Towards a quadtree based approach to learn local plans in robotic motion planning

Abstract

In this paper, a novel approach to the local planning of mobile robots and autonomous vehicles is discussed. This paper introduces a motion planning architecture utilizing both a conventional (Hybrid A*) and a learning-based planner, inspired by the recent results of reinforcement learning. The presented approach relies on a grid-based representation of the environment which is simultaneously used for planning and learning of such trajectories. The representation grid is derived from a quadtree representation of the environment and the definition is extended with convex polytopic description, to produce grid-based and Voronoi diagrams. The paper also discusses the possible integration of more sophisticated soft-computing-based control, like TP-model transformation as a basis for the heuristics used by motion planning components.