Development of static neural networks as full predictors or controllers for multi-articulated mobile robots in backward movements - new models and tools

Abstract

This article presents a new method and tools for the development of full neural predictors and controllers, with fixed time horizon, based on static multilayer feedforward networks, when describing backward movements of multi-articulated mobile robots, in the configuration space. The predictors are necessary for robot's assisted tasks and useful to be used as cores in simulators to synthesize and validate strategies for control or navigation. The controllers can be generated by real and modelled data or using the predictor. The proposed systematic and the developed tools are general. The training data set is composed by real data and by data generated from circular singular condition models. The article presents original models for the singularities and for the critical angles. These models were deduced from general movement equations of a MAMR with on-axle or off-axle hitching and with front and rear traction on the truck. The use of models for singularities is necessary because the singular conditions are situations of unstable equilibrium, which makes impossible to obtain enough data from open loop real systems. The model for critical angles is used to define the range for data acquisition before the jackknife. An interface is created to assist users for training and validation of predictors and controllers. This interface helps to find a prediction horizon for the model and to make easier the processes of creating, training and validating a neural network and setting its training parameters.