Why Hard Code the Bionic Limbs When They Can Learn From Humans?

Abstract

In this paper, we propose a task-generic learning-based model for the control of a powered ankle exoskeleton. In contrast to the traditional state machine-based control approaches that hard codes the transition heuristics for the different states and motion conditions during gait, we propose to learn the finer constraints of gait from multiple demonstrations of human gait. We validate our proposed approach on a dataset of ten subjects walking on various inclines and at multiple speeds. We deploy our model on an ankle exoskeleton, and conduct user studies on able-bodied subjects who perform gait scenarios across varying speeds and inclines. We conduct multiple online experiments to validate our learning-based approach for different motion conditions, e.g., normal walking, walking at different speeds and inclines, turns, cross-overs with variable speed and cadence, walking on a treadmill as well as on level ground. We find that our proposed learning-based model has the capability to extrapolate its learned decision rules to support untrained gait conditions, for, e.g., walking at higher speeds and inclines not seen during training. The subjects were able to adapt to the different gait scenarios comfortably without loss of stability.