Meeting Abstract

57-3  Tuesday, Jan. 5 10:45  Estimating predictive dynamical models of legged locomotion from data BANJANIN, B*; BURDEN, SA; MOORE, TY; REVZEN, S; FULL, RJ; University of Washington, Seattle; University of Washington, Seattle; Harvard University; University of Michigan, Ann Arbor; University of California, Berkeley sburden@uw.edu

We seek dynamical models of legged locomotion that provide quantitative predictions for the response to novel perturbations and experimental treatments. Purely data-driven models cannot provide predictions for behaviors not included in the training dataset. Analytical reduced-order models crafted to study steady-state behaviors may give poor predictions following perturbations that induce transient behavior. We propose a method that mediates between these two extremes by estimating parameters for a family of spring-mass models directly from data. To test the proposed method, we made use of an experimental dataset involving large lateral perturbations and inertial load treatments applied to running cockroaches (Blaberus discoidalis), where kinematics of body and limbs were measured and dynamic quantities (velocities and accelerations) were estimated using a tuned Kalman filter. By fitting parameters for a potential energy function using an ensemble of stepwise acceleration time series, we estimate a predictive model that is piecewise-Hamiltonian, a piecewise-defined dynamical system that switches between conservative subsystems at step transitions. The resulting dynamical model provides better agreement with population-level average kinematic measurements and dynamic estimates observed in our dataset than existing analytical models. In future work, we will compare the estimated model's response to the measured perturbation and experimental treatment. More broadly, parametric models fit to empirical data provide an intermediary between data and analysis, enabling the experimentalist to generate new testable hypotheses and the analyst to reject inconsistent model mechanisms.