Dynamic Cell Structures for Calibration-free Adaptive Saccade Control of a Four-Degrees-of-Freedom Binocular Head
This paper describes calibration-free adaptive saccade control of a four-degrees- of-freedom binocular head by means of Dynamic Cell Structures (DCS) . Incremental growth of this Radial Basis Function (RBF) based neural network model up to a pre-specified precision results in very small networks suitable for realtime saccade control. By learning and exploiting the topology of the input manifold the controller output calculation is particular fast. Training of the DCS is based on biological inspired error feedback learning and proceeds in two phases. In the first phase we use a crude model of the cameras and the kinematics of the head to learn the topology of the input manifold together with a rough ap proximation of the control function off-line. Different to e.g. Kohonen-type adaptation rules the distribution of neural units is shown to minimize the control error and not to merely mimic the input probability density. In the second phase, the operating phase, the linear output units of the DCS continue to adapt on-line. Besides our TRC binocular head we use a Datacube image processing system and a Staeubli R90 robot arm for automated training in the second phase. The controller is demonstrated to successfully correct errors in the model and to rapidly adapt to changing parameters. The paper also includes a comparison with a conventional (calibrated) inverse kinematics based controller.