Abstract Requirements for verifying spatial relations in robot workcell in terms of accuracy and repeatability are increasing. Improvements in the performance of industrial robots have extended the range of applications in to new fields in which flexibility, high payload, accuracy and repeatability are needed. To satisfy the requirements of overall geometric performance and flexibility in a robot system, a sensor-based, intelligent robot can be used. One of the goals of this thesis was to develop a flexible, CAD-based robot system. Modern applications and cost-effective production require off-line programming, and the difference between off-line programming systems and actual robot workcells has to be illustrated somehow in order to verify the gap between simulation models and actual robot systems. A method for modelling spatial uncertainties in a robot system is presented here, based on Bayesian-form estimation of model parameters and of the spatial uncertainties in the resulting parameters. The calibration of the robot workcell consists of several phases: hand-eye calibration, localization of the work object and estimation of model parameters for the work object surface. After localizing the work object, a finalization task can be carried out, e.g. inspection, manufacturing or assembly. A synthesis method of sensing planning that uses the same form of modelling spatial uncertainties is also presented. The deviation between covariance propagation models and actual systems is reduced by using detailed noise models of the robot system, including measured noise, at different phases in the calibration. The methods developed here were tested with simulation and extensive actual tests in each phase. The evaluation criteria used were eigenvalues in the directions of eigenvectors of the error covariance matrix. A careful analysis of spatial uncertainties was carried out to test the reliability of the covariance propagation method when the level of noise is changing, the results suggesting that the method is also applicable in such cases. The sensing planning method was compared with different types of sets of samples and the results analysed by considering the a posteriori error covariance matrix for the estimated parameters.