Connection-Based Data-Driven Gait Modeling of a Quadruped

We have recently shown that connection-based models arising from geometric mechanics apply to legged systems, both biological and robotic, whether they slip or maintain non-slip contacts with the substrate. A key assumption that underlies these physics – that friction annihilates momentum quickly – breaks down for large trotting quadrupeds, raising the question of how well data-driven connections approximate their observed motion. We report initial results from a Ghost Robotics Vision 60 robot, measuring ego motion using iterative closest point (ICP) estimation from LIDAR data to solve for the body velocity, and using encoder data for body shape. In its trotting gait the robot approximately conserves momentum around the line connecting its stance feet, yet the connection term of the reconstruction equation, which ignores momentum, accounted for 97% of observed forward body velocity. The Vision 60 system will allow us to test how well the connection captures the physics of real-world surfaces and motions.