Neural Networks Based Navigation and Control of a Mobile Robot in a Partially Known Environment

Abstract

A complete integrated system for navigation and control of a mobile robot in a partially known environment is developed. The neural network in the path-planning algorithm can automatically build up the collision penalty function (as an obstacles' potential function approximation). The training procedure uses the information of the environment, no matter how many obstacles there are and how they look like. This makes the algorithm effective, but because of its gradient character, it meets the local minima problem. The local obstacle avoidance controller performs very well under accidentally encountered obstacles. Probably this is due to the property of the artificial amygdala to be the short term memory for the 'fear' from the encountered obstacles. The trajectory tracking controller is highly adaptable to uncertaities and changes in the robot dynamics. The recursive least squares technique using Givens QR decomposition is a comparatively fast and easy way for training the RBF net's weights. Future work will include the development of a modification of the pathplanning algorithm, which aims at filling the local minima along the path to calculate and thus to ensure reaching the goal reliably. This modification, as well as the overall integrated control system, will be further verified on a real mobile robot.