A neural network controller of nonholonomic systems using a specialized training technique

摘要

A novel method of constructing training patterns for nonholonomic control problems is proposed. The patterns are designed to produce neural networks which can guide certain constraint restricted wheeled mobile robots. The task of guiding such vehicles from one state configuration to another can be highly nontrivial, and has been termed the nonholonomic path planning problem. Interestingly enough, many such problems arise in everyday situations. Parallel parking, and backing up a trailer, constitute two fairly challenging examples most of us have had some experience with. This paper describes a training method which renders neural networks capable of finding approximate solutions to the nonholonomic path planning problem.