Human Demonstrations Enable Efficient Solutions to Sequential Manifold Planning Problems

Breanne Crockett, Carl L. Mueller, Bradley Hayes

Abstract: Constrained motion planning algorithms generate solutions to planning problems that require robots to adhere to rigid behavioral restrictions. Traditionally, these problems are segmented and framed as requiring adherence to a single set of constraints throughout, a limiting compromise that results in motion planning over a single manifold. When a trajectory must comply with multiple sets of constraints in a single motion planning problem, each constraint set defines an implicit manifold within the planning space and solutions must sequentially traverse manifold intersections. This is known as a Sequential Manifold Planning Problem (SMPP). The choice of manifold intersection points plays a critical role in solving SMPPs, as a particular intersection point may not admit a path to a subsequent constraint manifold, preventing motion planners from finding solutions in reasonable or even finite time. Many works assume intersection point independence, requiring all intersection points to lead to viable solutions. We show how Learning from Demonstration models intrinsically define an SMPP and contribute an algorithm for Intersection Point Dependence Relaxation using distributions extracted from these models near constraint set transitions. These distributions, learned from human demonstrations, supply candidate points for an optimization process to identify intersection points that admit solutions, solving SMPPs with greater efficiency than uninformed approaches and relaxing intersection point dependence even when the demonstrator is noisy (i.e., out of adherence to constraints).