and intent (ambulation mode) and scale torque assistance accordingly in real time. In this study, we build off this work by proposing and validating a real-time framework for adaptive walking speed context estimation. Our system makes use of the general similarity in human gait patterns and iterates subject-independent ML models used for prediction towards subject-dependent models by method of batched retrospective labeling and retraining. Offline validation for walking speed estimation has been completed using seven amputee subjects' data, showing an average subject-independent MAE of 0.063 being reduced to 0.043 m/s, a 31.7% improvement. In addition, we discuss and present preliminary results for walking speed estimation and several alternative methods of retrospective labeling.