Waste Not, Want Not: Lessons in Rapid Quadrupedal Gait Termination from Thousands of Suboptimal Solutions

Abstract

Elaborate trajectory optimization models with many degrees of freedom can be a useful locomotion-planning tool, as they provide rich solutions that take advantage of the robot's specific morphology. They are, however, prone to falling into local minima. Depending on the seed that initializes the solver, the trajectories themselves and the extent to which they minimize the cost function can vary widely, making it impossible to judge the quality of any solution without generating many more. In this paper, we argue that this perceived drawback can actually be a powerful advantage in exploratory studies, since the resulting set of diverse motions can reveal which features tend to be associated with good performance, and therefore aid in the formulation of strategies for executing challenging maneuvers. We selected rapid gait termination from a high-speed gallop as our case study - a dangerous and scarcely-researched movement. By analyzing a set of over 3000 monopedal and quadrupedal trajectories, we were able to extract conclusions about how braking and sliding should be performed to reduce the stopping distance, and identify a hindlimb action that creates large braking forces.