A time series approach is used to model robot trajectory deviations. Trajectory deviations of the end effector are measured on a PUMA 560 robot programmed for straight line horizontal motion. Autoregressive Moving Average (ARMA) models are fitted to the data and used for predicting trajectory errors in the workspace. It is shown that while a model found for a certain trajectory is optimal only for that trajectory, it still has good prediction capabilities elsewhere in the workspace since the dynamics contained in different models are similar. The characteristics of the models in the frequency domain are also examined. The performance of models found using different velocities and different payloads are also compared. It is found that models obtained for a certain velocity adequately predict trajectory errors when used on data corresponding to a different velocity. Similar results are found using different payloads. These results show that ARMA models can be found off-line and used for predicting trajectory errors even in changing operating conditions.

The prediction errors are treated as differential translations of the end effector which are then converted into differential joint rotations using inverse kinematics. These corrections implemented in a controller would permit accurate tracking of robots.

Since the ARMA models perform well throughout the workspace and in changing operating conditions, robots can adapt to their working conditions using a method which is simple and efficient.