THE TRAJECTORY DETERMINATION AND THE KINEMATIC PROBLEM SOLVING FOR A HUMAN ARM MOVEMENT: A NEURAL NETWORK APPROACH *M. Costa, *P. Crispino, **A. Hanomolo, *E. Pasero *Politecnico di Torino, Italy, Dipartimento di Elettronica 24, Corso Duca degli Abruzzi, Torino, Italy **Universite Libre de Bruxelles, Faculté des Sciences Appliquées Service d’Automatique, Av. F.D. Roosevelt 50, 1050- Bruxelles, Belgique e-mail: ahanomol@labauto.ulb.ac.be ABSTRACT The article is the result of a preliminary study concerning the possibility to simulate human arm movements. The research work will be continued on a larger scale in the frame of an europena project -ANNIE. The objective of the work was to implement a neural network based methodology for simulating human movements in order to check and ensure ergonomical properties in differnet environments like work cells, cars (interior?) etc. If most of the studies are concerned with robot like movements, only a few have dealt with human movements. The neural networks potentiality s proved in this case, too: feed-forward nets are employed and the resilient backpropagation is used as training algorithm. INTRODUCTION CAD tools for 3-D modelisation can be used to simulate working environments and study ergonomic parameters for a correct positioning of the objects. The best way to verify the man-object interaction is to have a human-like mannequin moving in the environment. Robotic mannequins are available to simulate human actions. The robotic kinematics is well defined by several equations system, but the human movements are not easy to model. The simulation of the mannequin-objects interaction is therefore affected by the lack of precision of this approach. Some works suggest to use Artificial Neural Networks (ANN) to simulate the human movements. ANN demonstrated their efficiency when used to emulate specific human jobs, such as handwritten recognition, voice recognition etc. The main advantage of these systems is that they don't need any model of the job they have to emulate. An example based training is sufficient to have a network able to replicate the human behavior. The main drawback is that the generalisation of these systems is related to the number of examples you use to train the network. The more examples you use the best performance you have. The paper is part of a complex work consisting in the simulation of a human arm movement. The objective is to build up a methodology for simulating human arm movements in order to check the ergonomic features of various environments (a workstation, a car's interior etc). The work consists in the accomplishment of a procedure which using experimental data will provide a neural interface able to control a model build in a CAD. In our case the CAD is ROBCAD implemented by Tecnomatix which allows the simulation of 3 dimensional robots. STATE OF THE ART IN NEURAL NETWORKS APPROACH FOR THE ARM TRAJECTORY FORMATION Experimental observations of human unconstrained point-to-point reaching movements have indicated that these movements are characterized by straight hand paths [Cru90],[CB87] and symmetric bell-shaped velocity profiles that tend to remain invariant, despite variations in movement direction, speed and initial position. In order to control voluntary movements one must solve