A modular neural network architecture for step-wise learning of grasping tasks

This paper proposes a neural network architecture for learning of grasping tasks. The multineural network model presented in this work, allows acquisition of different neural representations of the grasping task through a successive learning over two stages in a strategy that uses already learned representations for the acquisition of the subsequent knowledge. Systematic computer simulations have been carried out in order to test learning and generalization capabilities of the system. The neural activity at different subparts of the artificial neural network during its performance phase, is compared to the activity of populations of real neurons in areas AIP and F5 of the distributed parieto–frontal biological neural network involved in visual guidance of grasping. A more biologically plausible development of the model presented here is also discussed. The proposed model can be also used as a high level controller for a robotic dextrous hand during learning and execution of grasping tasks.