Matt Maly, Guiding Motion Planners through Low-Dimensional Projections

The traditional problem of motion planning for a robot with dynamics requires computing a trajectory from a starting position to a goal that respects the robot's model of dynamics and avoids collisions with obstacles in the environment. SyCLoP (Synergistic Combination of Layers of Planning) is a framework that was proposed to solve challenging motion planning problems with dynamics by exploiting the discrete nature of such problems. SyCLoP consists of two layers: (1) a continuous layer, which includes a robot's state space and model of dynamics as well as a sampling-based motion planner, and (2) a discrete layer, which creates a graph-based decomposition of the robot's environment. By sharing information between its continuous and discrete layers when constructing a motion plan, SyCLoP significantly improves the performance of a traditional sampling-based motion planner, with reported speedups of up to two orders of magnitude.

This talk will present a generalized version of SyCLoP that constructs its discrete layer over a subspace defined by any projection from the continuous state space. We will discuss experimental results with various types of projections, including those obtained by applying dimensionality reduction to the continuous state space, and we will compare SyCLoP's performance with these projections to performance with the original environmental projection.