A distributed robotic control system based on a temporal self-organizing neural network

A TEMPORAL SELF-ORGANIZING NEURAL NETWORK GUILHERME BARRETO1, ALUIZIO ARA UJO2, CHRISTOF D UCKER3, HELGE RITTER4 1;2 Dept. Ele tri al Engineering, University of S~ ao Paulo, S~ ao Carlos, SP, Brazil fgbarreto, aluizioag sel.ees .s .usp.br 3;4 Neuroinformati s Group, University of Bielefeld, Bielefeld, Germany f hrisd, helgeg te hfak.uni-bielefeld.de Abstra t|A distributed robot ontrol system is proposed based on a temporal self-organizing neural network, alled ompetitive and temporal hebbian (CTH) network. The CTH network an learn and re all omplex traje tories using two sets of synapti weights, namely, ompetitive feedforward weights that en ode the individual states of the traje tory, and hebbian lateral weights that en ode the temporal order of traje tory states. Ambiguities that o ur during traje tory reprodu tion are resolved using temporal ontext information. Also, the CTH network saves memory spa e by maintaining only a single opy of ea h repeated/shared state of a omplex traje tory. A distributed pro essing s heme is proposed to evaluate the CTH network in point-to-point real-time traje tory ontrol of a PUMA 560 robot. The performan e of the ontrol system is dis ussed and ompared with other neural network approa hes. Keywords| Self-organization, neural networks, temporal sequen es, roboti s, distributed ontrol. I. Introdu tion Control of movements in both biologi al and arti ial systems demands the availability of several sensorimotor transformations whi h onvert sensory signals into motor ommands that drive a set of mus les or roboti a tuators. Su h transformations are highly nonlinear and it is very diÆ ult to express them in a losed analyti al form. Artiial neural networks (ANNs) an be used to learn one or more sensorimotor transformations required to perform a given roboti task without pre ise knowledge of the robot parameters [1℄. Usually, roboti tasks have a well-de ned sequential nature in the sense that a given robot arm should assume spe i on gurations (states) su essively in time along a prede ned path. Usually, this temporal hara teristi is not in orporated into the learning pro edure, whi h implies that only stati sensorimotor transformations an be learned by the network. In these ases, the temporal order of the roboti task at hand is set in advan e by the the ANN designer. An alternative is to use temporal ANNs whi h an dire tly take into a ount sequential aspe ts of the roboti task. Su h networks should learn to asso iate onse utive states of a traje tory and store these state transitions for posterior reprodu tion. Usually, for purpose of re all, the network re eives the urrent state of the robot and responds with the next one, until the traje tory is ompleted retrieved [2℄, [3℄. In this work we aim to emphasize the feasibility of applying temporal self-organizing neural networks to real-time, distributed ontrol of roboti manipulators. The performan e of the learning algorithm is evaluated based on its ability to learn and retrieve a urately and without ambiguities a omplex traje tory. The remaining of the paper is organized as follows. In Se tion II, the neural network is presented. In Se tion III, the distributed ontrol plataform and its main omponents is introdu ed. In Se tion IV, tests with a PUMA 560 robot are des ribed. The paper is onluded in Se tion V. II. The Ar hite ture of the Neural Network The proposed neural ar hite ture, alled Competitive and Temporal Hebbian (CTH) network, is shown in Fig. 1. The CTH network basi ally onsists of feedforward and lateral weights that play distin t roles in its dynami s; it also has ontext units at the input and delay lines at the output. The delay lines, however, are needed only for training in order to learn state transitions.