Application of neural networks and evolutionary programming to path-planning of planar manipulators

Abstract

In this thesis, a new approach to path-planning is proposed. A proposed path-planning uses a neural network and evolutionary programming as its learning algorithm. A neural network is used to produce a control sequence of a robot manipulator. Evolutionary programming is adopted as a learning algorithm of a neural network. This method takes the robot dynamics into consideration while the path is searched with a cost function that includes input energy, fixed time constraint, position error, and so on. The advantage of this method is that it can simply find a cost optimal path of a robot manipulator without any complex mathematical formulation. Depending on cost functions, the proposed path-planning can result in not only a simple path but also a collision-free path. In addition, this method result in a universal path-planning. The neural network learns several paths that lead to fixed points in advance. Then a universal path-planning refers to obtaining the path that leads to a point in the vicinity of fixed points without learning phase of a neural network. A universal path-planning owes the generalization capability of neural networks.