When using the EKF (or standard KF) on a real robot, you will want to tell the filter how much noise there is in each measurement, and in the process.
The purpose of this is so that the Kalman filter can decide how much it "trusts" each source of data, and therefore, the weighting to give each measurement in its final estimation.
For real robot data, the noise is already in the measurement. I think when you say "noise matrix", you are referring to the covariance matrix. This is not the actual noise per se, but rather, the noise covariance matrix describes the magnitude of the noise (that can be expected by the Kalman Filter), and the correlation between different noise terms, for a normal distribution of noise. You will generally want as accurate a noise covariance as possible, however, it can simply be estimated. When working with real data, you can perform a quick experiment to estimate the noise covariance, or you can also estimate it by consulting datasheets, or select a somewhat sensible value. Where there is not much data available, the covariance will normally be a diagonal matrix (ie. no correlation). The diagonal elements of the covariance matrix is also referred to as the variance. That means that you are telling the Kalman filter what the variance of the different noise sources are (square of standard deviation of the noise).
If on the other hand, you are wondering why the models related to the Kalman Filter may have a noise term (as opposed to a noise covariance), they are only models, and those equations are not actually used in the algorithm. The equations used by the algorithm will have terms representing the noise covariance (not the actual noise - which is unknown), which it normally keeps an online estimate of.