Towards a computational toolbox for resolving the multiscale dynamics of animal and robot behavior

The complexity of animals' and robots' coupled, nonlinear hybrid (contact-making and -breaking) dynamics can be mitigated by postulating that agile behaviors arise from composition of favorable postural subspaces within which constituent tasks are encoded as "templates" (low degree of freedom dynamical abstractions) and down toward which the high degree of freedom body and limbs are "anchored" (quickly controlled to converge). To take the next step towards discovery of the template representations that enable agile animal locomotion, we developed a computational pipeline that extracts such low dimensional attracting invariant postural submanifolds from high dimensional trajectory data. Growing evidence from robotics suggests the utility of hierarchical compositions whereby spatially active bodies anchor sagittal plane templates, which, in turn, anchor still lower dimensional cycles whose postural loci within the plane determine the abstract behavioral pattern, for example, the difference between bounding, trotting or pronking. This talk will discuss our recent work refining the selection of pipeline parameters suitable for resolving such nested multiscale hierarchical compositions.