Learning to Plan Near-Optimal Collision-Free Paths

Abstract

A new approach to find a near-optimal collision-free
\npath is presented. The path planner is an implementation
\nof the adaptive error back-propagation algorithm
\nwhich learns to plan “good”, if not optimal,
\ncollision-free paths from human-supervised training
\nsamples.
\n
\nPath planning is formulated as a classification
\nproblem in which class labels are uniquely mapped
\nonto the set of maneuverable actions of a robot or
\nvehicle. A multi-scale representational scheme maps
\nphysical problem domains onto an arbitrarily chosen
\nfixed size input layer of an error back-propagation
\nnetwork. The mapping does not only reduce the size
\nof the computation domain, but also ensures applicability
\nof a trained network over a wide range of
\nproblem sizes. Parallel implementation of the neural
\nnetwork path planner on hypercubes or Transputers
\nbased on Parasoft EXPRESS is simple and efficient,
\nSimulation results of binary terrain navigation indicate
\nthat the planner performs effectively in unknown
\nenvironment in the test cases.