A query based learning algorithm is proposed in this paper to obtain a valid neural emulator of the robot manipulator using radial basis function networks. This algorithm is centred around the concept of network inversion for which we have proposed an extended Kalman filtering (EKF) based algorithm. Initially the training data set is generated by tracking various random trajectories in robot work space. During this process, control input (or joint actuating torque) vector is computed using a fixed PD control scheme with added dither signal. The RBF network thus learned is tested for its validity using the inversion based recall process. In case the neural network is partially trained, additional input/output data pairs are generated. The data so generated are used to retrain the RBFN model of the robot manipulator. Experiments showed significant improvement in the neural model after retraining. © 1996 Taylor & Francis Group, LLC.