Robust gait stability analyses using dynamical machine learning

While animals move through the world with apparent ease and stability, underlying these movements are complicated neuromechanical dynamics that we have had difficulties replicating in simulations or robots. In particular, understanding how gait is stabilized has proven difficult given the complex and intertwined dynamics of muscles controlling locomotion and the neural processes that allow an animal to adapt to perturbations in its environment. Floquet Theory is a promising data-driven methodology for quantifying and understanding gait stability, combining neural and mechanical dynamics into a single framework. However, its application has been limited due to the amount of data it requires and its sensitivity to measurement noise. Here we leverage dynamical system theory and generative machine learning models, combined with finite-size corrections, to enhance the accuracy, and robustness of Floquet calculations starting from realistic data set sizes. We show applications of this method from fly, mouse, and human gait dynamics. This approach opens avenues for the exploration of neuromuscular mechanisms of gait stabilization and can be readily applied to the study of periodic dynamics more broadly.