A neural-based model for the control of the arm during planar ballistic movements

A software model simulating the learning process of planar ballistic movements of the arm was developed, using the following scheme: an artificial neural network (modelling the neural system), a pulse generator (a computational block driving the biomechanical model of the arm), a two degrees of freedom manipulator guided by a six-muscles model. The learning scheme was implemented in an unsupervised way, thus not back-propagating the error information on the arm final position with respect to the expected target, but associating movements between two space positions (network inputs) to muscular activations (network outputs). After a training consisting of about 45.000 simulated movements, the model reached a mean distance error consistent with the experimental data found in typical ballistic movements.