Path Planning for Six-joint Manipulators by Generalization from Example Paths

This paper presents a novel kind of generalizing neural storage tailored for the problem of global motion planning for six-joint manipulators in complex, changing environments. Paths are stored in a growing map of neurons with adaptive ellipsoidal receptive eld, called the Ellipsoidal Map. Path planning is based on nding the best-matching neurons for start and goal and a graph search in the neurons' connectivity graph, which gets adapted to the topology of free space. The Ellipsoidal Map remains adaptive all the time: Paths can always be stored and information about collision in planned paths, which occur in particular due to changes of the environment, lead always to corrections of the map. An application to a simulation of the Rotex robot with six degrees of freedom demonstrates highly satisfactory planning skills.