Reducing model fidelity for statistical path characterization of an insect-scale robot

Studies of both insect and robot locomotion have shown the usefulness of open-loop control for traversing heterogeneous terrain. An unanswered question is how to incorporate the effectiveness of open-loop control on feature-laden terrain with path planning techniques. We previously studied open-loop running across 3D printed models of feature-laden terrains with the 1.4~g quadrupedal Harvard Ambulatory Microrobot. We tested running frequencies both at and above the robot's center-of-mass resonances. When running near resonance, random perturbations from limb-ground interactions resulted in a mean-square displacement characteristic of diffusion at time scales of a few steps, resulting in trajectories with stochastic properties. In this work, we explore a modelling technique for an insect scale, robotic locomotor that abstracts the complex leg and foot-ground interactions by leveraging past experimental data. The model developed treats the problem of trajectory planning as a statistical one, abstracting away complex dynamics models which become increasingly inaccurate as terrain complexity leads to harder to predict foot-ground interactions. This reduction of model fidelity leads to a statistical characterization of the robot's trajectory for use in terrain exploration and mapping.