Unsupervised learning methods were developed and applied to sensorimotor models, which in turn were used to solve perceptual tasks. For each sensorimotor model the following three steps were pursued. First, data were collected and preprocessed. These data are distributed in a sensorimotor space, whose dimensions comprise all sensory variables and all motor variables. Second, this distribution was approximated by a simplified and generalizing representation. Third, upon this approximation, recall mechanisms were developed that could complete a partially given input pattern, and thus allowed the association of patterns. The advantage of the new learning methods over a multi-layer perceptron and a look-up table was demonstrated. Finally, it was shown how sensorimotor models can be used as a basis for the perception of object-shapes and space.