Adaptive Symmetry Learning and Data-Driven Predictive Models for Personalized Control of Robotic Ankle-Foot Prostheses

Abstract

The main goal of this research is to develop a personalized and adaptive control scheme for powered ankle-foot prostheses. Individuals with below-knee amputation generally have gait patterns that are more asymmetrical about their two limbs when compared to the general population. This is likely due to a lack of proper prosthetic ankle function. It has been observed that these individuals compensate for this lack of mobility in ways that can lead to a host of long-term musculoskeletal impairments (e.g., osteoarthritis in the knee and hip of the intact limb). Powered prostheses have the ability to better emulate biological ankles since they are capable of generating power, unlike clinically-prescribed limbs that are completely passive and can only return stored energy. However, current powered prosthesis control methods are over-reliant on able-bodied data, require extensive amounts of tuning by experts, and cannot adapt to the user's unique gait patterns. This work directly addresses all the listed limitations with an adaptive and data-driven control strategy. The controller utilizes time-invariant control signals enabled by phase-based state estimation, an impedance-based feedback loop, and trajectory planning that iteratively adjusts the motion of the prosthetic ankle to match the intact ankle motion. This robotic ankle controller was experimentally evaluated on two individuals with below-knee amputation. The control scheme successfully increased ankle angle symmetry about the two limbs, significantly increased peak prosthetic ankle power output at push-off, and significantly reduced factors associated with osteoarthritis in the intact limb. In addition, control signals were bounded, personalized, and unique across subjects without the use of able-bodied data or extensive tuning. This research also applied deep learning to build models capable of accurately predicting prosthetic ankle torque values from inverse dynamics. Computing inverse dynamics is a time-intensive task, is limited to a laboratory setting, and requires a large array of motion capture markers. The models developed in this research enable total prosthetic ankle torque from inverse dynamics to be observed using signals only from common wearable sensors. This application of machine learning provides an avenue to the development of model-based control systems for powered limbs aimed at optimizing prosthetic ankle torque.