The abstract RNN can be used to learn the kinematics of a robot arm (Hoffmann and Möller, 2003). Here, a pattern consists of the coordinates of the end-effector, the joint angles of the arm, and a binary collision variable. By completion of a pattern that only includes the end-effector coordinates and the collision state, a set of joint angles can be obtained; this is the inverse direction. Analogously, the forward direction maps from the joint angles to the end-effector coordinates and the collision state. For a given end-effector position, redundant arm postures exist. Apart from performance tests for the inverse and forward directions, two further tests show that the abstract RNN can cope with additional noise dimensions (see also section 3.1) and that the performance of the abstract RNN depends on the number of input dimensions.