Motion Planning Networks: Bridging the Gap Between Learning-based and\n Classical Motion Planners

Abstract

This paper describes Motion Planning Networks (MPNet), a computationally\nefficient, learning-based neural planner for solving motion planning problems.\nMPNet uses neural networks to learn general near-optimal heuristics for path\nplanning in seen and unseen environments. It takes environment information such\nas raw point-cloud from depth sensors, as well as a robot's initial and desired\ngoal configurations and recursively calls itself to bidirectionally generate\nconnectable paths. In addition to finding directly connectable and near-optimal\npaths in a single pass, we show that worst-case theoretical guarantees can be\nproven if we merge this neural network strategy with classical sample-based\nplanners in a hybrid approach while still retaining significant computational\nand optimality improvements. To train the MPNet models, we present an active\ncontinual learning approach that enables MPNet to learn from streaming data and\nactively ask for expert demonstrations when needed, drastically reducing data\nfor training. We validate MPNet against gold-standard and state-of-the-art\nplanning methods in a variety of problems from 2D to 7D robot configuration\nspaces in challenging and cluttered environments, with results showing\nsignificant and consistently stronger performance metrics, and motivating\nneural planning in general as a modern strategy for solving motion planning\nproblems efficiently.\n