Neural Networks for Robot Eye-Hand Coordination

In this paper learning the eye-hand coordination with Artificial Neural Networks is discussed as well the requirements and relevance to its practical application. A system to position an endeffector in 3D to grasp objects with an eye-in-hand camera system is presented both in simulation and in practice. The accuracy is discussed in relation to the number of learning samples and the adaptation period. In particular when objects have to be tracked high demands are encountered for the vision system. A vision system is presented which is able to detect moving targets in a cluttered dynamically changing environment, based on a multi-resolution scale. In is shown that also dynamic visual information can be used to control the robot acceleration.