Skip to Main content Skip to Navigation

Approche neuromimétique modulaire pour la commande d'un système robot-vision

Abstract : The laboratory works on the neural network control of a robot-vision system. This thesis concerns the application of artificial neural networks to control a robotic arm with visual servoing. The study focuses on modular learning to carry out neural networks controllers. We used self organizing maps to learn robotic functions which had been chosen for the simplicity of their algorithm as well as their capacity of generalization. These advantages enable us to have an online learning process, in real time. However these map can't deal properly with the important size of the input space following from the complexity of the task. To resolve the dimensionality problem we propose to divide the task into modules, each of them being composed of small-size artificial neural networks. With a sequential decomposition the learning phase of the whole structure becomes more difficult. Indeed not all desired inputs and outputs (which are examples for the learning process) of each module are available. Therefore a learning process is enabled by introducing the modules into the inverse flow which can estimate other introduced intermediary variables. The learning process in then bidirectional. The modular approach can therefore be considered as a neural network controller. The visual servoing has been validated on two complex robotic tasks: determining the angular configuration of a three- axis-robot and estimating the effector orientation of a four-axis-robot.
Document type :
Complete list of metadata
Contributor : Gilles Hermann <>
Submitted on : Friday, January 11, 2008 - 11:36:40 AM
Last modification on : Wednesday, July 22, 2020 - 4:04:02 PM
Long-term archiving on: : Thursday, September 27, 2012 - 2:05:49 PM


  • HAL Id : tel-00203857, version 1



Gilles Hermann. Approche neuromimétique modulaire pour la commande d'un système robot-vision. Automatique / Robotique. Université de Haute Alsace - Mulhouse, 2004. Français. ⟨tel-00203857⟩



Record views


Files downloads