What is the difference between any old number and a numerical feature? Well, in reality, any numerical data can be used as an input variable. However, in a machine learning model, you learn about a vector of weights for each feature. The weights play a role in mapping feature values to an outcome or target variable (in the case of supervised learning models).
Thus, we want to use features that make sense, that is, where the model can learn the relationship between feature values and the target variable. For example, age might be a reasonable feature. Perhaps there is a direct relationship between increasing age and a certain outcome. Similarly, height is a good example of a numerical feature that can be used directly.