Parallel neural processing for the visual servoing of a robot arm

We propose a neural approach to visual servoing to improve the characteristics and the interest of this technique in robotics. The neurobiological analysis of the human arm suggests the separation of the arm motor controls into two centers of decision: 1) for the positioning of the hand; and 2) for its configuration. Considering this analogy and seeking the amelioration of the neural efficiency, we introduce the notion of NeuroModule for the visual control of a robot end-effector. Using the same visual input features, we show that the control of the arm can be realized either with a single module, or better with two parallel processing modules using each a part of these features. The use of two interconnected modules reduces the size of the neural networks involved in the process and facilitates the synchronization of the reaching and grasping preparation tasks. The approach is illustrated on a robot-vision system with both simulated and implemented experiments