Data-driven discovery of hierarchical neuromechanical systems

Dynamic locomotion behaviors result from high dimensional, nonlinear, and dynamically coupled interactions between an animal or robot and its environment. The templates and anchors hypothesis resolves this complexity by postulating that preferred postural degrees of freedom for a specific task are anchored - rendered stable and robust to perturbations - by fast neural and mechanical feedback control forces. However, due to the lack of a general methodology for identifying templates from behavioral data, the use of templates to study dynamic locomotion behaviors in animals, and translate such behaviors to robophysical systems, has largely been limited to a few well-studied examples. Further, reliance upon existing analytic models limits the ability to discover new mechanisms in rich datasets. The promise of big kinematic datasets from new automated labeling methods motivates the aim for a general, data-driven paradigm to identify templates in motion data. We present a framework for identifying these dynamic posture principles grounded in a local model of the hypothesized template-anchor dynamics, enabling generalization to a range of both periodic and transient locomotion tasks. Using interpretable models of both continuous and hybrid template-anchor systems, we find that our approach is capable of effectively identifying template submanifolds with a surprisingly wide range of geometric and dynamical properties, and is robust to measurement noise and parameter errors.