In our behaviour-based robot system a manipulator has been equipped with a monochromatic video camera, fastened onto the robot hand. Based on image processing and neural network learning the system executes goal directed perception-action cycles and thus attains versatile skills. Especially this work reports on close range manipulator navigation, i.e. searching collision-free trajectories of the robot hand to approach and handle goal objects. Neural network learning with radial basis functions (RBFs) is involved twofold. First, a function is learned for reconstructing from the optical flow of detected obstacle points their three-dimensional positions. Second, a function of the inverse manipulator kinematics is learned which is used for describing the non-rigid space occupied by the agile manipulator. Furthermore, based on the goal position and the continually detected obstacle positions a vector field is created dynamically by using the gradient of RBFs as basis fields. The vector field simulates attracting and repelling forces for navigating the manipulator hand. To overcome the curse of dimensionality and reach acceptable efficiency in function learning we applied mixtures of RBF neural networks and strongly emphasized divide-and-conquer strategies. The parallel approaches for neural learning (and image processing) are implemented on a four-processor general purpose workstation.